What roles should humans retain in an AI-enabled organization?

Assign responsibilities within the work

“Which jobs will AI replace?” encourages a false choice between people and software. Jobs are bundles of tasks: some are repetitive, some involve retrieval or classification, and some require deciding what should be done when the evidence is incomplete and the stakes differ among people. Redesigning a job task by task is more useful than assigning a whole occupation to “human” or “AI.”

An AI-enabled organization therefore needs neither a rule that a person must touch every output nor a rule that speed justifies autonomous action. It needs an explicit operating model for each workflow. The model should identify the outcome being pursued, the allowed inputs and actions, the people affected, the likely failures, the decision owner, the escalation path, and the evidence used to evaluate results.

This is also the realistic reading of labor evidence. The International Labour Organization’s 2025 update finds that one in four workers globally are in occupations with some degree of generative-AI exposure, but says most jobs are more likely to be transformed than made redundant because they still require human input. It recommends managing that transition through social dialogue. ILO, Generative AI and jobs: A 2025 update The implication for leaders is not to preserve every old task. It is to preserve and strengthen the capabilities that let a team set direction, notice harm, correct a system, and adapt the work as conditions change.

The human roles worth designing for

Set goals and define what good means

AI can optimize a target or propose options, but it cannot legitimately choose the organization’s purpose, resolve a conflict among objectives, or decide what tradeoff is acceptable. A customer-support system can be asked to reduce response time, improve resolution, protect margins, and treat customers fairly. Those goals may conflict. “Close tickets quickly” can produce premature closures. “Maximize refunds avoided” can make a support experience feel adversarial. “Minimize fraud” can burden legitimate customers.

Leaders and accountable domain owners must therefore define objectives and constraints in terms that can be tested. Specify what the system may do, what it may recommend, what it must never do, which outcomes are unacceptable even if an aggregate metric improves, and when a human must decide. NIST’s framework places this work in the context-setting stage: intended purposes, uses, norms, requirements, benefits, costs, and risk tolerance need to be understood and documented before a system is deployed. NIST AI RMF Core

This role includes deciding whether AI is appropriate at all. A stable, low-harm task with a clear success measure may justify extensive automation. A task that sets policy, allocates a scarce opportunity, affects someone’s livelihood, or involves contested values needs much more than a model score. The human contribution begins before a prompt or model is chosen.

Exercise judgment under uncertainty

Judgment is not an unexplained gut feeling. It is the disciplined ability to weigh incomplete evidence, recognize when a case lies outside normal conditions, seek missing context, and make a reasoned decision that can be defended. AI may be valuable evidence in that process, but a confidence score or fluent explanation is not a complete basis for a decision.

Keep humans in charge of cases where the rule is ambiguous, information is unreliable, the decision changes a person’s rights or opportunities, a rare event has large consequences, or reasonable people can disagree about the appropriate outcome. Give them access to the source information and relevant policy, not only an AI summary. They need a usable way to record why they accepted, changed, or rejected a recommendation.

This does not imply that humans are flawless. Human decisions can be inconsistent and biased too. The organizational response is not blind faith in either side. It is to compare human-only, AI-only, and combined performance against the outcome that actually matters, inspect who bears errors, and revise the workflow when the evidence changes. NIST notes that human-AI interaction can, depending on the task and configuration, amplify bias or produce complementary performance. NIST on human-AI interaction

Own accountability and the right to say no

An AI system cannot bear organizational, professional, legal, or moral accountability. A vendor cannot absorb the organization’s duty to customers, employees, patients, residents, or shareholders merely by supplying the tool. Name a senior business owner for each consequential use case, alongside a technical owner, risk or compliance partner where appropriate, and an operational owner who sees the daily failures.

The owner must be empowered to change the workflow, pause it, allocate remediation resources, and report tradeoffs upward. The owner should not be the person whose compensation depends solely on volume or cost reduction. Otherwise the organization creates a predictable incentive to rationalize harmful outputs rather than investigate them. NIST explicitly assigns executive leadership responsibility for decisions about risks associated with AI systems and calls for roles and communication lines to be documented. NIST AI RMF governance outcomes

For systems covered by the European Union’s high-risk AI rules, specific human-oversight requirements apply. Article 14 requires, in its applicable scope, that assigned natural persons be able to understand relevant capabilities and limitations, interpret output, disregard or reverse it, and intervene or stop the system. Regulation (EU) 2024/1689, Article 14 This is not general legal advice. Organizations should obtain legal guidance for their jurisdictions and use cases. The operational principle is widely useful even when the law does not apply: authority without the ability to intervene is not accountability.

