Meta says its layoff decisions were “made by people, not AI.” But that response leaves a harder governance question unanswered: Was the full decision system—the business objective, the measures, the safeguards, the human authority, and the escalation pathways—designed to prevent harm?

Hi, I’m Dasheika. I work in responsible technology and AI safety and governance, helping organizations address the upstream conditions that create business risk and harmful outcomes before they become incidents.

Rather than attempting to determine what happened inside Meta, I use these allegations as a stress test to examine where governance gaps can emerge and what leaders and practitioners can put in place before risk becomes harm, liability, or the next headline.

Learn More

Below are five areas of decision-system oversight—1) upstream governance, 2) full business-outcome design, 3) measure validity, 4) contestability and reconstructability, and 5) meaningful human judgment—grounded in established AI governance, employment, and risk-management practices. For each, I examine the reported trigger, what sound technical and human governance should require, and a proactive step organizations can take to surface risk earlier.

This is operational guidance, not legal advice. In this piece, “high-risk AI” refers to technology that materially informs or influences a consequential employment decision.

1. Governance must begin upstream, before the system is built.

Trigger

The complaint alleges that many of the scores and ratings used to inform Meta’s layoff selections could not be accumulated by employees on protected medical or family leave or whose measured output was reduced by disability.

One plaintiff also alleges that a manager discouraged him from taking approved medical leave and warned that doing so would result in his selection for the layoffs. The AP report presents these as allegations, which Meta disputes.

Together, the allegations raise whether employment protections and governance expectations were translated across the full decision lifecycle, and whether the measures and incentives surrounding the process reinforced those protections in practice.

AI Safety and Governance Baseline: Technical and Human

Governance should begin when the use case, data, measures, and system requirements are defined. That means translating employment protections and foreseeable risks into data rules, system logic, limits on use, exception pathways, monitoring, and escalation from development through everyday use.

Developers, providers, and deployers must understand the operating context, where outputs may be incomplete or misleading, and when the technology should no longer inform the decision. Without that context, a system may operate as designed while still producing information that is incomplete or inappropriate for the decision in front of it.

The technical foundation alone is not enough.

The people developing, approving, implementing, and using the system need defined responsibilities, authority to question or stop the process, and accountability for the outcome. One accountable owner must connect those responsibilities across every handoff so that risk does not disappear between functions.

Those responsibilities must also be reinforced by the organization’s incentives. Protections can exist in policy while managers and reviewers are rewarded for visible activity, uninterrupted output, speed, or compliance. Technical controls cannot compensate for weak authority or conflicting incentives, and people cannot correct risks they cannot see or are not supported in challenging.

Federal guidance prohibits discouraging eligible employees from using FMLA leave and using an employee’s request for or use of protected leave as a negative factor in employment decisions. NIST guidance similarly calls for differentiated roles, clear accountability, continuing governance, and organizational practices that support critical thinking and responsible challenge throughout the AI lifecycle.

Together, those expectations point to the same requirement: the protections expressed in policy must remain intact as they move through the technology, workflow, management practices, and final decision.

Proactive Practice

Before high-risk AI informs a consequential decision, require a decision-system governance map that identifies:

• the protections the process must uphold;

• how those requirements appear in the data, technology, workflow, and review process;

• who owns each lifecycle stage and what is their process;

• who can intervene or stop the use, how, in what way, what are the paths and procedures for escalation and intervention;

• what incentives are in place to support or hinder the process;

• who monitors whether policy and practice remain aligned; and

• which executive remains accountable for the final outcome.

Use that map as an approval requirement, not simply as documentation. Approval should depend on named ownership, real intervention authority, and evidence that the organization’s technical controls, human responsibilities, incentives, and oversight mechanisms reinforce the same protections in everyday practice.

Without that connection, the organization may govern the model while leaving the consequential decision around it exposed.

Governing the model and governing the decision are not the same accomplishment.

2. Efficiency is not the full business outcome.

Trigger

The complaint alleges that Meta used internal AI systems and quantified performance measures to help determine who would be selected for layoffs.

It further alleges that many scores could not be accumulated by employees on protected medical or family leave or whose measured output was reduced by disability, and that the process did not account for those circumstances before selections were made. The allegations remain unresolved.

Separate reporting quoted an April memo from Meta Chief People Officer Janelle Gale, who said the layoffs were part of an effort “to run the company more efficiently and to allow us to offset the other investments we’re making.”

The efficiency objective was explicit. It does not establish how the process was designed, but it sharpens the question raised by the allegations: Was efficiency treated as one requirement of the outcome, or as the outcome itself?

AI Safety and Governance Baseline: Technical and Human

The full business outcome should be defined before the process is automated. Speed, consistency, and efficiency may be legitimate objectives, but they cannot become the complete definition of success. The process must also satisfy the legal, regulatory, human, operational, and risk requirements governing how the result is produced.

In an employment context, that means translating protections related to leave, disability, accommodation, pregnancy, and discrimination into process and system requirements from the beginning. Those requirements should shape:

• how the workflow treats absences and reduced activity;

• which circumstances require additional context;

• where, in what ways, and to what exten technology may inform the process; and

• where its use must be limited.

These are not downstream compliance checks to apply after the decision logic has already been established. Department of Labor guidance prohibits using protected FMLA leave as a negative factor in employment actions. EEOC guidance under the ADA requires covered employers to provide reasonable accommodation for the known limitations of qualified employees with disabilities unless doing so would create undue hardship. The Pregnant Workers Fairness Act similarly requires reasonable accommodations for known limitations related to pregnancy, childbirth, or related medical conditions, subject to its requirements and exceptions.

In practice, developers and providers need a clear understanding of the intended use and the conditions the system must handle. Deployers remain responsible for determining whether the technology is appropriate for the process, how it will be configured, and whether the underlying workflow can produce the full outcome.

Legal, HR, risk, technical, and domain expertise should therefore remain connected through development, deployment, implementation, and adoption. Without that connection, the operational objective can gradually displace the requirements that make the outcome valid.

Proactive Practice

Before developing or deploying high-risk AI, require a cross-functional review of the complete business outcome and the process designed to produce it.

That review should define the operational objective, translate the legal, regulatory, human, and risk requirements into the workflow and supporting technology, and assign clear ownership for confirming that those requirements remain intact through implementation and use.

Do not approve the technology simply because it can improve speed, consistency, or cost. Approval should depend on evidence that the process and technology can achieve the operational objective without creating preventable harm, noncompliance, or avoidable liability.

When efficiency is separated from the conditions governing how it is achieved, the organization may improve one metric by shifting the consequences somewhere else.

A business outcome is not successful when efficiency is achieved by transferring risk elsewhere.

3. Measures require evidence of validity, not just precision.

Trigger

The complaint alleges that Meta used keystroke and activity-monitoring data, AI token-usage dashboards, and algorithmically assisted performance rankings to help inform layoff selections.

It further alleges that employees on protected leave or whose output was affected by disability could not, by design, accumulate some of those measures and were disproportionately selected for layoffs.

The AP report attributes these claims to the lawsuit and notes that Meta disputes them.

The allegations raise whether those measures were valid indicators of contribution for this employment decision and whether they performed appropriately across the populations and circumstances the process affected.

