Explore our AI courses, practical training for non-technical teamsExplore courses →Explore AI courses →
AI Decision MakingAI for LeadersHuman Oversight

A Decision Matrix for What to Hand to AI and What to Keep Human

Score a task on five things, apply three overrides, and land on hand off, hand off with review, AI assists or keep human. Ten tasks worked through, plus a worksheet.

TLDR: Whether a task goes to AI is a decision to score, write down, give an owner and revisit on a date. Score consequence, judgment, data, checking and volume from 1 to 3, apply three overrides, and read off one of four outcomes. The criteria are our judgment, not research. Sources read October 26, 2026.
5criteria, each scored from 1 to 3.
4outcomes: hand off, hand off with review, AI assists, keep human.
10worked tasks across six roles, with a blank worksheet.

Share this article

The Short Version

Score each task from 1 to 3 on consequence, judgment and relationship, data sensitivity, difficulty of checking, and volume. Add them, apply the overrides, and land on hand off, hand off with review, AI assists or keep human. Name an owner and a re-score date for every row.

Most teams draw the line by feel, and feel differs from person to person

In many teams the line between what goes to AI and what stays with a person gets drawn one task at a time, by whoever happens to feel confident or cautious that day. The recruiter who likes the tools lets them write every candidate email. The finance analyst at the next desk won’t let them near a spreadsheet. Both have reasons, and neither set of reasons is written down.

A support lead finds out that two agents have been answering warranty questions in different ways. One pastes the policy into an assistant and sends what comes back. The other writes every reply by hand. Neither broke a rule, because there wasn’t one, and the customers who got the faster answer also got the one that was never checked against the policy.

There was just no line.

A decision matrix replaces that with something a team can look at together. You score a task on five things, apply three overrides, and land on one of four outcomes: hand off, hand off with review, AI assists, or keep human. The criteria and scores here are Future Factors judgment, not research findings, and the numbers are there to start a conversation. What the matrix adds is a written answer, a named owner and a date to look at it again.

If you want a quick fit test for a single task you’re about to do yourself, our piece on when not to use AI at work is built for that. This guide is for the person who sets the line for a whole role or team: it scores a list of tasks, records the outcome and says when to revisit it. The checklist that goes with a review step lives in our guide to human-in-the-loop guidelines, so I won’t rebuild it here.

Five things to score, four places to land

Each criterion is scored from 1 to 3, and a 3 points toward keeping a person involved. Volume is the one to read carefully. A task that repeats every week scores 1, because it’s worth building a checked process around. A one-off scores 3, because setting that process up costs more than doing the work.

The five criteria, scored 1 to 3

Criterion1: points toward hand-off23: points toward a person
Consequence if wrongFixed in minutes, and nobody outside the team sees itCosts time or money, or reaches a customer or colleague, and can be put rightHarms a person, breaks a rule or a contract, or can’t be taken back
Judgment and relationshipFollows rules or a clear patternNeeds some judgment about context or toneRests on trust, negotiation or a decision about a person
Data sensitivityPublic, or already fine to shareInternal, or ordinary personal detailsSensitive personal, confidential client, regulated or privileged
Difficulty of checkingCheckable in minutes against a source you holdTakes real time or specialist knowledgeCan’t be checked without redoing the work
Volume and repetitionSame shape every week or moreMonthly or quarterlyRare or one-off

A Future Factors working rubric. Score the task as it is done today, with the tools you actually have approved.

Add the five scores and read the band. The total runs from 5 to 15, and the outcomes mean this:

  • 5 to 7, hand off. AI does the task from start to finish, and a person spot-checks a sample on a set schedule.
  • 8 to 10, hand off with review. AI produces the output, and a named person reviews each one against a checklist before it’s used.
  • 11 to 13, AI assists. A person does the task and owns the result. AI helps with parts of it, such as research, structure or a second look, and nothing AI-made goes forward unedited.
  • 14 to 15, keep human. A person does it without AI, by decision, and the decision is written down.

A total can hide one dangerous score, so three overrides sit on top of the bands. Apply them after you add up, and if more than one applies, the strictest wins.

  • A 3 on consequence rules out a plain hand-off, whatever the total. The lightest outcome left is hand off with review.
  • A 3 on checking rules out hand off and hand off with review alike, because a review only works if someone can actually do the check. The lightest outcome left is AI assists.
  • A 3 on data sensitivity means keep human unless your approved tool is cleared for that kind of data. If it is cleared, treat it like a 3 on consequence.
The Override Rule

One high-risk score outranks the total.

