AI governance is having its sensible-adult moment.
Policies are being written. Committees are being formed. Approved tools are being listed. Risk registers are filling up. Somewhere, someone is almost certainly adding a new column to a spreadsheet called “Human Oversight Required?” and feeling, with some justification, that progress has been made.
Progress has been made.
After the first rush of experimentation, organizations are getting much better at asking the questions they should have been asking all along: What data can this system use? What decisions can it make? Who approves the use case? What happens when it fails? Who owns the outcome? Where does human review belong?
Those are essential governance questions. They are also only part of the problem.
Because once the policy is written and the controls are designed, the organization still has to operate them. People have to make decisions inside the boundaries. Managers have to supervise AI-supported work. Employees have to know when a strange result is merely strange and when it is dangerous. Someone has to decide when an agent has crossed the line between useful initiative and the digital equivalent of “I thought that was what you meant.”
That is where the conversation starts to move from AI governance toward Human-AI Workforce Resilience.
The next enterprise challenge is not simply governing AI. It is building a workforce capable of operating safely, effectively and accountably as AI becomes a normal part of how work gets done.
Governance can define the boundary. It cannot make every decision inside it.
There is a comforting fantasy buried inside a lot of governance work: if we write sufficiently good rules, sufficiently clearly, people will encounter the appropriate rule at the appropriate moment and do the appropriate thing.
Anyone who has ever written a policy has probably recovered from this fantasy.
Real work is messy. Context changes. Two perfectly reasonable requirements conflict. An AI system produces an answer nobody anticipated. A low-risk experiment suddenly becomes business-critical because everyone starts using it. A manager discovers a faster way to get something done and, entirely rationally, keeps doing it until somebody notices that the workflow now carries more authority than anyone intended.
Governance matters because it establishes boundaries and decision rights. It tells people what they may do, what they must not do, which situations demand review and who owns which parts of the system.
But policies eventually reach the edge of what they can specify, and at that point, someone has to exercise judgment.
That judgment might be as simple as deciding not to paste sensitive information into a public model. It might involve questioning an AI-supported recommendation because the conclusion looks plausible but the assumptions do not. It might mean stopping an autonomous workflow because the system is technically operating inside its permissions while producing a result nobody actually wants.
The policy can help define the perimeter, but it cannot stand beside every employee whispering, “This is the bit where you should probably stop.”
The human role is getting harder, not easier
There was an early version of the AI productivity story in which technology handled the boring work and people were freed to concentrate on higher-value thinking.
And true, that is still one of the great opportunities.
There is just a small complication: higher-value thinking is usually harder, from a cognitive perspective.
McKinsey’s recent work on the judgment bottleneck captures part of the problem. AI can generate analysis, scenarios and recommendations far faster than organizations can absorb them. The scarce resource starts to move away from production and toward interpretation, judgment and decision-making.
A person supervising AI-enabled work may need less time to create the first answer and considerably more skill in deciding whether the answer is worth anything. They need enough domain knowledge to spot what is missing, enough AI literacy to understand what the technology can and cannot reliably do, enough risk awareness to see the consequences beyond the immediate task, and enough organizational context to know when the problem belongs to somebody else.
This is not quite the simple world promised by “AI will handle the repetitive stuff.”
(The repetitive stuff may have been doing more work than we realized.)
It taught people how systems behaved. It exposed them to weird exceptions. It gave junior employees a few hundred low-stakes opportunities to be wrong before someone gave them authority over something expensive.
As AI absorbs more of that work, organizations will have to become much more deliberate about how those capabilities get built. Otherwise we may produce a generation of beautifully productive employees who can generate excellent work and have remarkably little idea what to do when the machine stops being excellent.
Resilience is not another word for training
The most obvious response to a workforce capability gap is to train people, and training will absolutely be part of the answer. The risk is assuming that it can carry the whole problem simply because it is the most familiar lever organizations know how to pull.
A resilient AI-enabled workforce is not one where everybody has completed the “Responsible AI 2027” course and remembers not to upload the customer database into whatever shiny model launched that week. Resilience is broader than that because the difficult moments rarely arrive as cleanly packaged knowledge questions.
What matters is what people do when the answer is ambiguous: whether they recognize risk, whether they trust AI too much or too little, whether they understand their own authority, whether they feel safe challenging a recommendation, whether escalation works, whether leaders reward good intervention and whether the organization learns when something goes wrong.
We think of Human Risk Management this way already because human-related risk is rarely caused by a single missing piece of knowledge.
An employee may understand the policy perfectly and still behave in a risky way because the process is unrealistic, the deadline is absurd, the authority is unclear, the team norm points in another direction or the quickest way to get the job done happens to bypass the control. AI tends to amplify those same organizational dynamics while introducing new ones around reliance, delegated authority, invisible work, supervision and the speed at which a small judgment can become a consequential action.