AI Safety and Governance Baseline: Technical and Human

Validity depends on the measure, the conclusion drawn from it, and the context in which it is used. Data can be collected accurately and applied consistently while still failing to represent contribution, performance, or suitability for a consequential employment decision.

Before use, an organization should establish what each measure represents, whether it is job-related, how strongly it supports the decision, and where the inference may fail. Validation should examine subgroup performance, protected circumstances, foreseeable proxy effects, potential disparate treatment, and disparate impact.

It must also account for the conditions that produce the data. Role design, work assignments, access to opportunities and technology, leave, disability, and accommodations can affect who is able to generate a measured signal. A seemingly neutral measure can therefore reproduce or compound historical and systemic disadvantage.

That context matters legally as well as technically. Under Title VII, when a selection procedure has an adverse impact based on race, color, religion, sex, or national origin, the employer must be able to show that the procedure is job-related and consistent with business necessity. The Uniform Guidelines describe methods for establishing validity.

NIST’s AI risk-management guidance similarly connects validity to context of use, performance under real operating conditions, documentation of limits, and continuing measurement as the system and its environment change.

The assessment therefore cannot remain within the data or technical function. Employment, legal, data, technical, operational, and risk experts must determine together whether the measure reflects the work being evaluated, how organizational conditions shape the signal, and whether the resulting inference is defensible for the intended use.

Proactive Practice

Before any quantified measure or AI-supported score can influence a consequential employment decision, require a documented, use-specific validation and impact review.

The review should address:

• Meaning: Define what the measure represents, what it cannot establish, and how much weight it may carry.

• Relevance: Confirm that it is job-related and relevant to the specific decision.

• Opportunity: Examine whether affected employees have comparable opportunities to generate the signal.

• Context: Account for how leave, disability, accommodations, role design, work assignments, access, and other historical or organizational conditions shape the data.

• Impact: Test for subgroup differences, proxy effects, potential disparate treatment, and disparate impact.

• Interaction: Evaluate the measure individually and in combination with other scores, thresholds, rankings, or decision rules.

• Alternatives: Compare whether another measure or method could achieve the business objective with less risk.

• Limits: Establish the conditions that require adjustment, restriction, revalidation, suspension, or removal.

Require accountable approval from the employment, operational, technical, legal, and risk functions responsible for the decision, with one owner accountable for whether the measure is appropriate for the intended use.

After deployment, monitor actual decision outcomes, not only whether the measure continues to operate as designed. Review subgroup patterns, recurring exceptions, changes in the work or affected population, and evidence that the measure is being used beyond its validated purpose.

If the organization cannot demonstrate that a measure supports a valid and defensible conclusion for the decision, population, and circumstances in which it is used, it should not be allowed to influence that decision. Precision cannot cure a weak inference or make an inappropriate measure fair.

Measurability is not validity. A measure must support a defensible conclusion without reproducing or compounding disadvantage embedded in the surrounding system.

4. Consequential decisions must be challengeable and traceable.

Trigger

The complaint alleges that Meta did not account for protected leave or disability-related circumstances when considering employee scores and did not provide an individualized review that was neutral to those circumstances before layoff selections were finalized.

The AP report attributes these claims to the lawsuit and notes that Meta disputes them.

The allegations raise whether affected employees had a meaningful opportunity to correct or contextualize the information used in the decision before the outcome became final. They also raise whether the organization could later account for how the measures, system outputs, safeguards, and human review produced the result.

AI Safety and Governance Baseline: Technical and Human

As a governance control, contestability gives an affected person a meaningful opportunity to correct inaccurate data, provide relevant context, and challenge how a measure or recommendation is being applied before a consequential decision becomes final.

Reconstructability serves a different but connected purpose. It creates a contemporaneous record of how the organization reached the decision, including the information and system output used, the limitations or exceptions identified, the context considered, who exercised authority, and why the final outcome was approved.

Neither safeguard is sufficient on its own.

Contestability has limited value when a person can raise a concern but the organization cannot show what the reviewer considered, whether the challenge affected the analysis, or why the original recommendation was accepted or changed. Reconstructability may create a complete record of a harmful decision while providing no meaningful opportunity to prevent it.

That means the technical system and the human process must support both prevention and accountability.

The technology and workflow must be able to surface missing or conflicting information, identify circumstances that require additional review, preserve the relevant data and system version, and pause the process when an exception is triggered.

The human process must establish when review is required, how the affected person can provide context, what information the reviewer must consider, and who has authority to alter or stop the outcome.

The central question is not merely whether an appeal or override exists. It is whether the challenge occurs early enough to affect the decision, whether the person reviewing it has real authority, and whether the organization preserves enough evidence to demonstrate how that authority was exercised.

Proactive Practice

Before high-risk AI or quantified measures inform a consequential decision, establish a contestability and reconstructability protocol with five components:

• Pause and escalation triggers: Define the missing information, conflicting evidence, protected circumstances, unusual patterns, or other conditions that automatically pause or escalate the decision.

• A meaningful opportunity to respond: Give the affected person a clear way to correct information, provide relevant context, and challenge how the measure or recommendation is being applied before finality.

• Review authority: Identify who must examine the challenge, what information that person must consider, and who can alter, override, or stop the outcome.

• A complete decision record: Preserve the data, measure, system output, version, known limitations, exception triggers, interventions, and final rationale used to reach the decision.

• Monitoring and organizational learning: Track whether required pauses occur, whether challenges influence outcomes, whether exception pathways are bypassed, and whether recurring challenges or overrides indicate that the measure, workflow, or original validation must change.

A consequential decision should not become final until the required challenge and review processes are complete. The organization should then be able to reconstruct the full path from the original measure or recommendation to the approved outcome.

Together, these safeguards create an opportunity to prevent avoidable harm before finality and a record that makes the organization’s exercise of authority accountable afterward.

Contestability can prevent an avoidable outcome. Reconstructability makes the decision accountable. A consequential decision system needs both.

5. Human involvement is not the same as meaningful human judgment.

Trigger

Meta denies that AI made the layoff decisions and says, “Workforce management and organizational decisions were and are made by people, not AI.”

That distinction matters, but the presence of a person in the process does not by itself establish effective oversight.

The allegations raise what organizations should require from the people expected to interpret, review, approve, or challenge AI-supported employment decisions.

AI Safety and Governance Baseline: Technical and Human

Meaningful human judgment requires more than a final approval, acknowledgment, or opportunity to override a system output.

A reviewer must understand the decision and its operating context, have access to the relevant evidence and known limitations, and have enough time to conduct an independent assessment. The reviewer must also have clear authority to question, pause, change, or reject the recommendation and be accountable for the quality of the review and the resulting decision.

The technical system must make that judgment possible. It should show the basis of the recommendation, missing or conflicting information, known limitations, relevant uncertainty, exception triggers, and the boundaries of the measure’s approved use.

Timing is equally important; human review must occur while intervention can still change the result. A reviewer who enters only after the ranking, recommendation, or outcome is effectively fixed may validate the process without meaningfully governing it.

Formal authority is also not enough when the surrounding incentives discourage its use. A reviewer may technically be permitted to intervene while being rewarded for speed, consistency, cost reduction, or agreement with the system. In practice, meaningful judgment depends on whether the organization expects, supports, and protects responsible challenge.

