Human-Led AI · The Autonomy Spectrum
How much should the AI do? There are five answers, and only one of them is wrong.
Every AI task in your business sits somewhere on a five-level scale. Most vendors sell you Level 4 and call it efficiency. Most businesses want Level 1 and call it caution. Both miss the point: the level isn't a setting, it's something a task earns. Here's the whole scale, with examples, and where I'd start you.
The scale at a glance
| Level | Name | What happens | Who's accountable | Failure looks like |
|---|---|---|---|---|
| 0 | Manual | A person does the whole task | The person | Slow, inconsistent, slips when busy |
| 1 | Assist | AI drafts, a person edits and sends | The person | A weak draft you fix in a minute |
| 2 | Approve | AI completes it, a person approves before it goes out | The person | One bad item you reject before it leaves |
| 3 | Audit | AI acts, a person reviews a sample afterwards | The person, by exception | A handful of bad items you catch in the weekly sample |
| 4 | Autonomous | AI acts, nobody checks | Nobody, until something breaks | Two hundred bad items, discovered by a client |
Where I build: Levels 1, 2 and 3. Level 4 exists on this page because pretending it doesn't would be dishonest. Some tasks genuinely earn it. I'll tell you when one has, and the decision is yours.
Level 0 — Manual. A person does the whole task.
This is where every task starts, and where plenty of tasks should stay. If it happens twice a year, involves judgement you can't write down, or the cost of a mistake is your biggest client, leave it here. Not everything needs AI. A surprising amount of my job is saying so.
Stays at Level 0: anything with a legal signature, anything involving a person's employment, first contact with a major prospect.
Level 1 — Assist. The AI drafts, your person edits and sends.
The AI produces a first version. A person reads it, changes what needs changing, and sends it under their own name. The person still does the task; they just start from 80% instead of zero.
Good for: proposals from templates, monthly report first drafts, meeting notes and actions, position descriptions, policy Q&A drafts.
What it feels like: your admin lead opens her morning and the three replies she'd have spent forty minutes on are waiting as drafts. She fixes a tone, adds a detail, sends. Twelve minutes.
When it graduates: when the edits stop. If your person is sending drafts unchanged for a fortnight, the task is ready for Level 2.
Level 2 — Approve. The AI completes the task, your person approves before it goes out.
The AI does the whole thing: reads the input, does the work, prepares the output, queues it. Nothing leaves until a person clicks approve. The person's job shifts from doing to checking.
Good for: invoice follow-ups, enquiry replies to common questions, onboarding pack assembly, CRM updates from emails and calls, compliance form pre-fill.
What it feels like: a queue of twelve invoice reminders, each showing the client, the amount, the days overdue and the message. Your bookkeeper scans the list, holds one because that client is in a dispute, approves eleven. Ninety seconds.
Why this is the level most businesses live at: it gives you almost all of the time back with none of the "what did it send?" anxiety. Human-Led AI is mostly Level 2.
When it graduates: when approvals become rubber stamps. If nothing's been rejected in a month, and the cost of a rare mistake is recoverable, the task is a candidate for Level 3.
Level 3 — Audit. The AI acts, your person reviews a sample afterwards.
The AI does the work and sends it. A person reviews a sample — every tenth item, or everything from the last week, or anything the AI flagged as unusual. Accountability stays with the person, exercised by exception instead of every time.
Good for: appointment confirmations and reminders, booking confirmations, routine document requests, logging calls and emails to the CRM, roster reminders.
What it feels like: Friday afternoon, a one-page summary: 340 confirmations sent this week, 4 flagged as unusual, here they are. Your person reads the four, nods, closes the tab.
Guardrails that have to exist first: a clear rule for what "unusual" means, a hard stop on anything the AI shouldn't attempt, and an audit log so you can answer "what did it send to whom" in thirty seconds. These are Groundwork layers 4 and 5. Without them, Level 3 is just Level 4 with better intentions.
When it graduates: rarely, and only if you want it to. Most tasks never need to.
Level 4 — Autonomous. The AI acts, nobody checks.
No draft, no approval, no sample. The task runs and the only signal you get is when something breaks.
I don't dismiss Level 4. Some tasks earn it: internal notifications, data moving between two systems you own, a reminder that goes to you and nobody else. If the worst outcome is mild internal inconvenience, Level 4 is fine and I'll say so.
What I won't do: is start a task here, or move a task here because it's cheaper. A task reaches Level 4 by graduating through the three below it, on your data, with your evidence, and with your decision at the end.
A task moves up when the evidence says so. Not when the vendor does.
Three questions decide it, and you answer them with data from your own business rather than a demo:
1. Has the human stopped changing anything?
At Level 1, are the drafts going out unedited? At Level 2, has anything been rejected in the last month? If your person is still catching things, the task isn't ready. That's not failure; that's the level doing its job.
2. What's the worst single mistake, and can you undo it?
A wrong appointment reminder is a phone call. A wrong payment instruction is not. Recoverable mistakes can move up. Unrecoverable ones stay where a person sees every item.
3. Would you know?
Is there a log? Is there a sample review? Is there a flag for "this one looked odd"? If the answer to "what did it send last Tuesday" is a shrug, nothing moves.
Graduation is written down: which task, which level, who decided, when it's reviewed. That record lives in your Groundwork agent register, so the decision survives the person who made it.
One task, four months.
Month 1 — Level 1.
The follow-up assistant drafts overdue-invoice reminders. Your bookkeeper edits about a third of them, mostly for tone with long-standing clients.
Month 2 — still Level 1, then Level 2.
Edits drop to one in twenty. We add the tone rules she was applying by hand. She agrees to move to Approve: the assistant sends nothing until she clicks.
Month 3 — Level 2.
She approves the queue each morning in under two minutes. One rejection all month: a client in a payment plan. We add the payment-plan check as a hard rule.
Month 4 — Level 3 for standard reminders.
First and second reminders now send automatically; she reviews the week's list every Friday. Final notices stay at Level 2, because those are the ones that damage relationships.
Hours back: about four a week. What she did with them: the supplier reconciliation that had been three months behind.
FAQ
Isn't Level 1 just using ChatGPT?
Sometimes, and if that's all you need, I'll tell you. The difference is where it runs and what it can see. A Level 1 assistant built on Groundwork works from your actual data, inside systems you own, under a policy your team has agreed to. Someone pasting client emails into a free consumer tool is Level 1 with none of that, and it's how data leaks happen.
Why not just start at Level 3 and save money?
Because you don't yet know what "unusual" looks like for your business, and neither does the AI. The two months at Levels 1 and 2 are where you find out. Skipping them doesn't save money; it moves the cost to the day something goes wrong.
Can different tasks be at different levels?
They should be. A business might run confirmations at Level 3, invoice reminders at Level 2 and proposals at Level 1, all on the same foundation. The level belongs to the task, not the business.
Who decides when a task moves up?
You do. I'll bring the evidence and a recommendation. The decision, and the record of it, is yours.
What if we want full automation from day one?
Then I'm probably not the right fit, and I'd rather say so now. If you want Level 4 on a task with real consequences, there are vendors who'll sell it to you. I build the version where you can see what happened.
Start wherever you're comfortable.
Low commitment
Fixed price, one week, a scorecard and a roadmap you keep whether or not we go further.
Direct
No pitch. Worst case you leave with two or three things worth fixing.