People using AI isn't a system. Four decisions turn it into one, and you can check yours in five minutes.
An AI-human system has four parts: roles, handoffs, checks and an owner. If you can’t name each one for a piece of work, you have people using AI, not a system. Score your own work with the eight-question test, then draw one workflow as a map. The framework is Future Factors’ working definition, not a standard.
An AI-human system is a piece of work that people and AI do together by design. Each has a defined role, the handoffs between them are explicit, someone checks the output before it matters, and one named person owns the result. Take away any of those four and what’s left is usually people who happen to use AI.
That sounds fussy until you compare two teams using the same assistant. In the first, a support lead pastes a complaint into the tool, tidies the reply and sends it. In the second, the assistant drafts the first reply for each ticket, a named agent reads it against the policy page, refunds go to the team lead, and the lead looks at a sample on Fridays. The tool and the draft are identical. Only the second has been designed, because someone decided who does what.
My view is that the term earns its place only when you can point to those decisions. Below: the four parts, a test for your own work, how the neighbouring terms differ, one workflow drawn out in full, and a design checklist. “AI-human system” isn’t a formal standard term, so I’ll say where other vocabularies overlap.
If you can’t name the role, the handoff, the check and the owner, the system isn’t designed yet.
If you only want the review side, Human-in-the-Loop: How to Actually Write the Guidelines goes deep on it. This piece is the wider frame that review sits inside.
Much of what gets called AI adoption stops at the first part, a person with a tool. The four parts below turn that into something a team can rely on, hand over and improve.
Notice what’s missing from the list. Nothing in it depends on how autonomous the AI is, which tool it runs in, or how big the team is. A system can be one prompt template, one reader and a Friday sample, or an agent with its own permissions. Only the size of the parts changes.
A finance controller opens the month-end pack and finds variance commentary an assistant drafted. It reads well. Nobody can say who was supposed to read it before it went in, or who would answer if a figure were wrong. That’s the owner and the check missing at once.
Better prompting wouldn’t have fixed it.
The owner isn’t always the person doing the work. A content lead can own the process while a copywriter edits each draft. What matters is that the name exists and the person has agreed to it.
The phrase has several neighbours that come from different fields, which is why they get swapped around in meetings.
The most useful primary source I found is the US National Institute of Standards and Technology. Its AI Risk Management Framework describes AI systems as inherently socio-technical, shaped by societal dynamics and human behaviour. It also describes human-AI configurations, which can span from fully autonomous to fully manual, and asks organisations to define the roles and responsibilities of the humans in them.[1] That’s close to what the four parts do in plain words. NIST doesn’t use “AI-human system” as a defined term, and I’m not claiming it does.
Here is how the neighbours line up:
That mapping is our own reading of common usage, and definitions vary by field and author. None of the sources I read gave one agreed definition of the phrase, so treat ours as a working one.
If a colleague says “human-in-the-loop” and you mean the whole arrangement, you’re both right and talking about different sizes of the thing. Why Human in the Loop Matters makes the case for the check part in more detail.
Take one piece of work you or your team do with AI at least weekly, and score it.
Be strict. Score what you could show someone, not what you intend.
| Part | Question | Yes (2) | Partly (1) | No (0) |
|---|---|---|---|---|
| Roles | 1. Can you say in one sentence what the AI does here and what the person does? | ☐ | ☐ | ☐ |
| Roles | 2. Is there something the AI is explicitly not allowed to do in this work (send, approve, decide)? | ☐ | ☐ | ☐ |
| Handoffs | 3. Is the input written down as a brief or template, not retyped from memory each time? | ☐ | ☐ | ☐ |
| Handoffs | 4. Does the output land in a set place, with a set person picking it up? | ☐ | ☐ | ☐ |
| Checks | 5. Does someone look at the output before it has an effect on a customer, a colleague or a decision? | ☐ | ☐ | ☐ |
| Checks | 6. Can that check change the outcome (reject, edit, escalate), and has it done so in the last month? | ☐ | ☐ | ☐ |
| Owner | 7. Can you name one person who answers if the output is wrong? | ☐ | ☐ | ☐ |
| Owner | 8. Can that person pause the AI step without asking permission? | ☐ | ☐ | ☐ |
Copy it into a document and tick one box per row. Maximum 16. Question 6 asks about the last month on purpose: a check that has never caught anything is worth a second look.