The relevant test is therefore not whether a person appeared somewhere in the workflow. It is whether that person had the knowledge, evidence, time, authority, independence, and organizational support required to reach a reasoned conclusion and affect the outcome.

Proactive Practice

Before assigning a human reviewer to a consequential AI-supported decision, establish a documented human-review standard for the role.

That standard should include:

• A role and competence profile: Define the knowledge, experience, and domain expertise needed to assess the decision and understand the limits of the system or measure.

• A decision-information package: Specify the evidence, contextual information, system limitations, uncertainty, conflicting information, and exception indicators the reviewer must receive.

• An effective intervention point: Position the review early enough for the reviewer to change the recommendation or outcome before it becomes final.

• Decision rights and escalation: Grant explicit authority to pause, investigate, escalate, override, or reject the recommendation, supported by clear criteria and escalation pathways.

• A reasoned review record: Document what the reviewer considered, what questions were raised, what changed, and why the recommendation was accepted, modified, or rejected. Recording only that a person approved the decision is not enough.

• Effectiveness testing: Monitor whether human review changes outcomes, identifies recurring weaknesses, surfaces missing context, or functions primarily as confirmation. Use those findings to strengthen the reviewer role, the underlying system, or the broader decision process.

The organization should be able to demonstrate that the reviewer understood the evidence, exercised independent judgment, and had real authority to affect the result.

Without those conditions, the human role may provide legitimacy to the system’s recommendation without providing meaningful oversight of it.

A human-in-the-loop label is not evidence of oversight. Meaningful human judgment exists only when people are equipped, empowered, and supported to challenge the system and change the outcome.

What This Requires

This is an organizational challenge that no single function can solve. Legal, Compliance, HR, Risk, Product, ML, the Responsible AI function or Center of Excellence, and the governance committee each bring different responsibilities, evidence, incentives, and definitions of risk.

The work must begin at the top, where governance, decision rights, and incentives are set, then move down and across the organization to align implementation, escalation, ownership, oversight, and accountability.

Without that alignment, each function can complete its part while the full decision system remains exposed. Cross-functional responsibility must still lead to clear authority and one accountable owner for the outcome.

Together, the proactive practices above also reduce risk exposure.

Laws, enforcement priorities, and regulatory expectations will continue to shift. The organizations best positioned are not the ones relying on a particular enforcement posture. They are the ones whose systems and processes can withstand scrutiny regardless of which agency is prioritizing enforcement in a given year.

None of the risks described here requires a rogue algorithm.

Risk can emerge when systems used to measure activity enter a consequential employment process without the process design, legal context, technical safeguards, human authority, and ongoing oversight needed to govern the resulting decision.

The failure point can sit in the disconnect between what leadership intends, what the process rewards, what the system measures, what people are enabled and empowered to do and challenge, and what the organization ultimately produces.

It is to govern the full decision system:

• the business objective;

• the data and measures;

• the technical controls;

• the operating process;

• the incentives;

• the people exercising judgment; and

• the outcome.

That is how organizations reduce the likelihood that a preventable risk becomes harm, liability, or the next headline.

Share

If This Pattern Looks Familiar

The pattern in this piece is a consequential decision system whose technical controls, human capability, capacity and judgment, process design, incentives, and oversight are not fully connected.

If any part of that sequence sounds familiar, a the Execution Check is designed to surface where the disconnect is showing up.

It can help determine whether the cause still needs diagnosis or the need is already defined and ready to build.

If you already understand the stakes and want to work through the issue directly, skip the diagnostic and request an Executive Conversation.

Start the Execution Check | connect@cultureclusive.com

Leave a comment

Share

Last Thursday, 4,500 Google employees delivered a petition to CEO Sundar Pichai asking for voluntary exits before forced cuts, guaranteed severance, and an end to quota-driven performance ratings. Alphabet Workers Union president Parul Koul framed the Google petition as workers wanting enough security to do their best work and bring new ideas to life, instead of operating in an environment driven by fear.

Coincidentally, that same week, I had posted a take on Meta’s crisis: the CTO telling staff in an internal meeting that morale was “probably one of the worst it’s ever been” in his 20 years there, then writing in a follow-up memo that he hoped to rekindle a culture “where people have the psychological safety to take risks and do the right thing,” in the same year the company plans to spend $145 billion on AI infrastructure.

Two different companies ended up naming the same missing condition (Psychological Safety) from opposite directions. Meta’s leadership confessed to it from the top. Google’s workers demanded it from the floor.

The logic connecting them is simple once you see it. Safety produces curiosity, creativity and risk-taking, risk-taking produces innovation, and fear produces compliance, which keeps dashboards green while the value never actually arrives. An exhausted system doesn’t always collapse. More often, it just complies.

In April, I published “Your AI Investment Is Landing on an Exhausted System,” arguing that organizations which heavily invested in technology need to place their next bet on the human domain. The layer that determines whether AI produces results isn’t built by deploying technology but through sustained, deliberate investment in the conditions under which people actually work. I also wrote that any business demanding maximum performance owes its workforce a floor, a minimum set of operating conditions required for people to carry what’s being asked of them, and that for many organizations, that floor has never been defined.

The Google petition is what happens when it isn’t. When leadership won’t define the floor, the workforce drafts it themselves and delivers it straight to the CEO’s office.

The Condition Is the Constraint

This isn’t a two-company problem. In February, the National Bureau of Economic Research published findings from nearly 6,000 senior executives across the United States, United Kingdom, Germany, and Australia, and nine out of ten reported little to no measurable productivity or employment impact from three years of AI investment. I examined those findings in April: the constraint was never limited to the technology and its capability, it was the workforce capacity and workplace conditions expected to absorb it.

The memo names psychological safety. The petition reaches into the governance and incentives that determine whether psychological safety can exist in the first place.

Governance determines who holds power, who makes decisions, and who carries the risk when those decisions fail. Incentives teach people which behaviors are rewarded and which ones make them vulnerable.

Repeated consequences become culture, which is why an organization can’t ask people to take risks while making employment itself feel arbitrary, can’t say it values candor while rewarding visible compliance, and can’t repeatedly reduce headcount, destabilize teams, and force performance distributions, then expect the trust required for people to challenge assumptions, expose problems, and offer ideas whose outcomes are uncertain.

Extraction doesn’t just cost the people pushed out. It exhausts the system that remains.

Companies have spent heavily on technical capability while weakening the human conditions required to put that capability to meaningful use. They’ve removed people and institutional knowledge, expanded workloads, increased instability, layered new mandates onto unchanged operating models, and expected the workforce to keep carrying the strategy anyway.

The Floor

Every business demanding maximum performance owes its workforce a floor: the minimum operating conditions required for people to carry what’s being asked of them. When a business expects sustained productivity, adaptability, judgment, creativity, and risk-taking, it has to account for whether its governance, incentives, and employment conditions actually make those contributions possible. You can only get as much from people as you’re willing to enable and empower them to give.

For many organizations, that floor has never been defined. The Google petition is what happens next. When leadership won’t define the floor, the workforce drafts it themselves and delivers it straight to the CEO’s office.

Not an Outlier

Google is the most visible case, not an isolated one. This spring, union members at Google DeepMind’s London office voted almost unanimously to pursue formal recognition covering roughly 1,000 staff. In May, technical workers across the University of California system formed what’s being described as the largest tech worker union in the country, with AI guardrails and layoff limits among its explicit goals. Industry reporting this month finds that the single biggest thing tech workers now want from a contract isn’t pay, it’s protection against arbitrary termination, and the counterforce is quantified too: American employers spend roughly $1.7 billion a year on union avoidance consultants and law firms.

