The 30-60-90 Day Plan You Should Be Handed in Week One

Because week one is the last moment the engagement is cheap to walk away from.

Read the published 30-60-90 plans and every one of them starts its clock at day 30. Days 1 to 30 is one undifferentiated block, and it is always called discovery. Look at what that block actually commits anyone to and the answer is nothing. It has no output you can hold up, and it ends when the provider says it ends. And by the time it does, you have paid a month of a retainer to find out whether the person you hired can do the job.

Run the arithmetic on a normal fractional arrangement. Two days a week is about sixteen hours, so a thirty-day discovery phase is roughly seventy hours of somebody's attention. Seventy hours is not a survey of your business. It is a small project. It should produce something.

The thing it should produce is the plan, and the plan does not take seventy hours. It takes five days of asking the right people the right questions and then writing down what you heard. Discovery isn't a deliverable. It is a description of somebody reading.

There is a harder reason too. Week one is the only week where nobody has said anything yet that they need to defend. Ask for a ranked list on Friday and you get an honest one. Ask for it on day forty, after three leadership meetings have gone well and the sponsor has repeated a favorite idea twice, and you get a list with that idea near the top.

Everything else in an AI engagement is either an opinion or a system that does not exist yet. The plan is the one object that exists on day five, costs almost nothing to produce, and can be marked right or wrong later.

The published numbers on what companies actually have in writing are worse than most executives think. Grant Thornton's 2026 AI Impact Survey went to nearly a thousand US senior business leaders and found 78% who lacked full confidence that their organization could pass an independent AI governance audit within ninety days. Three in four boards had approved a major AI investment. Only 52% had set clear governance expectations to go with it, and just one in five had tested a response plan for an AI failure. More than half of executives, 51%, named strategy as the biggest driver of AI return. Only 22% of operations leaders had a fully developed and implemented AI strategy to point at.

A separate January 2026 survey of 365 senior leaders at companies with more than a thousand employees, run by TrendCandy for Larridin, found 92% of the C-suite fully confident in AI's impact on their business, 25% with governance fully implemented, and 62% who could not produce a complete inventory of the AI applications already running inside their own company. The single most cited barrier to measuring AI performance, at 58.2%, was unclear or fragmented ownership.

Put the two surveys side by side and not one of those gaps is a technology problem. Every one of them is a document nobody wrote. Who owns this. What counts as working. What we are not doing. Those are sentences, not systems, and they take an afternoon.

Five working days, five outputs, and a decision on Friday. None of this needs a discovery phase. It needs somebody to sit down with the right people and then write.

Monday is the kickoff, and its output is the out list. Ninety minutes with whoever signs the invoice. What is this engagement for, what does it not touch, and what would make you cancel it. The out list is the part people skip and it is the part that saves the quarter, because an engagement with no boundary drifts toward whatever the loudest department wants.

Tuesday is the systems walk. Not a survey and not a workshop. Sit next to the four or five people whose work is actually going to change, watch them do it, and write down every system they touch and who owns access to it. Two hours per person. You will find shadow tools nobody on the leadership team knows about, which is the point.

Wednesday is the ranked list. Every candidate workflow with a number against it: hours a week, items a month, or dollars. A ranking with no numbers is a preference, and preferences are what you already had before you hired anyone.

Thursday names the first build and writes its acceptance test. One workflow, the system it runs against, and a sentence describing what has to be true for it to count as working. The test gets written before the build, not after, for the same reason an eval gets written before you tune a prompt. Written after, a test is a description of whatever happened to come out.

Friday is the handover. The plan document, the cadence, decision rights, and the measurement baseline captured before anything changes. Then you do one of three things: sign it, send it back with the missing lines named, or stop.

The plan is twelve lines long, not twelve pages. If a line cannot be answered in two or three sentences, the answer is not known yet, and writing four paragraphs around it hides that.

1. Scope, in and out. Two lists. The out list is the load-bearing one. "We are not touching the billing system, the CRM migration, or anything customer-facing this quarter" is worth more than any amount of strategy language, because it is the only sentence in the document a stakeholder can be shown when they arrive in week six with a new idea.