AI literacy gets you through the front door. The harder part comes after.
AI literacy matters, but the term is beginning to carry more weight than it can reasonably bear.
Most organizations now understand that employees need some basic fluency: what the tools are good at, where they fail, what they should never be given, why confident output is not the same thing as reliable output, and when verification matters. That is useful groundwork. It helps people use the technology without treating it as either magic or radioactive waste.
The harder questions appear once AI is embedded in actual work.
A finance analyst may understand perfectly well that an AI-generated forecast could be wrong. The more difficult judgment is whether the forecast is wrong in a way that matters, whether the assumptions have drifted, whether the recommendation is still appropriate in the current business context, and whether the analyst has enough evidence to challenge it.
A manager reviewing AI-assisted hiring recommendations faces a similar problem. Knowing that models can be biased is literacy. Recognizing that this particular recommendation is leaning too heavily on the wrong signals, understanding the consequence, and deciding whether to intervene requires something richer.
This is where competence, judgment and authority start to intertwine.
People need enough knowledge of the technology to avoid naïve trust, enough knowledge of the work to know when something looks off, and enough understanding of consequence to know when “off” becomes important. They also need to know what they are allowed to do about it. (That last part is easy to underestimate.)
There is not much point teaching an employee to recognize risk if the organization has given them no credible way to respond when they see it.
Culture has a habit of showing up wherever the policy stops
We have spent years in cybersecurity learning that people do not behave inside neat boxes marked “policy,” “training” and “control.” They behave inside teams, hierarchies, deadlines, incentives and all the ordinary social machinery of work.
AI will not be exempt.
Imagine a junior employee who spots a problem in an AI-supported recommendation that happens to align beautifully with what the senior leader in the room already wants to do. The employee may understand the system, recognize the risk and know the escalation route. None of that tells us whether they will actually challenge the decision.
That depends on what the organization has taught them through experience.
Has somebody been rewarded for stopping something before? Does the manager genuinely welcome challenge, or merely say that they do? Is raising a concern treated as part of doing good work, or as the thing you do when you enjoy being invited to fewer meetings?
The same dynamic matters at scale. A workforce that learns that speed is celebrated and friction is punished will eventually find efficient ways to make inconvenient controls disappear into the background.
That is why psychological safety belongs in the AI governance conversation. So do leadership behavior, group norms, confidence, perceived control and the small signals organizations send about what happens when someone says, “I’m not comfortable approving this yet.”
Those factors may sound softer than permissions, model inventories and technical controls. They are considerably less soft when the governance model depends on someone speaking up at exactly the right moment.
Agentic AI turns the authority question up several notches
The move toward agents makes this more interesting because AI is no longer confined to producing something for a human to inspect. Increasingly, systems are being given permission to act.
They can call other systems, modify information, communicate, initiate workflows and pursue objectives within whatever authority an organization has delegated to them.
At that point, the important unit of analysis is not simply the tool. It is the relationship between the agent, the human supervising it, the authority available to each and the consequences that can follow.
Enterprises already understand delegation. Managers delegate work every day. What makes AI agents unusual is that the delegated actor may operate continuously, at machine speed, across systems, without the ordinary social cues that tell a human employee, “This situation is getting weird.”
That puts much more weight on the surrounding human design.
Someone has to decide what the agent is allowed to do, how much autonomy is appropriate, what requires confirmation, what counts as an exception and what happens when the system behaves correctly according to its instructions but produces an outcome nobody intended.
Saying “the human remains accountable” does not solve much by itself. Accountability becomes meaningful only when the human can see enough of what is happening, understands enough to challenge it and retains enough authority to change course.
Without those conditions, accountability is mostly a name written next to a problem after the fact.
This is why we have been doing so much work around human-agent accountability. As agents gain capability, organizations need a much sharper understanding of who is supervising whom, where decision rights sit and whether the people carrying those responsibilities are equipped for them.
Human-AI Workforce Resilience is really about the system around the person
This brings us to what we mean by Human-AI Workforce Resilience.
For us, it is the organizational ability to keep people effective, appropriately skeptical, accountable and capable of adapting as AI changes the work around them. That includes recognizing risk, exercising judgment, challenging AI where necessary, understanding authority, escalating uncertainty, recovering from mistakes and learning as new patterns emerge.
The important word there is organizational.
There is a tendency in workforce programs to push responsibility downhill. Teach the employee. Remind the employee. Test the employee. Send another communication when the employee does not behave exactly as planned.
Sometimes that is appropriate. Sometimes the employee is simply operating inside a badly designed system.
