The short answer.
Do not give AI three kinds of work on your first plan: work that happens only a few times a year, work whose whole value is one person’s judgment in the moment, and work where a mistake cannot be caught before it leaves the building. Everything a person decides stays with a person, in every workflow, at every stage. To sort the rest, ask four questions of each piece of work: does it repeat, can it be written down, is it checked before it goes out, and does a named person decide?
Work that passes all four goes on the plan. Work that fails one gets help from the assistant one question at a time, or waits for next quarter’s review.
Three kinds of work to leave off the first plan.
The first plan leaves out rare work, in-the-moment judgment and work whose mistakes go out unseen, each for a different reason.
- Work that happens a few times a year.A workflow earns its keep by running the same way every week or month. The annual budget or the yearly insurance renewal runs too rarely to test well, and by the next time it comes round the inputs have changed. The assistant helps with rare work as a one-off conversation.
- Work whose whole value is one person’s judgment in the moment.A negotiation on a live call, a hard conversation with an employee, the read of a room in a client pitch. The value of that work is the person, live. A workflow prepares the brief before the call and drafts the notes after it. The moment itself stays with the person.
- Work where a mistake cannot be caught before it leaves the building.A message that goes straight to a customer, a payment that leaves the account, a record changed with no one looking. If nobody reviews the output before it goes out, a miss reaches the customer first. Add a review step and the work often passes; until then, leave it off.
Each of the three is a poor first workflow, and there is plenty of work that repeats. Repeatable work and ad hoc work has the sheet for sorting a team’s week.
What stays with people in every workflow.
In every workflow, a person makes every decision: what to approve, what to send, what to pay and what to change. A workflow prepares, checks and drafts. Anything that sends, pays or changes a record waits for a person to approve it.
Write that line into each workflow by name. “The controller approves every posting.” “The account lead sends every client email.” “The office manager signs off every schedule change.” A workflow designed to pull from your accounting system and push data back does so with guardrails and a person approving. Every workflow runs in your company’s own ChatGPT, Claude, Copilot or Gemini, on your account, run by your people.
Three things stay with people outright, whatever the workflow:
- Decisions about a person.Hiring, pay, performance and anything that affects someone’s job.
- Commitments on the company’s behalf.A price, a contract term, a promise to a client.
- The call when the facts are unclear.When the inputs disagree or something looks wrong, the workflow flags it and a person decides.
The four-question test, as a sheet to copy.
The four-question test decides whether a piece of work goes on your first AI plan. Copy the sheet below into a document, one row per piece of work, and answer yes or no. Four yeses put the work on the plan. Any no keeps it off for now.
- Does it repeat?The work happens at least once a month, the same way each time. Yes or no.
- Can it be written down?Your best person can say what they start from, the order they work in, what they check and what good looks like. Yes or no.
- Is it checked before it leaves the building?One person already reviews the result, or a review step can be added, before anything reaches a customer, a supplier or a record. Yes or no.
- Does a named person decide?You can write one name next to the decision in the work: who approves, who sends, who signs. Yes or no.
For each row, fill in:
- Piece of work._______________________
- Answers to the four questions.___ ___ ___ ___
- What happens next.On the plan / Ask the assistant / Revisit next quarter
Work with four yeses then gets scored for order. Choose your first AI wins has the scorecard for deciding which goes first.
What to do with work that fails the test.
Work that fails the four-question test goes one of two ways: the person who does it uses the assistant ad hoc, or the plan revisits it next quarter.
- Use the assistant ad hoc.Rare work and in-the-moment work get better help from ChatGPT, Claude, Copilot or Gemini one question at a time: a first draft of the annual letter, a list of questions before a hard call, a summary of last year’s file. The person asks, reads and decides. Nothing is built.
- Revisit next quarter.Work that failed on question three or four often passes once someone adds a review step or names an owner. Put it on the plan’s “waits” list with the reason it failed. At the quarterly review, ask again.
The “waits” list shows the team the work was seen and set aside on purpose. We have a plan. How do we build the first AI workflow? takes the work that passed and turns it into a tested workflow.
One company runs the test.
Illustrative example.
Take a 45-person property management company with three teams: leasing, maintenance and accounts. The operations director runs the four-question test with each team lead.
In leasing, the weekly vacancy report passes all four: it repeats, the leasing lead can describe it, the lead reviews it and the lead decides what goes to owners. It goes on the plan. Lease negotiations with commercial tenants fail question two, because the value is the lead’s read of each tenant on the call. The lead keeps using the assistant ad hoc to prepare for each call.
Maintenance lists the work-order triage. It repeats daily and can be written down, but today the dispatcher sends vendors straight from the inbox with no review. It fails question three. The team adds a step where the maintenance coordinator approves each dispatch, and triage goes on the “waits” list for next quarter.
Accounts lists the yearly owner statements for tax season. They fail question one and stay with the assistant ad hoc. The monthly owner statement passes all four, and the accounts manager approves every one before it is sent. It goes on the plan.
The plan ends the week with two workflows, one item waiting and two pieces of work kept with people on purpose.
What is different in a company of under 200 people.
In a company of under 200 people, the four-question test gets answered in one meeting, and three things change how you read the answers.
- The rare work sits with the same few people.The owner does the budget, the operations director does the renewals. Their rare work stays with the assistant ad hoc; their weekly work goes on the plan first.
- One person often sends without a second look.Large companies have review layers built in. A smaller company often has one person doing the work and sending it. Question three catches that, and the fix is one line naming who reviews.
- The judgment work is often the founder’s.The founder’s read of a client or a deal is the company’s edge. Keep that work with the founder, and give the founder workflows for the preparation around it.
Most published advice here is written for large enterprises with risk committees. A company your size settles it with a sheet and one named person; Who should own AI in a company your size? says who.
The four-question test is one part of the first step in How organizations plan their use of AI: plan it, develop the workflows, deploy them. FullyPowered Planner is where the plan lives: you set out your teams, their roles and the work between them, and choose what to build first, with the waiting work kept beside it.