An AI-assisted role should explain the outcome a person owns, what AI handles, which decisions the person can make, and how their contribution is evaluated. Write those commitments before listing tools. A job description that promises responsibility also needs to explain the authority, information, and time that make responsibility possible.

Imagine hiring a support specialist to “own customer outcomes with AI.” During interviews, you describe judgment, empathy, and room to improve the service. Once hired, the person spends the day approving suggested replies, needs permission to change a refund decision, and gets challenged whenever careful investigation increases handling time.

The problem is in the job you designed. You asked someone to be accountable while keeping the decisions elsewhere.

A paper currently being discussed on [Hacker News](https://news.ycombinator.com/item?id=49601814) raises a related question: what happens to the meaning of work when a machine could perform a task that a person still does? It offers a useful reason to examine your next role before you advertise it. It does not establish that AI destroys purpose, or that a better job description repairs it.

## What changes when AI can do work you still perform?

Automation can change how people understand their contribution even when their job remains. That possibility is the subject of Joshua Gans's July 2026 working paper, [*Replaceable but Employed: Automation and the Meaning of Work*](https://www.nber.org/papers/w35559).

Gans models workers who value both useful output and their own contribution to producing it. Within the model, a credible automated alternative can weaken that second source of value before the employer replaces anyone. The work still happens, but the worker's view of what they contribute can change.

**This is a theoretical mechanism, not a measured workplace trend.** The paper does not estimate how many employees experience it. Its discussion of wages depends on assumptions about how compensation adjusts. Its later sections propose empirical tests rather than reporting that the mechanism has already been confirmed.

The distinction matters when you turn research into hiring advice. “People need to believe machines cannot do their jobs” would be a poor conclusion. Gans explicitly distinguishes meaningfulness from indispensability. Teaching, reviewing, caring for someone, and providing a useful service can matter even when somebody else could do them.

Nor should you preserve pointless manual tasks to manufacture a sense of contribution. A support specialist does not need to type every sentence to help a customer. They may contribute through finding a missing fact, choosing a fair response, or noticing a product defect that the draft never mentions.

The hiring question is concrete: **what can this person change, and who benefits when they change it?** If you cannot answer, “human oversight” is an unfinished specification. If you can, describe that contribution without promising that it will remain technologically unique forever.

## What does the evidence say about meaningful work and AI?

The evidence warrants examining job design, but does not show that every use of AI makes work worse. Studies of industrial robots, earlier workplace AI, and generative assistance describe different technologies and populations.

In [*Robots, meaning, and self-determination*](https://www.sciencedirect.com/science/article/pii/S0048733324000362), Milena Nikolova, Femke Cnossen, and Boris Nikolaev analyze 14 industries in 20 European countries between 2005 and 2021. Their estimates connect robotization with lower perceived meaningfulness and autonomy. This is relevant evidence about work quality, but it concerns industrial robots. It does not test what happens when your team adopts a chat assistant, or verify Gans's proposed effect before adoption.

The OECD provides an important counterweight. Its [2023 employer and worker surveys](https://www.oecd.org/en/publications/the-impact-of-ai-on-the-workplace-main-findings-from-the-oecd-ai-surveys-of-employers-and-workers_ea0a0fe1-en.html) covered 5,334 workers and 2,053 firms in finance and manufacturing across seven countries. Results were generally positive about performance and working conditions. Training and worker consultation were associated with better outcomes, although the surveys do not establish that those practices caused them.

The OECD's [job-quality chapter](https://www.oecd.org/en/publications/oecd-employment-outlook-2023_08785bba-en/full-report/artificial-intelligence-job-quality-and-inclusiveness_a713d0ad.html) also reports that 58% of finance AI users and 59% of manufacturing AI users said the technology increased their control over task sequence. Smaller shares reported less control. Choosing the order of tasks is narrower than deciding what your team should do, but these findings challenge a simple story of inevitable lost autonomy.

Those surveys predate the current generative-AI wave. They capture employed workers' perceptions, leaving out people who had already departed. Treat them as evidence from a particular setting, not a forecast for every new role.