Unions have begun negotiating AI guardrails directly into contracts, a frontier that didn’t exist a few years ago. Read the wave for what it is: the workforce building the floor the system never provided, one demand, organizing effort, and agreement at a time.

The Next Bet

This is why the next bet must be on humans, not as an HR initiative but as a strategic correction of the constraint. It will take more than HR it needs the engagement, ownership, of the entire C-suite and boardroom.

Leadership has begun naming what the organization needs. Workers have begun naming what providing it will require. The starting point isn’t more analysis, it’s defining the floor: the minimum operating conditions an organization owes its workforce before it expects judgment, risk-taking, and innovation in return.

Governance and incentives already exist inside every organization. Under extraction logic, where shareholder primacy is treated as a legal command rather than a strategic choice, they’re producing exactly what they’re built to produce: layoffs, eroded morale, and a

workforce that have learned to comply instead of contribute. A better snack budget won’t change what the system is optimized to do. Only a defined floor changes what governance is required to hold, and what incentives are built to reward.

Meta and Google aren’t the end of this story. They’re early proof of what happens when the floor stays undefined. Correcting course starts with naming it.

If This Is Landing for You

The pattern in this piece isn’t abstract. Capability funded. Capacity depleted. Compliance rewarded. Judgment withheld. If that sequence sounds like your organization, the Execution Check is built to surface exactly where it’s showing up, whether the cause is still unclear and needs diagnosis, or the need is already defined and it’s time to build.

If you already know the stakes and want to talk it through directly, skip the diagnostic and request an Executive Conversation.

Start the Execution Check · Request an Executive Conversation

Notes and sources

This piece draws on reporting from KQED and The Guardian on the July 16, 2026 Alphabet Workers Union petition; Business Insider, Wired, and Fast Company on Meta’s June 2026 all-hands remarks and internal memo and its 2026 AI infrastructure plans; the February 2026 National Bureau of Economic Research working paper surveying nearly 6,000 senior executives, examined at length in Your AI Investment Is Landing on an Exhausted System; Computerworld’s July 2026 series on tech worker unionization; coverage of the Google DeepMind London recognition vote and the formation of the UPTE technical workers union; and the May 2026 Economic Policy Institute and LaborLab report on employer spending on union avoidance. The floor argument and the capability and capacity distinction are developed in Your AI Investment Is Landing on an Exhausted System, and the company-level pattern is documented in The Post-Layoff Trap.

Organizations are encouraging AI adoption. Some are measuring it. Some are incorporating it into performance expectations. Educational institutions are adapting around it. Technology companies are embedding it into the products and services we use every day. Institutions are increasingly using it to inform decisions in hiring, lending, healthcare, policing, and other areas that shape opportunity and participation. Workers and students are expected to understand it and adapt.

Whether we actively choose it or not, AI is increasingly becoming part of the infrastructure through which people learn, work, create, communicate, and participate.

At the same time, I am noticing how quickly AI use can become a basis for questioning credibility, intelligence, originality, expertise, effort, and legitimacy.

The contradiction itself is interesting.

Society is increasingly integrating AI into the conditions of participation while simultaneously questioning the legitimacy of participation that involves AI.

We are being told to adapt. And then being judged for adapting.

The more I sit with that tension, the more it reminds me of conversations that existed long before AI.

Signals and Legitimacy

Historically, people have not always been evaluated on the substance of their contributions. They have often been evaluated through signals.

• Accent became a proxy for intelligence.

• Dialect became a proxy for intelligence.

• Educational pedigree became a proxy for intelligence.

• Professional language became a proxy for intelligence.

• Executive presence became a proxy for competence.

The signal became a proxy for the capability. The form became a proxy for the substance.

Sociologists have explored this through concepts like cultural capital. Linguists have examined language ideology and linguistic profiling. Others have written about respectability politics, code-switching, and the invisible labor many people perform to be perceived as credible within dominant institutions.

Different disciplines. A similar observation.

Institutions reward particular forms of expression, behavior, and knowledge. Over time, those standards become normalized. Eventually we stop experiencing them as assumptions and begin experiencing them as common sense.

As merit.

As professionalism.

As objectivity.

As simply “the way things are.”

That process fascinates me because it is often invisible while it is happening.

The standards institutions reward eventually become the standards individuals internalize.

We inherit them.

We reproduce them.

And eventually we stop recognizing them as standards at all.

We begin treating them as truth.

The standards institutions reward eventually become the assumptions individuals carry.

Which makes me wonder whether AI is colliding with a set of assumptions we rarely examine because they have become so deeply embedded in our institutions and, eventually, in ourselves.

Many of us were taught that intelligence is individual.

That expertise is owned.

That authorship is singular.

That merit is objectively observable.

That professionalism is neutral.

That standard forms of expression are inherently superior.

These ideas are often treated as universal truths.

Yet many emerged from specific historical, educational, professional, and economic contexts. They are not natural laws. They are social constructions that have become normalized through repetition and reinforcement.

Perhaps that is part of what makes this moment feel so charged.

Not because questions about authorship, attribution, expertise, effort, and truth are unimportant.

They are.

What interests me is something slightly different.

What AI May Be Disrupting

I find myself wondering what happens when those questions become a substitute for evaluating the work itself.

Because increasingly, I encounter moments where the existence or suspicion of AI seems to shift the inquiry away from understanding the work and toward determining its legitimacy, its credibility, and even its worthiness of evaluation, interpretation, and understanding.

The analysis, the argument, and the idea itself become secondary.

The focus shifts to legitimacy.

“Did they really write it?”

“Could they have produced this on their own?”

“How much assistance did they receive?”

“Should this contribution count?”

The questions are no longer only about the work.

They become questions about the person.

And that feels significant.

Perhaps because legitimacy has never been distributed evenly.

Legitimacy Has Never Been Distributed Evenly

Long before AI, many people learned that how they spoke, wrote, presented, dressed, and expressed themselves could influence whether their ideas were taken seriously.

Code-switching was never simply about communication. It was often about legitimacy. About learning which forms of expression would be recognized as professional, credible, intelligent, or worthy of being heard.

The work mattered.

But so did the presentation of the work.

The idea mattered.

But so did whether the institution recognized the person presenting it as a legitimate contributor.

That reality is not evenly distributed.

Which is part of why this moment feels familiar.

For some people, these questions may feel new.

For others, they echo older experiences.

Experiences of learning which signals were rewarded, which forms of expression were considered professional, and which versions of themselves would be recognized as credible.

AI is arriving in a society where legitimacy has never been distributed evenly.

A society where some people receive the benefit of the doubt more readily than others.

A society where assumptions about intelligence, expertise, effort, and professionalism already exist before the technology enters the conversation.

The question is not whether AI creates those dynamics.

The question is what happens when it collides with them.

Early research suggests that trust, adoption, and perceptions of AI are not distributed uniformly across populations. Different groups are approaching these technologies with different levels of access, confidence, concern, and scrutiny.

That alone is worth paying attention to.