Handle exceptions and restore service

Automation is strongest in its expected operating range. People should retain responsibility for the edges: unusual cases, conflicting evidence, system degradation, customer hardship, security incidents, and policy conflicts. A well-designed exception path does not treat an escalation as an embarrassing failure. It treats it as a source of operational knowledge.

Make exception handling concrete. Define triggers for escalation, maximum time before a person responds, which actions are reversible, what information an operator receives, and which cases require a specialist rather than a general queue. Log the reason for each escalation and the final disposition. Recurrent exceptions may show a policy gap, poor data, a model limitation, or a workflow that was automated too aggressively.

Humans are also essential when the organization needs to recover from a failure. Someone must recognize that observed performance no longer matches the expected range, coordinate a pause or rollback, communicate honestly with affected people, and decide what evidence is sufficient to resume. NIST recommends continuous risk management, testing before deployment and regularly in operation, plus monitoring and documented response to incidents. NIST AI RMF measurement and management guidance

Supervise systems rather than merely review outputs

Supervision means checking whether the system remains fit for its purpose. It includes sampling outputs, tracking error patterns and drift, testing for known failure modes, reviewing changes in inputs and surrounding processes, and comparing outcomes across affected groups where relevant. It also includes knowing when automated activity has become too fast or too broad for a person to control.

Output review is one tool within supervision, not the whole job. A human can catch an occasional bad answer while missing a systematic problem in who gets routed, prioritized, priced, hired, or monitored. Monitor the decision process and downstream outcomes, not only model accuracy. For example, an accurate support-routing model may still be unacceptable if it makes it difficult for people with complex cases to reach a person.

Independent review can be valuable for a system that generates revenue, affects people materially, or creates a conflict for its business owner. NIST says independent review can improve testing and mitigate internal bias or conflicts of interest. NIST AI RMF Core The intensity of supervision should rise with the scale, autonomy, reversibility, and impact of the action.

Maintain relationships and legitimacy

Relationships affect how well the work gets done. Customers disclose context when they believe they will be heard. Employees report a broken process when doing so is safe. Partners cooperate when commitments are credible. Managers resolve conflict by understanding history, incentives, and concerns that may not appear in structured data.

AI can summarize a customer history, draft a communication, translate a document, or surface patterns for a manager. It should not be used to pretend a person reviewed or empathized with a message when no one did. Make the mode of interaction clear, provide a route to a person where the decision or relationship warrants it, and ensure human staff have discretion to repair harm rather than merely repeat an automated policy.

This matters inside the organization as well. Algorithmic management can improve consistency or manager efficiency, but it can also make accountability unclear and the tool’s logic hard to follow. These are concerns reported by managers in an OECD survey covering more than 6,000 firms in six countries. OECD, Algorithmic management in the workplace Human managers should spend less time compiling routine status reports and more time coaching, resolving tradeoffs, developing people, and challenging an automated assessment that does not fit the individual’s real work.

Learn the domain and preserve the ability to operate without automation

An organization that delegates every first draft, diagnosis, or decision to AI can lose the expertise needed to evaluate it. This is a capacity risk, not nostalgia for manual work. If no one understands how claims are evaluated, contracts are interpreted, code is debugged, or a plant is operated, the organization cannot audit an output, recognize a novel failure, negotiate with a vendor, or recover during an outage.

Preserve domain learning through paired work, rotation, case review, simulations, and periodic unaided exercises. Teach operators the system’s purpose, inputs, failure modes, uncertainty, and escalation routes, not only the user interface. NIST frames operator and practitioner proficiency as a process that should be defined, assessed, and documented. NIST AI RMF Core

The appropriate reserve capacity depends on risk. A low-impact marketing draft does not require a full manual fallback. A high-consequence service should have named people, workable procedures, and enough practice to continue safely if the model, provider, data feed, or integration fails. Design the human fallback before turning on autonomy.

Give workers a real voice in redesign

Workers see the messy inputs, workarounds, customer harm, and hidden dependencies that a process map misses. Involving them early improves system design and gives people a credible way to challenge metrics or automation that distort the work. Consultation is not an announcement after tools and staffing decisions are final. It is a forum with enough information, time, representation, and influence to change the design.

Invite affected workers and their representatives, where applicable, into task mapping, pilot design, safety and privacy review, training plans, metric selection, and post-launch evaluation. Ask what work is invisible to the current data, which errors are costly, who will carry the exceptions, which parts of the job teach the next generation, and what new monitoring feels unsafe or unfair. Publish what changed as a result of the input and explain any disagreement.