2. The ranked opportunity list. Every candidate carries a number you could argue with. Six hours a week of manual re-keying. Eleven hundred invoices a month. Four days of month-end close. The ranking then has a stated rule behind it, usually some mix of hours saved and how contained the workflow is, and the rule is written down so a reshuffle later has to justify itself.

3. The first build, named. One workflow, named, with the system it runs against, rather than a theme, a category, or something like "finance automation". If the plan says the first build is "customer operations", nobody has decided anything and the first two weeks of month one will be spent deciding it at your expense.

4. The acceptance test, written before the build. A sentence a stranger could use to mark the evidence pass or fail. "At day thirty, the invoice path has a verified inventory entry, named owner, current baseline, approved data path, initial risk class and acceptance measures, so the gate can authorize one bounded validation or delivery path." That is testable without pretending the workflow is live. "Improve invoice processing efficiency" is not. This is the same discipline as writing the evals before you trust an agent, and a provider who resists writing it down is telling you they expect to negotiate the definition of success later.

These four are the ones that quietly decide whether the first build ships on time, and all four are usually missing.

5. The data and access checklist. Every system the first build touches, with a named human who can grant access and a date that access lands. Not a department. A person. Access is the single most reliable cause of a slipped month, and it slips because somebody wrote "we will get IT to sort it out" instead of naming the one person in IT who can grant it and the day they will.

6. Decision rights. Who approves a build going live, who is consulted before it does, and who can stop one mid-flight. Three names. The Larridin survey put unclear or fragmented ownership at the top of the list of barriers to measuring AI performance, at 58.2%, and this line is the cheap fix for it.

7. The measurement baseline. The current number, captured this week, before anything changes. How many hours, how many exceptions, how many days. Miss this and you cannot prove anything later, because the honest comparison is gone. Every argument about AI return that goes nowhere goes nowhere for this reason: nobody wrote down what it was like before.

8. The cadence. Which meetings happen, how often, who is in them, and what comes out of each one. A written cadence is what stops the engagement from turning into a monthly update call where somebody talks over a slide for forty minutes.

9, 10 and 11. Two readiness checks and three formal gates. The week-two readiness check confirms a verified inventory, named process owner, current baseline, approved data path, initial risk class, acceptance measures and external delivery owner. It assembles evidence; it does not authorize continuation.

The day-30 formal gate can authorize one bounded validation or delivery path after that inventory, owner, baseline, data, risk and acceptance evidence is recorded. It does not require a live or production workflow. The week-six readiness check then asks whether representative approved cases have been exercised in controlled validation, explicitly not production, with documented quality, cost, review, failure and access evidence.

The day-60 formal gate decides whether the workflow enters bounded production using evaluation results, controls, an exercised runbook draft, a rollback path and operating-owner evidence. The day-90 formal gate accepts, accepts with a dated condition, redesigns or stops operation, then funds, defers, rejects or investigates the next-quarter portfolio. Each gate has a written consequence. We covered what tends to slip through them in our record of what actually ships in the first ninety days.

12. Who does the work. Named people and their hours. "A senior engineer" is not a name. If the plan says the provider will bring in delivery capacity, the plan says who, for how many hours a week, starting when. This is where the difference between an advisor and an operator becomes visible on paper, and it is a question worth settling before you sign rather than in month two.

And the exit. What you keep, where it lives, and what happens on the last day. Repo access, credentials, a runbook, and the name of the person inside your team who owns the thing after the engagement ends. An AI system with no internal owner is an outage waiting for a quarter with no retainer in it.

That is the whole template. Twelve lines, one page, five days.

Four rhythms carry a ninety-day engagement. Any more and the provider is spending your retainer talking to you about the work instead of doing it.

The deliberate one is the fortnightly written update, which has no meeting attached. Writing forces a position. A person who has to put "the invoice build is two days behind because access to the ERP sandbox landed on Thursday instead of Monday" in a document will not write it three times before somebody fixes access. The same person in a call will say it is going well.

And the monthly gate is short on purpose. Forty-five minutes, three people, one decision, recorded in writing afterwards. A gate that runs an hour and ends with everyone feeling positive isn't a gate.

Half of scoring a plan is spotting the filler, and the filler is always the same five things.