Because if legitimacy was unevenly distributed before AI, there is little reason to assume those patterns disappear once AI enters the equation.

The Question Beneath the Question

AI may be disrupting some of the signals we have historically relied upon to make judgments about intelligence, expertise, effort, and credibility.

Signals that many people spent years learning to navigate.

Signals that institutions have long rewarded.

Signals that have never been applied equally.

If legitimacy was unevenly distributed before AI, and early evidence suggests trust, adoption, and scrutiny remain unevenly distributed as AI emerges, then the question may not simply be whether the signals are changing.

The deeper question is whether the assumptions beneath them are changing too.

If AI changes the signals, what happens to the assumptions underneath them?

Do we become more thoughtful about how legitimacy is assigned?

Or do we simply create a new set of proxies and repeat familiar patterns under a different name?

Which is why I am increasingly convinced that this is a sociotechnical problem we are exploring in fragments.

This is a conversation about participation.

It is a conversation about credibility.

It is a conversation about legitimacy.

And perhaps most importantly, it is a conversation about the assumptions we bring with us when we decide whose contributions are worthy of trust.

The more I observe this moment, the less interested I become in whether AI was used and the more interested I become in what our reactions to AI might reveal about ourselves.

About our institutions.

About our definitions of intelligence.

About our definitions of expertise.

About our definitions of merit.

About who receives the benefit of the doubt and who must continually earn it.

Society is increasingly integrating AI into the conditions of participation. Organizations are redesigning work around it. Educational institutions are adapting around it. Technology companies are embedding it into everyday life. Participation increasingly requires adaptation. Yet adaptation itself is becoming a basis for scrutiny.

That feels like a contradiction worth examining.

Not because AI should be beyond criticism.

But because the ways we assign credibility, expertise, intelligence, and legitimacy have never been neutral.

The technology may be new.

The questions feel much older.

I am still following the question.

TL;DR

High‑risk AI and automated decision systems are already helping decide who gets hired, who gets a loan, who gets care first, who waits in jail, and who goes home. For the people these systems were not designed for, the harm is not abstract or future tense. It is already here—showing up in the same patterns, in the same communities, with the same concentrated weight.

Most organizations now have an AI policy. Very few are actually ready to use high‑risk AI safely. The space between those two realities is where unaddressed risk turns into harm.

Closing that gap requires a shift from what AI can do to what it should do, from principles, policy manuals, and documents to the conditions that surface and mitigate business risk and harm.

Share

Subscribe now

As the corporate AI race continues, high-risk AI systems are being deployed across finance, employment, healthcare, and criminal justice to inform or determine consequential decisions in hiring, lending, healthcare triage, bail, and sentencing.

These domains were rife with discriminatory and harmful practices long before digital acceleration. AI is now inheriting and automating those same failures.

The infographic below documents consumer harm already happening across employment, lending, housing, healthcare, facial recognition, workplace surveillance, and criminal justice. These are harms that deny resources and opportunities, distort or erase how people and groups are seen, distribute quality of service unequally, and place direct consequences on individual lives, freedom, and dignity.

Not hypothetical. Not future tense. The harm is not coming. It is already here.

Citations are included in the graphic.

This requires a shift from what AI can do to what AI should do. AI governance and safety need to move from stated principles to built infrastructure. That means creating the organizational conditions that make safety an outcome, not just a goal, especially where systems are classified or functionally operating as high risk.

Most companies can point to an AI policy or committee. Very few are genuinely prepared to deploy AI safely at scale.

For many organizations, the space between those two realities becomes unmanaged regulatory, operational, reputational, and human risk. When AI governance is treated as a real risk management discipline and embedded in how decisions are made, organizations are more likely to avoid major failures and create durable value from AI. Governance is not only protection from downside. It shapes the organization’s AI risk profile and its ability to compete.

The Readiness and Risk Gap

In a lot of current practice, the model is where we first see harm, but it is not where the harm begins. The potential for harm is already present in the data and in how the problem is framed. Harm becomes real when those risks are not surfaced, challenged, or designed against during development.

It often starts with data and problem framing. Teams work with the data they can access, the proxies that are available, and the histories the institution already treats as normal. In lending, hiring, healthcare, and criminal justice, those histories reflect patterns of exclusion and unequal treatment. When those patterns are carried forward as ground truth, the model inherits them before a single parameter is tuned.

The work then moves into design and development. Cross functional product and ML teams translate that data and problem statement into a system that will operate in the real world. Under deadline and resource constraints, they decide which users to optimize for, which scenarios to treat as edge cases, and which harms feel unlikely. When those teams do not reflect the people most affected by the system, blind spots in the data are reinforced by blind spots in the room.

All of this happens inside an environment with a low safety climate. When incentive structures and norms reward speed over safety, human fallibility stops being a risk to manage and becomes something the system leans on. A reviewer who raises a concern can be treated as a blocker. A manager who ships on time can be celebrated. Psychological safety is limited. Authority gradients make challenge difficult. Small lapses normalize until deviance no longer looks like deviance. People who see the harm have no safe path to surface it.

The model then moves through a handoff. When the organization that provides the model is not the same as the one that deploys it, accountability diffuses and remains largely unaddressed. The provider assumes the deployer will catch issues. The deployer assumes the provider built something safe. Neither owns the feedback loop. Technical metrics are monitored. Human outcomes often are not.

The gap between them is where harm lives without an owner.

By the time the system reaches the people it is meant to serve, bias may have entered through the data, been reinforced in design and development, passed through testing under pressure, and been deployed into institutions that already affect people unevenly. The results of those decisions then feed back as new data and influence the next version. Over time, the cycle tightens if nothing in the lifecycle is explicitly designed to interrupt it.

This is not about individual engineers or product managers acting in bad faith. It is about how normal constraints on people, such as time, information, incentives, and psychological safety, interact with organizational structures that were never built with AI safety in mind. When that interaction goes unaddressed, each release quietly adds to both the organization’s risk and the potential for harm.

What AI Safety and AI Governance Covers Today

AI safety and responsible AI governance are related but not identical. Both are necessary. Neither is sufficient in their current form.

AI safety, in most organizations, focuses on how the system behaves. It is about testing, evaluation, red teaming, bias and fairness checks, and other technical work meant to find problems before or after deployment. This work matters. It is strongest at the model layer. It is weakest at the conditions around the model, such as who is in the room, how decisions are made, what incentives apply, and who can say no.

Responsible AI governance focuses on how AI is allowed to be used. It shows up as principles, policies, oversight committees, review processes, and accountability frameworks. This work matters too. It is strongest at the level of documented intent. It is weakest at the level of day to day execution, where policies compete with production pressure, gaps in ownership, and limited resourcing.

That is the limitation.

AI safety has matured further on the technical side than on the organizational side. Responsible AI governance has matured further on the policy side than on the operational side. One can show you that a model behaves as expected under test conditions while leaving the surrounding system unchanged. The other can define responsibility on paper without giving people the time, authority, or support to act in practice.

This leads to familiar result.

• A model can pass technical review and still produce harmful outcomes.

• A governance program can look mature in a slide deck and still fail to prevent or catch harm once the system is in the world.

Until both practices extend into the real conditions of development, deployment, oversight, and feedback, safety will remain more aspirational than operational.

Four Questions That Expose the Gap

Subscribe now

Four questions make this gap visible.