I chose these five because each changes what a sensible person would do. They are how bad a mistake is, how much the work rests on a person, what information goes in, whether you can tell if the output is right, and whether the task repeats enough to justify a process. How capable the model is isn’t on the list, because it changes every few months, which is what the re-score step is for. What your tools are cleared for is set by policy, and our guides to writing an AI acceptable use policy and keeping company data out of the wrong tools cover that side.

Score the smallest task, not the whole job

The performance review is the usual trap. A manager who scores “performance reviews” as one task gets either a full hand-off, which is wrong, or a full ban, which throws away the useful part. Preparing for the conversation, by organising notes and anticipating questions, is one task. Holding the conversation and setting the rating is another, and they land in different places.

The test is quick to run. If two parts of a piece of work would get different answers, split them and score each.

The Smallest Task Rule

Score the smallest piece of work that could get a different answer.

Ten tasks across six roles, scored

Here are ten recurring tasks across six roles. Read the scores as C-J-D-K-V: consequence, judgment, data, checking, volume. They are one team’s reasonable answers, and yours will differ with your tools, customers and contracts.

Ten tasks, scored and placed

Role and taskScores C-J-D-K-VTotalOutcomeWhy it lands there
Recruiter: scheduling interviews and sending confirmations1-1-2-1-16Hand offCandidate contact details are ordinary personal data, and a wrong time is fixed with a reply.
Recruiter: shortlisting candidates from applications3-3-2-3-112AI assistsRejected candidates are invisible, so the recruiter can’t check what was filtered out. AI summarises each CV against criteria the recruiter wrote, and the recruiter picks the list.
Finance analyst: reformatting the monthly variance table2-1-2-1-28Hand off with reviewTotals reconcile to the ledger in minutes, so a review is cheap and real.
Finance analyst: explaining why the numbers moved2-3-2-2-211AI assistsThe reasons sit with the budget holders. AI can suggest questions to ask and a structure, and the analyst writes the explanation.
Marketer: adapting approved copy to each channel’s format1-1-1-1-15Hand offThe wording is already approved. Only length and layout change.
Marketer: drafting the monthly customer newsletter from approved content2-2-1-1-28Hand off with reviewIt reaches customers, so a named editor reads every issue before it goes.
Support lead: answering warranty and safety questions from the policy text3-1-1-1-17Hand off with reviewThe total says hand off. The 3 on consequence moves it up one outcome.
Manager: preparing for a performance conversation2-2-3-2-312AI assistsOrganising notes, anticipating questions. Assumes the tool is cleared for employee data.
Manager: holding the conversation and setting the rating3-3-3-3-315Keep humanThe task is the relationship, and there’s nothing to check it against.
HR business partner: summarising medical-leave documents3-2-3-2-313Keep humanThe band says AI assists. The data override applies because the example tool isn’t cleared for health information.

Scores are Future Factors judgment for illustration, not measured data. C = consequence, J = judgment and relationship, D = data sensitivity, K = difficulty of checking, V = volume.

Three rows are worth a second look. The warranty answers show the override doing its job. The task is easy to check, but a wrong answer on a safety question can’t be put right with a reply. The medical-leave summary shows the data override moving a task from AI assists to keep human.

When two people score the same task differently, the gap is the useful part. One of them usually knows something about the consequence or the data that the other doesn’t. Use the higher score until you’ve found out why they differ.

When the law has already drawn part of the line

Some rows aren’t only a judgment call. The EU’s AI Act lists AI systems used to recruit or select people, including to analyse and filter applications and to evaluate candidates, among its high-risk uses. It does the same for systems used to make decisions about terms of work, promotion or termination, to allocate tasks based on behaviour or personal traits, or to monitor and evaluate performance.[1] That puts the recruiter shortlist and the manager’s rating in territory with extra obligations if you operate in or sell into the EU. This isn’t legal advice. For those rows, agree the outcome with whoever handles compliance rather than leaving it to the matrix alone.

On the governance side, NIST’s AI Risk Management Framework asks organisations to have policies and procedures that define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems.[2] The owner and re-score columns in the worksheet below are one small way to write that down for a team. The link to this matrix is our reading, not NIST’s endorsement of it.

What hand off with review has to mean

The review outcome is where this tends to go wrong. A person is nominally in the loop, the output looks fine, and the check quietly becomes a skim. A review row needs three things on paper: what the AI does, what the human still owns, and how the check happens. Here are the three review rows from the worked example, filled in.