A more recent example comes from [*Generative AI at Work*](https://arxiv.org/abs/2304.11771), by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond. The updated study follows a staggered customer-support deployment involving 5,172 agents. Access to assistance increased issues resolved per hour by 15% on average, with substantial differences by skill and experience. Customer interactions improved too.

In that deployment, agents could edit or ignore suggestions while remaining responsible for the conversation. That is a useful example of assistance alongside discretion. **The study does not prove that discretion caused the gains**, measure the meaning mechanism above, or establish effects across the economy.

Read together, these sources leave room for very different experiences of automation. The practical recommendations below are proposals for making a role explicit. None of these studies tests this checklist or promises that using it improves retention, satisfaction, or hiring results.

## Define the decisions your AI-assisted hire will own

A person can be responsible for an outcome without having enough authority to affect it. Before opening the role, agree on the decisions the hire can make and the resources available to support those decisions.

Start after the decision to hire. Whether you need a person, a contractor, or automated capacity belongs in your [headcount planning](/blog/headcount-planning-humans-ai-agents-new-hr-role). Here, you have decided that a person belongs on the team. Your task is to describe a job they can actually perform.

Consider this **hypothetical support role**:

| Part of the work | What AI does | What the person can decide | What requires escalation |
|---|---|---|---|
| Routine customer reply | Retrieves policy and proposes a draft | Check facts, rewrite the reply, and send it | Conflicting policy or missing account evidence |
| Disputed refund | Summarizes the request and relevant history | Investigate and approve within the agreed policy | Exceptions outside that authority |
| Unreliable suggestion | Offers an answer that may be wrong | Discard it and use the manual workflow | Recurring failures that need an owner |
| Product feedback | Groups recurring customer issues | Bring evidence and recommend a change | Product prioritization remains with the product lead |

The table makes limits visible. “Own the customer experience” could imply that the specialist controls product priorities. This role does not. It gives them a route to present evidence while naming who makes the final call.

Now test whether the permissions survive a busy day. Can the specialist discard a draft without asking first? Can they access the account history needed to check it? Will you allow investigation time when the answer is unclear? Who answers an escalation when the manager is away?

A commitment in the brief means little if working conditions contradict it. Someone evaluated solely on speed may have a formal right to investigate but little practical room to use it. Someone told to escalate needs an available recipient who can decide.

Write the manager's commitments beside the employee's responsibilities. Those commitments might include access to policy records, training before independent work, a named escalation contact, and time to review recurring errors. Be specific about what is available now and what you still need to arrange.

Meaningful contribution also extends beyond catching mistakes. The person might explain a confusing policy so a customer can act, identify an unnecessary step, or teach teammates something learned from repeated cases. You do not have to invent a strategic project for every role. You do need to explain how ordinary work helps somebody.

## Rewrite the job description around real contribution

An AI-assisted job description should let a candidate picture the work and its limits. Replace broad promises of ownership with decisions, support, and evaluation criteria that the hiring manager is prepared to honor.

Here is the weak version of our hypothetical role:

> You will own customer happiness, use the latest AI tools, monitor quality, and thrive in a fast-moving environment.

A more useful version would say:

> You will help customers resolve account and billing questions. Our AI system retrieves policy and drafts replies; you check the facts, choose the response, and can discard a draft when it is unsuitable. You can approve refunds within the agreed policy and escalate exceptions to the support lead.
>
> You will also bring recurring customer problems to our product-feedback review. We assess resolution quality, the reasons behind escalations, and response time. During onboarding, you will practice on reviewed cases before handling the queue independently.

This is example copy, not a description of a particular employer. It is also a commitment. Do not paste it into an advert if your system sends replies automatically before the person can review them, or if your support lead cannot support the promised escalation path.

Name uncertainty as well. You might currently review every reply while testing automation for routine requests. Tell candidates which parts are settled, which are changing, and how the team will discuss changes to their responsibilities. “The role will evolve” is more informative when you explain what is under review.