1. If humans are biased and fallible, and institutions have never resolved their historical and systemic biases, how can AI avoid replicating those same patterns in its inputs?

2. If the teams developing and deploying AI do not reflect the populations affected, how will potential harms be surfaced and resolved during the development lifecycle and after deployment, as models and environments change?

3. If the environments leading AI development are misaligned in their incentives, norms, behaviors, operating models, and accountability systems, how will human in the loop, human on the loop, and human review move from policy statements to real practice?

4. If the model changes during and after deployment and the deployer is not the same as the provider, what mechanisms exist to ensure feedback loops that address both technical performance and the human aspects of prevention, remediation, governance, and maintenance over time?

The is a structural problem requiring system wide solutions. Even when no one intends harm, the way data, teams, incentives, and ownership are arranged can steer systems toward harmful outcomes.

These questions are a diagnostic.

They surface the organizational factors leaders must address to govern, enable, empower, and operationalize the conditions required to reduce harm. They are part of what it looks like to mature AI safety from a technical practice into a discipline that protects users and reduces risk for the organizations that deploy it.

From Review to Safety by Design

Here is the specific limitation that keeps showing up in practice.

Current practice treats human review, human in the loop, and human on the loop as pre-deployment events. A compliance moment or signature before launch. That setup assumes a human reviewer can catch everything that matters at a single point in time, under production pressure, with no ongoing support, and that their judgment will remain stable over the life of the system.

In my experience, that assumption ignores two things.

First, individual fallibility. Humans get tired. Humans normalize deviance over time. What looked concerning on day one can look routine by day one hundred. Humans under deadline make different decisions than humans with protected time. Current practice leaves no room for this.

Second, organizational climate and constraint. Even a well trained, well intentioned reviewer operates inside an environment. If incentives reward speed over safety, the reviewer will internalize that. If there is no psychological safety, the reviewer will stay silent. If there is no rotation, no redundancy, and no second pair of eyes, the organization is betting everything on one person not failing.

Human fallibility is not a bug you patch with a signature. It is a system property. It lives in individuals and in the structures around them. Current practice addresses neither.

This is not a gap you close once. It is a limitation in the practice itself. It is a missing capability that requires ongoing maturation. The work is not to add more checkpoints. It is to change how the system is designed.

That means moving beyond policy documents, pre-deployment checklists, and isolated rounds of testing toward governance and operations that enable safe practice over time. In practical terms, that looks like three shifts.

1. Human review is redesigned. It becomes structured, auditable, and protected from production pressure. Reviewers know when and how to raise concerns, and they have clear authority and backing when they do.

2. Human in the loop is redefined in high-risk contexts. The loop cannot only be engineers and compliance officers. It needs to include affected communities and domain experts, especially where allocative, representative, interpersonal, and quality of service harms have historically landed on vulnerable groups.

3. Human fallibility is treated as a system property. Organizations build redundancy, clear escalation paths, and organizational memory so that safety does not depend on one reviewer, one team, or one moment in the lifecycle.

This is not theoretical. In domains where harm has been most concentrated, such as lending, hiring, healthcare triage, and policing, a reactive posture has already failed. Safety by design in these settings means:

• Impact assessments before deployment grounded in lived experience, not just technical metrics.

• Ongoing governance with real authority to pause or remediate.

• Feedback loops that close not only technical gaps but also the human accountability gap between providers and deployers.

Share

We cannot prevent every harm. But organizations can stop pretending that more review cycles, better intentions, or faster patches will fix a system whose incentives are misaligned with safety. The work is to move from asking how to catch harm to asking how to design so harm is less likely to occur in the first place.

I engage with this problem because I believe the most under-leveraged intervention is building the organizational infrastructure and conditions to address the compounding effects of historic, systemic, and cognitive bias and human fallibility on the design, development, and deployment of AI — specifically the people doing that work and the conditions in which they do it. The AI safety field names organizational risks as catastrophic. They are not being prioritized. That is the gap I am working to close.

CultureClusive works at the intersection of AI safety, organizational design, and structural transformation, and human design — in the space between fields that do not talk to each other, inside organizations that do not have the infrastructure to do this work. It is the most consequential place I know to stand.

If that intersection is where you work — in any of these fields, across them, or in the gap between them — I want to hear from you.

This is the founding signal.

Dasheika Rainney is the Founder and Principal at CultureClusive, a boutique executive advisory practice at the intersection of AI governance, organizational design, and structural transformation. cultureclusive.com

AI is not under-performing because of the technology. It is under-performing because it is being deployed into systems that cannot absorb it.

That is the structural constraint. Everything that follows is how it expresses, what happens if it continues, and the only viable correction.

Boards did exactly what they are wired to do. Protect and grow shareh

older value. Over the last several years, that meant moving money into AI quickly. Funding tools. Rolling out copilots. Issuing usage mandates tied to performance reviews on the assumption that more adoption would automatically push productivity, efficiency, and profit in the same direction.

The logic was coherent. The results are not.

In February, the National Bureau of Economic Research published findings from nearly 6,000 senior executives across the United States, United Kingdom, Germany, and Australia.¹ Nine out of ten report little to no measurable productivity or employment impact over the last three years. Those same executives still expect AI to lift firm productivity by 1.4% and output by 0.8% over the next three years.¹

That gap between what leaders are reporting and what they are projecting is not optimism. It is a signal that the systems-level conversation about whether the organization can actually absorb and realize what Digital is being asked to deliver is not happening at the depth it needs to.

The Structural Constraint

Every organization operates through two layers.

The Business layer is what gets built and funded: tools, job design, workflows, governance, policies, metrics, and formal training. This is what gets reported to the board. This is where the AI investment has gone.

The Cultural layer is what determines whether any of it produces results: trust, incentives, behavioral norms, psychological safety, leadership credibility, and the real capacity the workforce has to carry change. This layer is not built by deploying technology. It is built by sustained, deliberate investment in the conditions under which people work.

Most organizations have funded the Business layer heavily and assumed the Cultural layer would follow. It does not follow automatically. The gap between what is built and what can be absorbed is where AI value is currently disappearing. Not in the models. Not in the deployment. In the space between what the organization has constructed and what the workforce can realistically carry and is actually rewarded for doing.

The Cultural layer has two components that determine how wide that gap is.

Capability is what the workforce can do. Its inputs are skills, knowledge, judgment, context, and institutional memory. Capability is what training addresses. It is measurable and it is real. But it is not sufficient on its own.

Capacity is how much load the system can run through people without breaking. Its inputs are time, headcount, role clarity, and the accountability structure — the norms, incentives, and consequences that determine which behaviors are safe, expected, and rewarded. The accountability structure sits upstream of almost everything leaders say they want from culture. It sets incentives and consequences. Those shape norms. Norms drive behavior. From that behavior, people infer whether it is safe to tell the truth, raise a concern, challenge a decision, try something new, or admit uncertainty.

Trust, psychological safety, and judgment are not separate programs to launch. They are outputs of that chain.

Most organizations have invested in capability. Very few have invested in capacity. That is the structural constraint the NBER data is measuring.

We say human-centered. We overindexed on Digital. The NBER data is what that cost looks like at scale.

How the Constraint Expresses

This did not happen all at once. It compounded.

