What to hand to AI first: an Ask, Map, Move worksheet for owners

Hand AI first the work that happens often, follows clear rules, is easy to check and can be reviewed before it reaches a customer; keep one-way, judgment-heavy decisions for last.

That is the whole answer. The rest of this guide is how to apply it to your own business without guessing: a short scoring test, then a worksheet built on the Of One method (Ask, Map, Move), worked through on an example company.

Why the order matters

AI is not uniformly good or bad at work. It is very good at some tasks and quietly bad at others, and the two can look alike.

A 2026 paper in Organization Science calls this the “jagged technological frontier.” In a preregistered experiment with 758 knowledge workers at Boston Consulting Group, people using GPT-4 on tasks within its abilities completed 12.2% more tasks and finished 25.1% faster, with significantly better quality. On a task chosen to fall outside those abilities, the people using AI were 19 percentage points less likely to reach a correct answer than those working without it.

A second study points the same way. Across 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers, with minimal effect on experienced and highly skilled workers (Brynjolfsson, Li and Raymond, NBER).

Two lessons for an owner:

• Start where the frontier is known. The experiment’s gains were on tasks inside GPT-4’s abilities. A task that looks similar can still fall outside them.
• Put the check where the risk is. When the task might sit outside the frontier, a person reviews before anything leaves.

Most small businesses have not started. In a Census Bureau report published in May 2026, between 17% and 20% of US businesses said they used AI over December 2025 to May 2026, and use was under 20% among firms with four or fewer employees. Picking the right first task is how you start without an expensive mistake.

The five-question score

Score each candidate task from 0 to 2 on five questions. Higher totals go first.

Question	0	1	2
How often?	A few times a year	Monthly	Weekly or more
Clear rules?	Mostly judgment	Some rules, some judgment	A written procedure, template or price list exists
Easy to check?	Hard to tell if it is right	Needs an expert to check	Anyone who knows the job can check it quickly
Reviewable before it acts?	Acts on its own, hard to undo	Reaches a customer, but can be corrected	Stays a draft until a person approves it
Inside the frontier?	Novel judgment, little precedent	Mixed	Drafting, summarizing, sorting, following a procedure

Break ties by time freed for your best people, and by closeness to revenue.

A task that scores low is not banned forever. It is a later task, after you have a working check and some evidence.

Worked example: a 25-person service business

Illustration only. This is a made-up company. It is not a client, and every number below is an example input, not a measured result.

The example is a 25-person commercial cleaning company. Twenty people work on crews. Five run the office: the owner, an operations manager, a scheduler, an account manager and a bookkeeper. The owner wants to know what to hand to AI first.

Step 1: score the candidates

Candidate task	Often	Rules	Check	Review	Frontier	Total
Turn walkthrough notes into draft bid proposals	2	2	2	2	1	9
Summarize crew inspection checklists and photos into client reports	2	2	2	2	1	9
Draft payment reminders for overdue invoices	1	2	2	2	2	9
Answer and route inbound service requests	2	1	1	1	2	7
Build the weekly crew schedule	2	1	1	2	1	7
Screen job applicants	1	1	1	0	1	4
Set prices for a new type of contract	0	0	0	1	0	1

Three tasks tie at 9. The tiebreakers decide it. Proposals take the operations manager hours every week, and every proposal is a chance at revenue. Proposals go first. Inspection reports go second.

Applicant screening scores low for good reason: it is hard to check for fairness, and hiring rules vary by place. This guide is education, not HR or legal advice.

Step 2: Ask

Ask is the first part of the Of One method. On of1.ai it starts with the eight starting questions, before anything is built, with a warning: “Every question you skip becomes a guess treated as a fact.”