Roughly, 0 to 5 means people using AI, which is fine for low-stakes personal work. From 6 to 11, parts of a system exist but something is missing, often the owner or a check that can change the outcome. From 12 to 16, you have a system. The bands are Future Factors’ working judgment, not a validated scale.
A low score isn’t a failure. A private brainstorm that stays on your desk can sit at 4 indefinitely. The test matters once other people rely on the result, and the decision matrix for what to hand to AI and what to keep human helps you decide which work that is.
Failures are rarely dramatic. A part exists on paper and has stopped doing anything.
An HR business partner approves AI-drafted summaries of screening notes for a few weeks and catches the odd slip. Then the volume rises and the approvals speed up. The summaries are right often enough that reading them closely starts to feel like wasted time, which is when a wrong one gets through.
NIST gives that pattern a name in its generative AI profile, listing human-AI configuration as a risk: arrangements between a human and an AI system that can lead to automation bias, over-reliance or the opposite, algorithmic aversion.[3] The point for design is that the person in the loop is a variable, not a guarantee. A system accounts for that with a sample read on a schedule, a rotation of who reads, and a stop rule.
A definition is easier to trust once you’ve seen it drawn. Here is a common back-office workflow, the monthly review of invoices that didn’t match their purchase orders, written out as a system. It’s an illustration, not a client story.
| Part of the system | In this workflow |
|---|---|
| Input | The accounts payable clerk exports the month’s unmatched invoices into a template sheet and pastes in the relevant contract summary for each supplier. |
| AI step | The assistant groups the exceptions by cause (price, quantity, missing purchase order) and drafts a one-line note per supplier using only the pasted contract summary. It may not email suppliers or change a payment status. |
| Human step | The clerk checks each note against the invoice and the contract line, fixes mismatches, and marks disputes. The finance controller decides which disputes to pursue. |
| Check | A second clerk reads the notes marked as disputes before they go anywhere. The accounts payable lead reads a handful of the unflagged notes each Friday. |
| Owner | The accounts payable lead, by name. Answers for the whole month’s output, including the AI-drafted notes. |
| Stop rule | If the weekly read finds a note that contradicts its contract, the AI step is paused until the template is fixed. The lead can pause it alone. |
A Future Factors worked example of the four parts, using roles, not people, so you can swap in your own.
The map doesn’t say which AI tool to use, and it doesn’t claim time savings. It says who does what, what moves between them, who reads it and who answers for it. Run it for a month, then change one row.
One template, one sample read and one name is already a system.
A team can build that internally, and plenty do. An expert-led session helps with the second step: getting a whole team to draw maps for their own workflows, to one standard for what counts as a check.
Ownership is the part people skip, because it feels like an org-chart question. It decides what happens on the day an output is wrong.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Drafts the first reply to each support ticket from the policy page | Sending it, and every refund or exception decision | The agent reads each draft before sending. The support team lead samples sent replies on Fridays. |
| Summarises the week’s interview notes for a hiring panel | Which candidates move forward, and what the panel records as its reason | The HR business partner compares the summary with the raw notes for any candidate who is rejected. |
| Drafts variance commentary for the month-end pack | The explanation the finance controller signs off, and what goes to the board | Controller checks each figure against the ledger. A named reviewer signs the pack. |
A Future Factors working split. In each row the person decides, and the check is by someone who can change the outcome.
The law has already put this into words for a narrow set of cases. Article 14 of the EU AI Act, which covers systems it classes as high-risk, says the people assigned to oversee them must be enabled to understand the system’s capacities and limits, stay aware of the tendency to over-rely on its output, interpret it correctly, decide not to use it or override it, and intervene or stop it.[2] That list applies to those regulated systems, not to every workplace use. Legal requirements differ by country, so ask your own counsel.
I still find the list a useful design prompt for everyday work. Can the person who checks actually say no? Do they know where the tool is weak? Can someone stop it?
A check that can’t change the outcome isn’t a check.
For ownership the practical version is short. Name a person, not a team. Write down that they can pause the AI step. Review the name when the workflow or the person changes. A vendor can be responsible for its product, but your customer or colleague will ask you who looked at the result, so the owner sits on your side of the line.
Contract terms differ, so read yours.
Every AI step has one human who answers for it, and who can switch it off.