Since the generative AI wave accelerated in 2022, organizations have poured capital into Digital capability on the assumption that deployment equals transformation. It does not. PwC has been consistent on this: technology delivers roughly 20% of the value in a transformation. The other 80% comes from redesigning how work actually gets done.² Most organizations made the first move. Very few made the second.

HR has been structurally excluded from transformation conversations for years, confined to compliance and administrative functions, brought in after strategic decisions are already made. The function that owns the domain every technology investment has to move through was not at the table where the investment decisions were made. When AI transformation accelerated, that pattern held.

And the workplace itself stayed static. Siloed. Designed for stability. Trying to absorb transformation as a permanent condition rather than a bounded event. Research on digitalization and role overload shows this pattern consistently: when technology deployment advances without corresponding redesign of roles and operating conditions, stress and strain accumulate in the workforce rather than dissipating.³

For the last several years, Digital has received the investment and the strategic attention. The Human domain — the workforce, the behaviors, and the workplace conditions that determine whether any investment actually produces results — received the mandate and the expectation that it would adapt. That is the gap the NBER data is measuring.¹

Then the AI mandate landed on top of an already strained system.

Over the last eighteen months, most organizations have been operating under layered pressure that has not resolved. Job cuts spiked back to levels not seen since 2020. Leadership churn accelerated. AI is increasingly being cited as a driver in both. At the same time, the basic cost of staying employed went up: health benefits became more expensive for employers and employees alike, and real wage gains failed to keep pace for nearly half the workforce.⁴ ⁵ ⁶ CEO turnover surged and interim appointments became more common, sending a steady signal of instability from the very top.⁴

These are not isolated statistics. They describe a system in which pressure has been accumulating faster than it can resolve, and the workforce has been absorbing that load the entire time.

Each of these pressures has its own impact. Together, they hit harder than the sum of their parts. Sustained pressure compounds faster than the system can recover.

These are not background conditions. They are active forces hitting the same workforce at the same time. All work runs through the Human domain. Technology and operating models can be designed, deployed, and reported on. But nothing moves without the workforce’s capacity and willingness to carry it. When that capacity is depleted, the investment does not produce what it was designed to produce. It stalls.

Three of the most common patterns make this visible in practice. They are not the only ones.

When usage is rewarded and judgment is not, people perform the metric instead of exercising discernment. AI output moves through without review because review is not what the rubric measures. Activity climbs. Quality does not. AI is everywhere in the reporting and absent from the outcomes.

When people do not feel safe raising concerns, errors stay invisible. Dashboards show green. Conditions on the floor are red. The signals that should travel from the point of work to the point of decision do not travel. Leaders are making strategic decisions on data that does not reflect operational reality.

When mandates for AI come from leaders who use it rarely themselves, each subsequent directive lands with less credibility. The say-do gap is visible to everyone below it. Engagement falls. Discretionary effort falls. The workforce learns to perform compliance rather than practice judgment.

These are not culture problems to be addressed with engagement programs. They are structural signals that the operating conditions are hindering, not enabling, the workforce.

Leadership capacity to carry this is also thinning. Research from Torch’s 2025 Leadership Evolution study found that managers are increasingly caught between organizational demands and workforce needs, with many reporting they lack the support and clarity to lead through continuous change effectively.⁷ When leadership is stretched, the signals from the workforce have nowhere to go.

What I see when I walk into organizations running this pattern is recognizable.

You ask a question and the room is silent. Everyone waits for the most senior person to speak, then agrees. Issues do not surface internally. They appear later as customer-facing incidents, compliance gaps, or errors that made it to market because nobody felt safe raising them before deployment. Deadlines are missed. Work product quality declines. Usage metrics climb and training completion rates look healthy while workarounds multiply quietly underneath. Goals, OKRs, and actual work being done do not connect consistently across levels and layers of the organization.

That is not resistance. That is capacity depletion. And it is what the system looks like when the pace of digital investment has been outrunning the pace at which the organization can actually absorb it for too long.

What Happens If It Continues

The default response to underperformance is already forming. Better models. More pilots. Stricter mandates. More training. Some leaders are quietly waiting for the next model release to close the gap and lower the labor cost permanently.

None of it addresses the constraint. All of it adds load to a system that is already running near its limit.

Research consistently shows this does not hold. Gartner finds that fewer than half of change initiatives succeed, and change fatigue is now one of the primary reasons transformation efforts stall.⁸ Leaders believe they can set a vision, issue a mandate, run change management and training, and the culture will move accordingly. The entire system cannot transform on a directive, especially without ongoing support for absorption and recalibration over time. Adding capability to a system with insufficient capacity does not produce results. It produces more compliance behavior, more signal failure, and more invisible accumulation of risk.

My read is that organizations carrying this load without addressing the capacity constraint are accumulating compliance exposure they cannot currently see.

The governance infrastructure is thin.

• Fewer than one in five organizations have enterprise-wide AI accountability in place ⁹

• One in five cannot describe how humans oversee AI ¹⁰

• More than half of HR professionals in states with active AI laws are unaware those laws exist ¹¹

Thin governance infrastructure combined with a workforce that has learned it is not safe to surface problems is not a theoretical risk. It is the conditions under which harm reaches the market before anyone with authority catches it. At some point the exposure surfaces — in customer-facing failures, in regulatory action, in liability that traces back to operating conditions that were visible and unaddressed.

That is not a forecast. It is a read on where the compounding leads when the constraint goes unaddressed.

The Only Viable Correction

The technology is not the problem. It is becoming one when capability is planned in isolation from the rest of the organization, and when the goal set for it is unclear and undefined. Nobody has decided whether AI is meant to handle defined parts of the work or to sharpen the judgment of the person doing it. Those are two different deployments. Most organizations are running one while measuring the other.

If the next three to five years are going to look different from the last three, the answer is not more of the same AI, faster. It is to address the system AI has landed on.

If a CEO called me right now I would tell them this directly. The NBER data is not telling you the technology failed. It is telling you the system underneath it was not ready to carry it. And that system is your workforce — the people who have absorbed eighteen months of compounding pressure and are now being asked to carry an AI transformation on top of it.

The next bet is the Human domain. Not as an HR initiative. As a strategic correction of the constraint.

Get an honest baseline on workforce capability. Not headcount. Not org charts. Whether the people you have today can actually execute the strategy you are describing for the next three years, and where the institutional knowledge that used to carry your operations has gone.

Get a real read on workforce sentiment and the operating conditions producing it. What the culture is actually rewarding right now — compliance, candor, or adaptation. Whether your accountability structures are aligned to the behaviors AI-enabled work actually requires or to the behaviors that made sense in a previous operating model.

Rebuild the workplace conditions that restore capacity. Decision rights that match the speed the business is now operating at. Governance that reflects what the work actually requires. Feedback loops that function. Escalation paths that people actually use because using them is safe.

Then recalibrate the pace of Digital investment to what the organization can actually carry. The pace of digital investment has to match the pace at which the organization can absorb it. Closing that gap is not a soft recommendation. It is a realization requirement.

The Floor the Business Has to Hold

There is something underneath this I want to say plainly.

If a workplace operating under shareholder primacy expects maximum performance, productivity, and output from the workforce, then that workforce can reasonably expect — and the business must account for — whether its culture and operating conditions actually support their ability to deliver that. You can only get as much from people as you are willing to enable and empower them to give.