Hand off with review: who does what, and how it’s checked

TaskWhat AI doesWhat the human still ownsHow it gets checked
Variance table (finance analyst)Reformats the export into the template and flags totals that don’t tieEvery figure, and the decision to share itThe analyst reconciles totals to the ledger before it goes out, and the finance lead picks one line item at random each month and traces it back
Customer newsletter (marketer)Drafts from approved content and offers subject linesEvery claim, offer and link, and what has been left outA second marketer reads each issue against the source content before it’s scheduled
Warranty answers (support lead)Drafts a reply using only the policy text pasted inThe answer, any exception, and the call to escalate when the policy is unclearThe agent compares the reply to the policy sentence by sentence, and the support lead reads a weekly sample, including replies that weren’t edited

A Future Factors working table. The checks are suggestions to adapt, not requirements from any standard.

Two details make a review real.

The reviewer has a name, not a team, and the check is against a source such as the ledger or the policy text, not against whether the draft sounds right. For how deep each review should go, see our guide to writing human-in-the-loop guidelines, and for what happens when the check exists only on paper, why human in the loop matters.

A blank worksheet, and a prompt to help you fill it in

You can run this with a team in under an hour, with the people who do the work in the room.

  1. Pick one role and list ten recurring tasks, small enough that each could get its own answer.
  2. Each person scores every task alone, before anyone compares.
  3. Compare, find the gaps, and use the higher score until you know why they differ.
  4. Apply the three overrides and read off the outcome.
  5. Write a name against each task as owner, and a reviewer where the outcome needs one, then set a re-score date.

Blank worksheet: copy it into a shared document

RoleTaskCJDKVTotalOutcome after overridesOwner and reviewerRe-score by
__________________________________________
__________________________________________
__________________________________________

C = consequence if wrong, J = judgment and relationship, D = data sensitivity, K = difficulty of checking, V = volume and repetition. Score each 1 to 3, using the rubric above.

If you’d like help drafting the first pass, an assistant can interview you and propose scores. Describe tasks in general terms and leave out names, figures and anything confidential.

A prompt to draft your scores

I’m a [role] on a [function] team. I want to decide which of my recurring tasks to hand to AI. Interview me one question at a time.

First, ask me to list up to ten recurring tasks, described in general terms with no names, figures or confidential details.

Then, for each task, ask me these five questions and wait for my answer each time: What is the consequence if the output is wrong? How much does it rest on judgment or a relationship? How sensitive is the data involved? How hard is it to check the output? How often does it repeat?

Score each answer from 1 to 3, where 3 points toward keeping a person involved, and add the scores. Bands: 5 to 7 hand off, 8 to 10 hand off with review, 11 to 13 AI assists, 14 to 15 keep human. Overrides: a 3 on consequence rules out plain hand-off; a 3 on checking rules out both hand-off outcomes; a 3 on data sensitivity means keep human unless my approved tool is cleared for it.

Show a table of task, five scores, total and outcome. Mark any score where you guessed. Do not decide for me: I will confirm or change every score.

Fill in the brackets and paste into your approved AI tool. The assistant proposes and you decide, which is AI assists applied to the matrix itself.

Teams can do this internally. When several teams are scoring at once, working with someone who has run the exercise before shortens the learning curve and builds a shared standard across teams instead of in one careful person. That’s the kind of session our corporate workshops are built around.

When to score it again

A hand-off is a decision made on today’s tool, today’s data and today’s stakes, and all three move. Check this list when the re-score date arrives, or sooner if something on it happens.

Re-score triggers: tick any that apply

TriggerWhat to look for
The tool changedA new model, an upgrade, a switch of vendor or a new plug-in
An error got throughA mistake reached a customer or colleague, or the review step is being skipped
The data changedA new kind of customer or employee information now flows into the task
The task changedA new audience, a new channel or higher stakes
The volume changedOutput has grown past what the sample check covers
The owner or reviewer moved onThe name on the worksheet is no longer doing the job
A rule changedCompany policy, a client contract or the law
The date passedThe re-score date is behind you

No ticks: keep the date. One or two: re-score that task this month. Three or more, or an error that reached a customer: pause the hand-off until it has been re-scored.

The Re-score Rule

A decision to hand off is a decision to look again, on a date.

Keep each decision in a form someone else can read in a minute. This is one row from the worked example, written up as a record.