Evidence supports treating this as operational work, not public relations. An OECD experiment in three German manufacturing firms found that consultations among workers, managers, and works-council representatives could produce technology designs participants considered to preserve productivity gains while improving job quality. The authors call for broader research, so this is encouraging evidence rather than a universal guarantee. OECD worker-consultation experiment The ILO likewise calls for social dialogue to help manage generative-AI transition in ways that support productivity and working conditions. ILO 2025 update

A usable division of work

The following pattern applies to many knowledge workflows. It is a starting point, not a template to copy into every context.

The table is not a claim that people must hand-review every recommendation. It identifies the decisions that must have an accountable human owner. A workflow can be highly automated at the intake stage while remaining firmly human-led at policy, exception, and consequence boundaries.

Example: AI-assisted customer refund triage

Consider an online retailer that receives thousands of support requests each week. The business wants faster responses without giving an AI agent unrestricted authority to issue refunds, expose customer data, or create policy through its behavior.

The human team sets the charter. A support leader, finance partner, security or privacy partner, and experienced agents define the purpose: resolve routine, low-value requests quickly while protecting customers and the business. They set limits on refund amount, action types, customer data exposure, languages or products that require specialist review, and indicators of harm. They define a genuine customer appeal path and nominate an accountable operational owner.

The AI handles bounded work. It reads an approved subset of account and order data, identifies the request type, drafts a response, proposes a resolution under fixed policy, and either completes a reversible low-value action or routes the case. It cannot alter policy, issue exceptions above the set threshold, contact a customer outside approved channels, access unrelated records, or make changes to payment settings.

Humans decide the meaningful exceptions. A trained agent handles disputed delivery, accessibility needs, suspected fraud, high-value accounts, repeat complaints, conflicting evidence, and any request outside the policy. The agent can overrule the recommendation, record the reason, and make a goodwill decision within their authority. A supervisor handles policy conflicts and trends, not just the largest queue.

Supervision checks the system, not just the queue. The team samples completed cases, measures recontacts and appeals, examines whether some customer groups or request types are being disadvantaged, monitors error and escalation rates, and reads agent feedback. A sharp increase in “routine” refunds, bad routing, appeals, or agent overrides triggers review. The operational owner can narrow the boundary or turn off the automated action while the cause is investigated.

This is a division of work, not “human in the loop” theater. The human role has actual decision rights at the beginning, exception boundary, and control layer. AI provides speed within limits, but it does not decide what the organization owes a customer when policy and context collide.

Avoid ceremonial human oversight

Ceremonial oversight is a person nominally assigned to review a system, but unable to make a considered or effective intervention. It often appears as a high-volume approval queue, a dashboard no one has time to inspect, a generic warning that output “may be wrong,” or a reviewer who sees only the model’s conclusion and is measured on throughput.

It fails because it transfers blame without transferring control. The reviewer may have seconds for a complex decision, lack the necessary domain knowledge, lack access to underlying data, fear being penalized for overriding, or have no authority to stop the system. In those conditions, asking someone to click “approve” creates automation bias and a false audit trail rather than independent judgment. Article 14 of the EU AI Act explicitly notes the risk of over-reliance and requires, for covered high-risk systems, capabilities to understand, interpret, disregard or override, and stop the system. EU AI Act Article 14

Use the following five tests before describing a control as human oversight:

1. Decision right: Can the person reject, change, pause, or escalate the action, and will the system obey?
2. Competence: Has the person been trained in the domain, the tool’s limits, known failure modes, and the escalation policy?
3. Evidence: Can the person see the relevant source information, uncertainty, and reasons needed to reach an independent view?
4. Capacity: Does the person have enough time and a manageable enough queue to inspect the case rather than rubber-stamp it?
5. Incentive and learning: Is the person safe to disagree, and are overrides, errors, appeals, and incidents analyzed to improve the system?

If the answer to any test is no, change the workflow. Reduce the automation boundary, automate a more routine portion, add a second review or domain specialist, improve the interface and evidence, slow the system down, or remove the claim of human oversight. NIST recommends defining, assessing, and documenting oversight processes and collecting data about the frequency and rationale for human overrides. NIST human-AI interaction appendix

Match human involvement to risk and autonomy

There is no sensible universal rule such as “all AI must be reviewed by a person” or “AI should work autonomously whenever it is accurate.” Choose a mode based on the action’s impact, reversibility, uncertainty, scale, and the organization’s ability to detect and repair harm.

The examples are illustrative. Sector-specific rules, contracts, professional standards, and local law may require stricter controls. The key point is that higher autonomy is not just a technical setting. It changes staffing, authority, control design, training, incident response, and the evidence an organization must maintain.

NIST makes this distinction directly: some AI systems, such as a model used to improve video compression, may not need human oversight, while other human-AI configurations require clearly defined roles and responsibilities. NIST human-AI interaction appendix Do not spend scarce expert attention on meaningless review queues. Spend it where human judgment can change an outcome.