That means there is a floor. A minimum set of operating conditions the business has to provide to maintain the capability and capacity of its workforce. Not as a social responsibility initiative. Not as an engagement program. As the bare minimum required for the system to function at the level the business is demanding.

Right now, for many organizations, that floor has not been defined. The social contract between employer and employee has been operating below any reasonable baseline of what people need to carry what is being asked of them. The AI mandate arrived on top of that. The productivity gap is, in part, the cost of that deficit showing up in the numbers.

The organizations that figure this out will not just produce more from their AI investments. They will be the ones that establish what it actually looks like to build a workforce that can absorb continuous transformation without fracturing. That is a competitive advantage. It is also, at minimum, what the logic of the system requires.

Rebalance the System

The business is listing because the load is uneven. Digital raced ahead. The Human and Operational system stayed where they were.

To right it, leaders have to start with the workforce. Not as a gesture toward culture. As a recognition that all work runs through the Human domain and the Human domain has been running near its limit for eighteen months while being asked to carry more.

The organizations that produce more value over the next three years will not be the ones that deploy the most AI. They will be the ones that increase their capacity to absorb, adapt, and execute under it. That requires a genuine shift — away from treating Digital capability as the primary lever and toward treating the operating conditions of the workforce as the constraint that determines whether any capability produces results.

The productivity gap is not a technology problem waiting for a better model. It is a capacity problem the next deployment will make worse.

Execution integrity is not a project to complete. It is a state to maintain. Right now, maintaining it starts with putting the workforce back at the center of the strategy and letting Digital and Operational design follow.

If This Is Landing for You

If you are a CEO, CHRO, or transformation leader and you suspect your AI investments are quietly landing on an exhausted system, the next step is a direct read on where this is sitting inside your organization.

The Signal Scan is a 20-minute conversation. It is designed to give you a clear picture of whether your current operating conditions are supporting or hindering what you are asking your workforce to carry. No deck. No proposal. A direct read and a clear next step.

Book directly here.

Sources and further reading

This article draws primarily on the National Bureau of Economic Research working papers on firm AI adoption and the modern productivity paradox, Fortune’s April 2026 coverage of those findings, Perceptyx’s 2026 employee experience analysis, Torch’s 2025 Leadership Evolution research, Harvard Business Review’s reporting on Gartner’s change data, academic research on digitalization, role overload, and job stress, and broader current commentary on change fatigue, operating conditions, and AI adoption failure.

STRUCTURAL THINKING · CULTURECLUSIVE

——

Something is happening inside your organization right now.

A reorg redistributed accountability — but not the context behind it.

Layoffs removed institutional knowledge alongside the people who held it.

A new leader stepped in without the rationale behind what was already in motion.

An AI deployment closed on time, on budget and is producing none of what it promised.

Maybe it’s one of these. Maybe it’s all of them. At once.

What makes this hard to navigate is that everything still looks like it’s working. Work is moving. Teams are active. Programs are progressing.

And still something isn’t converting.

So the response follows what’s visible.

Adoption slows → train harder.

Usage drops → improve the tool.

Outcomes lag → increase accountability.

Each move is reasonable. None of them close the gap.

Because the problem isn’t sitting where it shows up.

A stalled rollout looks like a people problem.

A clunky tool looks like a technology problem.

Missed outcomes look like an execution problem.

But those are expressions — not origins.

What’s actually happening is quieter.

The technology changed. The way work moves didn’t.

The expectations changed. The incentives didn’t.

Ownership shifted. Decision rights didn’t.

So the organization starts operating on two versions of reality at once. That’s when execution begins to drift.

Not because something failed but because the system can’t absorb what’s changing around it.

You can see it in how the same question gets answered differently depending on where you ask it.

Did it work?

The product team says yes — it shipped, it performs.

The program says yes — it delivered, it closed.

Leadership is still waiting to see it in the numbers.

All three are right. They’re just measuring different things.

Most deployments are designed to complete. Very few are designed to carry what was deployed.

In a stable environment, that gap is manageable. In an organization under continuous pressure (reorgs, leadership churn, knowledge loss, accelerating digital change) it compounds.

This is where I start.

Not with dashboards or activity reports. With three questions — asked plainly, in the room where the work is actually happening. Each one targets a different layer of the organization. Each answer tells you immediately where attention needs to go first.

Question 1

Where is the system allowed to act — and where is human judgment required?

If someone can answer this clearly — name the boundary, say who owns it, describe what happens when the organization changes around it — people know what they’re supposed to do. Decisions move. Risk stays visible.

If the answer is a policy document, a committee, or a long pause — people are hesitating at exactly the moment the organization needs them to move. Adoption stalls. Risk builds quietly.

After a reorg, this boundary shifts. After layoffs, the person who owned it may be gone. If no one has redrawn it, the gap is already open.

Question 2

If something went wrong today, what would happen in the next four hours?

If the answer is immediate — a specific person, a clear path, a designed response — the accountability structure is real.

If the answer depends on who notices, who has capacity, or who picks it up — it exists on paper. Every incident that gets improvised is proof that the structure wasn’t designed to hold what the organization is running on it.

Leadership churn makes this worse. The person who knew the escalation path may no longer be in the role. If no one has updated it, every incident starts from zero.

Question 3

What changed about how people are measured after this was deployed?

If the answer is specific — adjusted expectations, redesigned workload, new signals that reflect what the organization now requires — behavior will follow.

If nothing changed — the organization is asking people to work differently inside a structure still designed to reward the old way.

That isn’t resistance. It’s a rational response to an unchanged environment.

Knowledge loss makes this the most expensive gap of the three. When the people who understood the original design leave, what remains is the tool — without the judgment that made it safe to use.

This is the first piece in the Structural Thinking series. Each one is a standalone read on the conditions that determine whether organizations convert strategy into sustained value. Subscribe to follow the series →

Subscribe now

The pattern across those three answers tells you more than any dashboard metric.

Clear on the first.

Improvised on the second.

Silent on the third.

That’s not three separate problems. That’s one organization that was designed to complete and was never designed to carry what it deployed into the world.

The silence is always where to start.

It’s where the most load has accumulated without anyone owning it. It’s where behavior stopped matching what the system was designed to produce. And it’s where a single structural move: one boundary redrawn, one escalation path named, one accountability made explicit will produce the most immediate change in how the system operates.

The work isn’t to fix each symptom individually. It’s to find where the organization fell out of sync and intervene at the point where a change will actually hold.

You don’t need to map everything before you act. You need to know which part of the organization is carrying pressure it was never designed to absorb and where to direct attention first.

Not sure which of the three gaps is carrying the most load in your organization? The Execution Diagnostic surfaces the structural signals in about two minutes no email required.

If this is where your organization is — the Signal Scan is the right starting point.

Twenty minutes. You describe one situation where outcomes aren’t matching expectations. I’ll tell you what I’m seeing and where I think the problem is sitting.

No pitch. No deck.

Book a Signal Scan → Read the Executive Brief first →

The next piece in Structural Thinking continues this. Subscribe to follow the series.

Execution integrity is not a one-time design. It is a state to maintain.

Dasheika Rainney is the founder of CultureClusive. Structural Thinking is a series of standalone pieces on the conditions that determine whether complex organizations convert strategy into sustained value — regardless of where you are in the transformation.