Start with one workflow, not five. Pick something recurring that has a clear output, like a weekly report, a reply queue or a monthly pack, and draw the map before you change anything. The workflow diagnostic can help you choose where AI fits at all.
| Part | Done when | Done? |
|---|---|---|
| Roles | One sentence each for the AI step and the human step, in your own words | ☐ |
| Roles | A written list of what the AI step must not do | ☐ |
| Handoffs | The input is a saved brief or template, with the source material named | ☐ |
| Handoffs | The output goes to a set place and a named person picks it up | ☐ |
| Checks | A named checker who isn’t the person who prepared the input | ☐ |
| Checks | The checker compares against a concrete document, not a feeling | ☐ |
| Checks | A scheduled sample read for the outputs nobody flagged | ☐ |
| Owner | One named person, agreed in writing, who answers for the result | ☐ |
| Owner | A stop rule: what makes them pause the AI step, and a note that they can | ☐ |
| Review | A date, about a month out, to look at use, persistence and impact | ☐ |
Copy it into a document. Eight or more boxes ticked is a reasonable first version; the last two tend to get left blank. The threshold is our working judgment.
A month on, check whether the system is holding. Logins and prompt counts only tell you whether people use the tool, so ask three questions in order.
| Question | What you look at | Example |
|---|---|---|
| Use: is it running as drawn? | The template and the weekly sample read | Notes were drafted from the template each month, and the sample read happened. |
| Persistence: still running a month later, unprompted? | Who started it, and who did the second run | The clerk ran it again without a reminder, and the lead still did the read. |
| Impact: is the work better or different? | A before figure you wrote down, and what the check caught | Time to clear exceptions against last quarter; mismatches caught by the second clerk. |
Write the before figure down first, or the impact row has nothing to compare against.
If drawing the map from a blank page is the sticking point, an assistant can produce a first draft from your description. Use a tool your organisation has approved, and leave out customer data and personal details.
| Paste this, then fill in the brackets |
|---|
| I’m a [your role] at a [size and type of organisation]. I want to draw one recurring piece of work as a system where people and AI each have a defined role. The work is: [describe it in two or three sentences, including who receives the output]. Draft a one-page map with these rows: Input, AI step, Human step, Check, Owner, Stop rule. For each row, give a specific proposal using role titles, not names. List what the AI step must not do. Where you have to guess, write ASSUMPTION and tell me what to confirm. Don’t suggest tools. Finish with the two rows you think are weakest. |
Copy, fill in the brackets and paste. You still decide every row.
Who the owner is depends on the role. A finance controller will usually own a pack or reconciliation, an HR business partner anything that touches a hiring decision, a support team lead the reply queue, and a content lead voice and claims. Draw one map per workflow, not one for the whole department.
You can build this on your own, and a first map usually takes an afternoon. If you want a team to design theirs together, on their real work and to a common standard, that’s what our corporate workshops are for. They shorten the time to a shared way of working.
This week, pick one recurring piece of work, draw its map, run the eight-question test and name the owner.
Then look again in a month.
Not by itself. When one person prompts, reads and uses the answer, all four parts collapse into that one person. That’s fine for low-stakes work, and it becomes a system once someone else relies on the result.
Human-in-the-loop usually means a person reviews or approves before output takes effect. Here that is the check, and sometimes the owner. The system also needs roles, a defined handoff and a named owner.
No. Scale the design to the consequence. A private brainstorm needs none of it. A customer-facing reply, a hiring summary or a figure in a board pack needs all four parts.
The vendor can answer for its product, but the person who passes the output on answers for the work. Contract terms and legal duties vary, so check yours.
The four parts stay. The check tends to move earlier and into the setup: what the agent may touch, what it may not do, and a stop rule, because there may be no person reading each output. Our guide to what AI agents are covers the building blocks.
I read three primary sources for this guide on October 29, 2026: NIST’s AI Risk Management Framework 1.0 and its Appendix C on human-AI interaction, Article 14 of the EU AI Act as published by the European Commission’s AI Act Service Desk, and NIST’s Generative AI Profile (AI 600-1). None of them defines “AI-human system” as a term, so the definition and the four parts are Future Factors’ working framework, not a standard. I left out older academic work on levels of automation because I couldn’t read a primary copy while writing. I didn’t test any tool hands-on for this article, and the scenes and the worked example are illustrations, not client stories.