A maturity score with no decision hanging off it. Levels one to five, a diagram, a position on the diagram, and nothing that changes as a result. If the score does not move a build up or down the ranked list, delete it.

Tool selection in week one. Nobody knows enough on day four to pick the platform, and a plan that picks one has usually picked whatever the provider already resells.

A full governance framework before the first build exists. Rules written against an imaginary system are fiction, and they get rewritten the first time a real one behaves in a way the document did not imagine. Write the two rules the first build actually needs, then extend them when there is something to govern. Deloitte's 2026 State of AI in the Enterprise, which surveyed 3,235 leaders across 24 countries, found only one in five companies with a mature governance model for autonomous agents, and only a quarter had moved 40% or more of their experiments into production. Both numbers are low, and neither one improves by writing more rules before there is anything to govern.

Headcount plans. In week one, a proposal to hire two engineers is a proposal to spend your money on a problem nobody has scoped.

And a deck as the artifact. If the plan exists only as slides, it exists only inside the meeting where the slides were shown. Ask for the document.

Twelve checks, one point each, and a maybe counts as a no.

Eleven or twelve. Sign it. You have a provider who has done this before and is comfortable being marked against their own words.

Eight to ten. This is the common score and it is not a rejection. Name the missing lines in an email and ask for them by the end of the following week. A competent operator fills four gaps in a day or two, because the gaps are usually things they know and did not think to write. Somebody who needs three weeks to name the first build did not have one.

Under eight. You have bought a reading week. That is a real thing to buy and it is occasionally the right call, but then it gets priced and scoped as a paid assessment with a fixed end date, rather than sitting inside month one of a retainer that renews by default.

One useful test underneath the score. Take the plan, cover the provider's logo, and hand it to somebody in your finance team who knows nothing about AI. Ask them what will be different in thirty days. If they cannot tell you, the plan was not written for you. It was written for the next meeting.

Most people find this page in week three, not week zero.

Ask for it anyway, and frame it as a summary rather than a renegotiation, because that is what it is. "Can you put the first build, its acceptance test, the access owners and the day 30 condition in one document by Friday." Everything in that sentence is something you have already agreed to verbally or already paid for.

What comes back tells you what you bought. A document by Friday means you have an operator who was carrying it in their head. A document in three weeks means the decisions had not been made and are being made now, which is recoverable if the ranked list arrives with numbers on it. A deck means you should have the gate conversation early rather than at day thirty.

If nothing useful comes back, use the AI ownership and opportunity assessment capability to frame the fit, prioritization, and evidence questions that remain unresolved. Then require any provider to state a separate delivery scope and its acceptance evidence in writing.

We publish this template because a standard nobody can check you against is marketing. If our own fractional AI leadership engagement turns up on your Friday with nine out of twelve, that is a fair thing to say to us in the room, and the plan is the reason you can say it with a number attached.

The honest pricing picture separates ownership from delivery. CAIO Core is From $5,000/month for ongoing executive AI ownership, with builds separately scoped. CAIO + Delivery is From $10,000/month and adds an ongoing delivery pod with one active build stream. A Starter build is $1,500 to $2,500 fixed, and a Production sprint is $5,000 fixed for one full workflow live in about two weeks. The whole ladder is on our pricing page; the project offers are alternatives for companies that do not need ongoing executive ownership.

There is an anti-fit case here and it is a real one. If you already know the one workflow that is eating your week, you do not need a plan or a fractional AI officer to rank a list of one. Buy the build. If you already have an internal head of AI, this template is theirs to run, and handing it to an outside provider instead just adds a layer between the person accountable and the work. And if the real problem is that three departments each bought their own tool and nobody can say who owns the outcome, a plan will not fix that on its own, though writing line six down is a better first move than another vendor evaluation.

This plan is scoped to a non-software company where no employee writes software, firmware, or embedded code. It assumes executive ownership inside the company and externally delivered implementation when a build is approved. It is not an engineering plan for a software product, an AI startup, or a company with an internal code-writing team.

The plan is an operating specification, not a claim about a customer. Every baseline, gate, workflow, owner, and acceptance threshold must be filled with evidence from the company using it.