A completed decision record

FieldEntry
TaskAnswering warranty and safety questions from the policy text (support lead’s team)
Scores and total3-1-1-1-1, total 7. The band says hand off.
OutcomeHand off with review, because of the override on consequence
Owner and reviewerSupport lead owns the task. The agent on shift reviews each reply against the policy text.
CheckSentence-by-sentence against the policy, plus a weekly sample read by the support lead
Re-score byEnd of next quarter, or sooner on any trigger above

Copy the fields into your own tracker. One record per task keeps the reasoning with the decision.

To judge whether the matrix itself is working, use the lens we apply to any rollout: use, persistence and impact.

A quarterly check on the matrix: use, persistence, impact

QuestionWhere to lookYes means
Use. Does the team reach for the worksheet when a new task comes up?The shared worksheet and recent task requestsThe matrix is part of how work gets assigned
Persistence. Are owners and reviewers still doing their part a quarter later, without a reminder?The review log or sample readsIt has become a habit, not a launch event
Impact. Can you point to a change, such as fewer errors reaching customers, a review that now catches problems, or time freed on low-scoring tasks?Error reports and the owners’ own notesThe matrix is earning its keep

Score one point per yes. Three: keep going. Two: fix the weakest row. One or none: look at the process before blaming the people.

This week, pick one role on your team and list ten recurring tasks. Score them alone, then with the person who does the work, and put a re-score date on every row.

Sana Mian
Sana Mian, Co-Founder of Future Factors AI

Sana is an AI educator and learning designer specialising in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.

More about Sana →

Frequently Asked Questions

Who should do the scoring, the manager or the team?

Both, in order. The person who does the task scores it alone first, because they know where it goes wrong. The manager then compares scores, settles the gaps and owns the written record.

Does keep human mean never use AI on that task?

It means no AI on that task until the next re-score, with the decision written down so it isn’t re-argued every week. The tasks around it can still use AI, which is why the Smallest Task Rule matters.

What if two people score the same task differently?

Treat the gap as information. Ask each person what they assumed about the consequence and the data, then use the higher score until the difference is explained.

How does this fit with my company's AI policy?

The policy sets what your approved tools are cleared for, and that decides your data scores. The matrix sits underneath it, task by task. If a score conflicts with the policy, the policy wins. Our guide to writing an AI acceptable use policy covers how to set that foundation.

Does the matrix work for AI agents that take actions on their own?

Yes, if you score each action the agent can take, not the agent as a whole. Consequence and checking carry the weight, because an agent can repeat a mistake quickly. Any action that scores a 3 on consequence should wait for a person’s approval. For the wider picture, see how to be an agent boss on a marketing team.

About This Article

I read two sources on October 26, 2026: the European Commission’s AI Act Service Desk page on employment, cross-read against the Annex III wording of Regulation (EU) 2024/1689 as reproduced at artificialintelligenceact.eu, because EUR-Lex itself did not load for scripted access; and the AI RMF Core page on NIST’s AI Resource Center, which was last updated in June 2026 and notes that a revision of the framework is in progress. I did not test any tool hands-on for this article. The five criteria, the bands, the overrides, the re-score triggers and every score in the worked example are Future Factors judgment, not research findings, and nothing here is legal advice. The scenes are illustrations, not client stories.

Sources

  1. European Commission, AI Act Service Desk, Employment (guidance page on Annex III point 4, employment, workers’ management and access to self-employment; no date shown on the page), cross-read against the Annex III wording of Regulation (EU) 2024/1689 as reproduced at artificialintelligenceact.eu/annex/3/. Annex III 4(a): AI systems intended for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates. 4(b): systems intended to make decisions affecting terms of work-related relationships, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour. EUR-Lex itself did not load for scripted access, so the Annex wording was read on the reproduction. Read October 26, 2026. https://ai-act-service-desk.ec.europa.eu/en/node/464
  2. National Institute of Standards and Technology, AI Risk Management Framework, AI RMF Core (airc.nist.gov; page last updated 2026-06-18, with a note that AI RMF 1.0 is being revised): GOVERN 3.2, ‘Policies and procedures are in place to define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems.’ Read October 26, 2026. https://airc.nist.gov/airmf-resources/airmf/5-sec-core/

Psst, Hey You!

(Yeah, You!)

Want helpful AI tips flying Into your inbox?

Weekly tips. Real examples. Practical help for busy professionals.

We care about your data, check out our privacy policy.