#	Question	The example’s answer
1	Who makes the decision?	The operations manager. One person, not “the office.”
2	What outcome do they want?	Every proposal sent within one business day of the walkthrough, with no pricing errors
3	What is true right now?	Six to eight walkthroughs a week. Each proposal takes about 90 minutes. Proposals go out three to five days after the visit. Prices live in one spreadsheet.
4	What are the hard limits?	No price leaves without the manager’s approval. Client details go only into approved tools. The template’s contract terms do not change without the owner.
5	What can they actually change?	The walkthrough checklist, the proposal template, what the AI can see, and the review step
6	What will shift the outcome?	If walkthrough notes follow a checklist and the AI drafts from the template and price sheet, then the manager’s work drops to review, and proposals go out sooner.
7	How will we know it worked?	Days from walkthrough to sent proposal; minutes of review per draft; pricing errors caught at review
8	When must they decide?	After 20 proposals or four weeks, whichever comes first: keep, fix or stop

Step 3: the Map

The Map lays out what you know before you build. An Of One map has ten main kinds of parts; this example uses seven of them.

Part	The question it answers	The example’s entry
Fact	Where did this come from, and is it still fresh?	Walkthrough volume comes from the job log. The price sheet was last updated this quarter.
Claim	What do we believe, and how sure are we?	The AI can produce a usable draft from checklist notes. Sureness: medium.
Unknown	What do we not know yet?	Whether faster proposals win more work. Whether photos improve the drafts.
Kill test	What result would prove us wrong?	If review still takes more than half the old time after 20 drafts, the claim fails.
Option	What could we actually do — and can we undo it?	Draft-only with human send (easy to undo), or automatic sending (hard to undo). Choose draft-only.
Trigger	What new fact would reopen the question?	A new service line, a price sheet change, or a new AI model
Human gate	Which moves need a person to say yes?	Every price and every sent proposal. The manager approves both.

The Map already made one decision: the AI drafts, and a person sends. That single choice removes the largest risk.

Step 4: Move

Only now does anything get built. Move means taking the step the map supports, with anything you cannot undo held at a human gate. For a build like this one, this guide uses five steps.

1. Define outcome. A draft proposal in the company template, with prices pulled from the current sheet, waiting for review within an hour of the walkthrough notes arriving.
2. Prototype fast. Build the smallest version that tests the riskiest guess: can the AI read real checklist notes and pick the right line items? Use ten past walkthroughs whose correct proposals you already know.
3. Instrument. Log the three signals named under Ask: days to send, review minutes, errors caught.
4. Iterate. Fix what the numbers show. If drafts keep missing the same line item, fix the checklist, not only the prompt.
5. Deploy. Switch over when the kill test passes. Keep the human gate. Move to the second task, inspection reports.

What not to hand over first

Some work should wait until you have working checks and evidence, or should stay with people:

• Prices and contract terms. AI can draft; the owner decides.
• Hiring, firing and discipline. High stakes, hard to undo, and regulated.
• Payroll and tax filings. A system can prepare; a qualified person files.
• Anything that moves money or data out of the business without a person’s approval.
• Safety decisions on the job site.
• The one relationship that pays the bills. Your biggest client talks to a person.

Common mistakes

• Starting with the most painful task. Pain is not the same as fit. The most painful task is often the most judgment-heavy.
• Automating an undocumented process. If nobody can write down the steps, the AI cannot follow them. Write them first.
• Letting it act before it drafts. Run in draft mode until the kill test passes.
• No owner. A system without one named person drifts.
• Measuring hours saved on paper. Measure work finished and errors caught.

The blank worksheet

Copy this for your own first task.

1. Candidates: list five to ten recurring tasks. Score each on the five questions.
2. Pick: take the top score; break ties by time freed and closeness to revenue.
3. Ask: who decides, what outcome, what is true, hard limits, levers, what will shift it, how you will know, when you decide.
4. Map: facts with sources, your main claim and sureness, unknowns, a kill test, reversible vs one-way options, triggers, human gates.
5. Move: take the step the map supports and keep the human gates; to build it, define the outcome, prototype on past examples, instrument, iterate, deploy.
6. Next: after the first task works, pick the second. After four or five, redesign the roles around them. The AI-first company, for owners shows what that looks like.

The full method is published at of1.ai, and you can walk one decision through it with the free tool at of1.ai/ask. For how Utlyze runs this kind of system on its own work, and what broke, see What is a business of one?.

If you would rather work through the worksheet with someone who has done it, Business of One runs Consult blocks at $400 per hour in 5-hour blocks, and builds with owners for $15,000 a month, with build-with-you coaches and a shared 24/7 on-call rotation.