Your [job-description guide](/blog/writing-job-descriptions) can help turn those details into readable copy. Keep the substance intact when you shorten it. Removing the limits while retaining the word “ownership” gives candidates a different promise.

Avoid selling purpose as a substitute for pay, manageable workload, or competent management. Research about meaning does not justify a compensation discount. Candidates are entitled to assess the role as a whole, including the parts you would rather not feature in the opening paragraph.

## Let candidates see how judgment works on your team

Give candidates a realistic view of a decision before asking whether they want to make it every day. A recent example can reveal more than a list of approved AI tools.

Ask the hiring manager to walk through a sanitized case where a suggested answer was disputed. Show what information the person could access, what they changed, who approved any exception, and what happened next. If the manager cannot find such a case because the workflow is new, say that and use a clearly labeled hypothetical example.

Let candidates ask awkward operational questions. What happens when the person disagrees with the suggestion? Which decisions get reversed by a manager? Does slowing down to investigate count against them? What has the team changed after an employee identified a problem?

These questions are a preview of the relationship you are offering. A candid answer about limited authority is more useful than promising freedom the job cannot provide. A candidate may prefer a well-defined service role to a broader one with unclear expectations.

For an assessment, keep the same connection to real work. You could give the candidate an imperfect draft and enough fictional account information to evaluate it. Ask what they would send, what evidence they relied on, and what they would escalate. Disclose the permitted tools and criteria beforehand.

Score the decisions relevant to the role, including whether the candidate recognizes a missing fact or respects an authority boundary. Do not quietly reward taking an unauthorized action because it looks proactive. Conversely, do not call every request for clarification a failure of independence.

Our guide to [building a human review team](/blog/mturk-shutdown-human-in-the-loop-team) covers representative samples and reviewer selection in more detail. Here, the exercise has a narrower purpose: checking that your evaluation matches the authority you advertised. One exercise does not diagnose someone's motivation or establish that they will find the job meaningful.

## Check whether the job matches the promise

After the hire starts, compare the advertised responsibilities with the decisions the person actually makes. Automation can change the work faster than a job description gets revised.

Use a recent case in a manager conversation. Ask where the employee's judgment changed the response, what helped the customer, and where they were held responsible for something they could not decide. These are conversation prompts, not a validated measure of meaningful work.

Follow a specific case through to the decision. Perhaps the person repeatedly escalates the same exception and waits for approval. You might expand their authority after training, clarify the policy, or keep the boundary because the consequences require a different owner. Any of those choices can be reasonable; leaving responsibility vague is the part to fix.

Look at the time available too. If automation removes routine cases, the remaining queue may contain more difficult investigations. Review whether the old handling-time expectation still describes the job. Do not assume that a shorter queue means the remaining work is easy.

Revisit commitments when tools change. If the system starts sending answers without review, explain what the person's role becomes and whether their responsibilities should change with it. Someone should not remain answerable for a decision they can no longer inspect or stop.

The OECD's findings about consultation give you a reason to listen to workers' experience, while leaving the causal question open. The employee can describe what happens in the workflow. You still need to decide what to change, resource the decision, and explain it.

## Carry the role definition into your hiring process with Kit

Keep the public role, candidate instructions, and evaluation criteria connected to the same decisions. A hiring process can record those commitments; the manager must make them true in daily work.

In Kit, you can put the role's outcomes and authority boundaries in the job description, then explain the exercise in candidate-facing stage instructions. Process templates provide reusable stages and reviewer assignments. You can configure a Team Review stage with named criteria and collect recommendations, scores, and comments.

Kit's review interface hides colleagues' reviews until a reviewer submits their own assessment. That supports recording an independent view before comparing notes. It does not establish that the assessment measures meaningful work, or guarantee an unbiased hiring decision.

Kit also does not enforce an employee's authority over workplace AI tools after hiring. That remains a management commitment. The useful starting point is a role people can understand: what the tool does, what the person contributes, and what they are empowered to decide.

> [!CTA]
> **Define the decisions before advertising the role.** Start with a [Kit hiring template](/templates), then adapt the job description, interview instructions, and review criteria to the work your hire will actually own.