If a person is expected to challenge AI but has no time to inspect the output properly, the problem is not literacy. If they spot the issue but cannot interrupt the workflow, the problem is not awareness. If they raise it and are repeatedly told to stop slowing everyone down, another training module is unlikely to produce the miracle.
A resilient organization pays attention to the conditions around the person as much as the person themselves.
That means role design, leadership, decision rights, workload, incentives, escalation, technical controls and access to expertise all belong somewhere in the picture.
Measurement becomes much more useful when it tells you why the control is weak
This is where a mature Human Risk Management program becomes extremely useful. If organizations are going to rely on people as part of the AI control environment, they need better evidence than “97% completed the course.”
Completion tells you something. It does not tell you whether somebody can spot an inappropriate recommendation, whether they will challenge it, whether they know when to escalate, or whether their manager has created conditions in which any of those behaviors are realistic.
We separate competency, behavior, psychology, culture and organizational conditions in our own Human Risk Management work precisely because two people can exhibit the same risky behavior for entirely different reasons.
One employee may over-rely on AI because they do not understand its limitations. Another may understand the limitations perfectly but trust the system because it has been right for six months. A third may be uncomfortable with the recommendation but approve it because the senior person in the meeting clearly wants to move on.
Those are three different problems, so sending all three people the same AI awareness course would be wonderfully efficient administratively and not especially useful for a risk mitigation effort.
The interesting measurement question is therefore not simply whether people are using AI safely. It is what is shaping the behavior, where the organization is exposed and which intervention has a reasonable chance of changing the outcome.
Secure AI Enablement should make AI easier to use well
There is an important trap here.
As organizations discover more AI risk, the reflex can be to add controls until using the technology becomes so awkward that employees either stop using it or quietly find another way.
We have seen this movie before in cybersecurity. It does not improve with age.
The aim of Secure AI Enablement should be to create enough confidence in the system around AI that organizations can use the technology more ambitiously, not less.
Governance plays one part by setting boundaries and decision rights. Security provides technical protections. Human Risk Management helps ensure that people understand the work, can exercise judgment and operate inside a culture that supports the behaviors the controls depend on.
When that works, organizations should actually be able to give AI more freedom in the places where it is appropriate because they understand what is happening around it.
Where it does not work, organizations tend to compensate with blanket restrictions and approval queues. Everyone gets frustrated, adoption moves sideways, and somebody eventually discovers that half the workforce has found a more convenient tool anyway.
Secure enablement is therefore partly about earning the confidence to say yes.
Sometimes the correct governance decision will be to constrain a use case. Sometimes it will be to expand it. A resilient organization should be able to tell the difference.
The workforce is not one AI risk category
One of the most useful things leaders can do now is stop talking about “the workforce” as though everyone has been handed the same AI problem.
The risk is not evenly distributed because the work is not evenly distributed.
Someone using AI to summarize meeting notes occupies a very different position from a finance leader relying on AI-supported forecasts, a developer supervising an agent with production access, an HR professional making decisions about candidates, or an executive using AI to pressure-test a major strategic decision.
The relevant capability has to follow the work.
That means identifying consequential AI-enabled activities, understanding where human judgment still matters, examining the authority attached to the role and looking carefully at what happens if the decision goes wrong.
From there, organizations can ask much more useful questions. Does this person have enough subject-matter expertise to challenge the system? Have they seen enough exceptions to know what abnormal looks like? Are they actually expected to intervene, or merely available for ceremonial approval? What evidence would tell us that this arrangement is working?
Those answers should feed back into AI governance.
If a workforce is not yet capable of supervising a particular level of autonomy, that is information about how the use case should be designed. If a team demonstrates strong capability, good escalation and reliable oversight, that may support greater autonomy.
Governance should evolve with the evidence rather than sitting above it as a static set of instructions.
What we are really trying to preserve
None of this is an argument for keeping a human attached to every machine action forever. There will be plenty of work where AI should simply do the job, and organizations should be delighted when it can.
Nor should the future involve thousands of well-paid people ceremonially clicking “approve” on outputs they barely looked at so the organization can say human oversight was present. That is not resilience. It is governance theatre with better software. The hardest question right now is defining exactly where human judgment genuinely matters and what the organization needs to preserve around it.
Sometimes that will be expertise. Sometimes it will be skepticism. Sometimes it will be the confidence to say that an answer which looks perfectly reasonable is inappropriate in this particular context. Sometimes it will simply be knowing when the machine has wandered into territory where somebody else needs to take over.
That is the capability we mean when we talk about Human-AI Workforce Resilience.
The organizations that get this right will not be the ones that force humans to keep pace with machines. They will be the ones that become very deliberate about where human judgment still matters, build the conditions that allow it to work, and use that confidence to let AI do more everywhere else.