Redesign the workforce, not just the interface

AI adoption redistributes work. It can remove routine drafting or sorting while creating new work in policy design, data stewardship, quality assurance, exception handling, security, vendor management, customer repair, training, and system supervision. Leaders should make that redistribution explicit before presenting a tool as a productivity gain.

Start with a task inventory. For each job, identify tasks that can be eliminated, accelerated, augmented, shifted to review, or newly created. Then estimate the effect on workload, autonomy, skills, career progression, wellbeing, pay, and staffing. Ask a hard question: if automation removes the entry-level tasks through which people learned the domain, what supervised pathway will develop the next cohort of experts?

Create a workforce plan alongside the technical plan. It should include training time during work hours, access to approved tools, a route for people to report harmful or unreliable behavior, updated job expectations, and a fair plan for roles whose task mix changes. Avoid measuring success only by headcount reduction or raw throughput. Pair productivity measures with quality, rework, incidents, customer outcomes, workload, skill development, and retention. A system that increases volume while driving hidden rework, worker surveillance, or customer churn is not a durable improvement.

Worker voice belongs in this plan. The OECD reports that training and worker consultation are associated with better worker outcomes, and notes that consultation can mitigate risks while improving engagement and acceptance. OECD AI and work That does not mean every consultation will produce agreement. It means leaders should give workforce feedback genuine design influence and explain how it affected the final decision.

Visual brief

1. Define the work: leaders and affected people agree on purpose, limits, and measures.
2. Perform the agreed tasks: AI supports the defined process.
3. Keep human authority: people can approve, override, escalate, or stop consequential actions.
4. Review outcomes: include appeals, incidents, and worker feedback.
5. Adjust the process: the responsible owner acts on those findings.

This loop is the organizational architecture to preserve. It keeps people at the points where intent is defined, conflicts are resolved, harmful behavior is repaired, and the automated boundary is redrawn from evidence.

A 90-day leadership starting plan

First 30 days: map and choose

Inventory current AI tools and shadow uses. For each material workflow, document its objective, users and affected people, data, actions, autonomy level, possible harms, owner, and existing fallback. Prioritize a small number of use cases where benefits are clear and failure is detectable. Do not start with the workflow whose failure would be hardest to repair.

Create a simple decision charter for each pilot. State what the system can do, cannot do, when it must escalate, who can override it, and which measures would cause the pilot to pause. Include frontline workers and representatives early. Their task knowledge is necessary to identify hidden exceptions and unsafe workload shifts.

Days 31 through 60: build control into the work

Train operators, supervisors, and owners on the system’s intended use, known limitations, data boundaries, and escalation procedure. Build an interface that displays the evidence a human needs, not only a recommendation. Test exception paths, logging, rollback, and manual fallback before broad deployment. Set up monitoring for outcome quality, appeals, overrides, incident reports, and differential impact where it is relevant and lawful to assess.

Trial the human role under realistic load. If reviewers cannot meet their normal work while reviewing, the design is not ready. Measure the time per case, agreement and disagreement patterns, reasons for override, and whether humans are adding independent judgment or simply confirming the system. Adjust the boundary before scaling.

Days 61 through 90: evaluate and redesign

Compare the pilot against the baseline on benefits and harms. Review both aggregate results and important slices of the workflow, such as complex cases or groups most likely to be affected. Ask workers and customers whether the escalation path worked in practice. Publish a short internal decision record explaining whether the pilot will scale, change, pause, or end.

Then update roles, staffing plans, skills development, and performance measures. Name the ongoing owner and set a review date. An AI workflow is not “finished” when a model passes evaluation once. Its inputs, model behavior, surrounding policy, and real-world effects can all change.

Questions leaders should be able to answer

• What organizational goal does this AI use support, and which goals or rights does it not get to trade away?
• What may the system do autonomously, and what actions are always reserved for a person with named authority?
• Who is the accountable business owner, who is the technical owner, and who handles the daily exceptions?
• What information does a reviewer see, how much time do they have, and how can they override or stop the system?
• How will people affected by the system question, appeal, or correct an outcome?
• What data shows whether the system improves the actual outcome, rather than merely a convenient proxy or throughput metric?
• How will the organization preserve domain expertise, train people, and create a credible path for new workers to learn the work?
• Which worker, customer, legal, privacy, security, and accessibility voices helped design the workflow, and what changed because of their input?
• What is the safe fallback if the model, provider, data, or integration fails?

If these questions have no clear answers, the organization has not yet designed a human-AI operating model. It has bought or built a tool and hoped that an employee will absorb the uncertainty. The better approach is to make human responsibility concrete, resource it properly, and allow automation to take work only within boundaries people can actually govern.