METAHEURISTIC
In the first 90 days, a fractional Chief AI Officer must do four things. Find out how the company uses AI today. Select one business problem and measure it. Put one AI system into production for that problem, with evals, monitoring and cost limits. Write the rules that the next system needs. At the end of day 90, you must see a system, a number and an owner.
Many sellers describe a 90-day plan in three steps: diagnose in month one, implement in month two, measure in month three. The steps are correct. But the plan often stops at words. This page gives the details: what happens each week, what you must see at each checkpoint, and when to stop.
I use a plan like this one in my own fractional AI CTO work. It is not the only correct plan.
What must exist on day 90?
Write this list into the contract before day 1. It is the definition of success.
If the leader cannot agree to this list, ask why. Sometimes the reason is good. For example, the data is not ready, or the first process needs a long legal review. Then agree on a different list. But do not accept “a strategy” as the result of 90 days.
Before day 1: preparation
Do these things before the first day. They save one or two weeks.
- Give access. Accounts for the email, the documents, the code repositories, the cloud and the AI tools that the company pays for.
- Name one internal owner. This person works with the fractional leader each week. This person will own the system on day 90.
- Book the interviews. The leader must speak to the CEO, the CTO, and the heads of the two or three departments with the most manual work.
- Collect the AI bills. All invoices from model providers, AI tools and platforms for the last three months.
Days 1 to 30: inventory and selection
Week 1: Interviews and access
The leader speaks to the key people. The questions are simple:
- Which tasks take the most hours each week?
- Which tasks have the most errors or the most customer complaints?
- Which AI tools do you use? What data goes into them?
- What did you try with AI that did not work?
The leader also reads the logs and the setup of any AI system that already exists.
Week 2: AI inventory
The leader delivers a list of all AI use in the company. For each item:
- The tool or system.
- The team that uses it.
- The data that goes into it.
- The monthly cost.
- The owner, if there is one.
This list often surprises the management. Employees use many tools with their own accounts. Customer data goes into free chat tools. Nobody knows the total cost.
Week 3: Select the first problem
The leader makes a short list of three candidate problems. For each one, the leader estimates:
- The current cost in hours or money per month.
- The data that a system needs, and if the data is available.
- The risk if the system gives a wrong answer.
- The time to put a first version into production.
The leader recommends one problem. A good first problem has a high cost, available data, a low or controlled risk and a short build time. Customer-facing systems with high risk are usually not a good first problem.
Week 4: Measure the baseline and decide
The leader measures the current process for the selected problem. For example: “The support team answers 400 tickets each week. The average time per ticket is 11 minutes. 30% of the tickets are the same 20 questions.”
At the end of week 4, you have the first checkpoint.
Checkpoint 1 (day 30): You must see the AI inventory, the selected problem with a measured baseline, and a plan for the system. If you only see a strategy document, stop or reset the engagement now.
Days 31 to 60: build
Week 5: Design and eval set
The leader designs the system with the internal owner. The design shows the data flow, the model, the tools that the system can use and the points where a human approves an action.
In the same week, the leader and the owner build an eval set. This is a set of real examples from the process, with known good answers. For a support system, it can be 100 to 300 real tickets with good replies. The eval set is the most important technical asset of the project. It shows if the system is good enough, and it tests each later change.
Weeks 6 and 7: First version
The team builds the first version. The leader builds it, the internal team builds it, or an agency builds it under review. Each change runs against the eval set.
The leader also sets up:
- Monitoring. Traces of each request, and alerts for errors and high cost.
- Cost limits. A budget per day or per month, with an alert.
- Access control. The system can only read and change what it needs.
Week 8: Internal test
A small group of real users tests the system in their daily work. The leader collects every failure and adds it to the eval set.
Checkpoint 2 (day 60): You must see a working system with real users, an eval score, a cost per task and a list of open problems. If the system is still a demo, ask what blocks it. Data access and approvals are the usual reasons.
Days 61 to 90: production, measurement and rules
Weeks 9 and 10: Launch
The system goes to all users of the process. The team watches the monitoring each day. The leader and the owner fix the most common failures.
Week 11: Measure
The leader measures the same number as in week 4. For example: “The average time per ticket is now 6 minutes. The 20 common questions get a draft answer that agents approve in one minute.”
The leader also measures the cost: the model cost, the tool cost and the hours of the team. The result can be negative. That is acceptable. A negative result after 90 days is better than an unknown result after a year.
Week 12: Rules and handover
The leader writes the rules that came from the work:
- An AI usage policy. Which tools are approved, for which data.
- An approval path for new AI tools, which takes days, not months.
- A short runbook for the system: how to run it, how to change it, how to stop it.
The internal owner runs the system alone for one week. The leader only watches.
Week 13: Review
Checkpoint 3 (day 90): You must see the system in production, the before and after numbers and the cost. You must also see the internal owner who runs it and the written rules. Then decide: stop, continue with the next system, or reduce the hours.
What goes wrong in a 90-day plan?
These are the most common problems. Each has a fix.
| Problem | Fix |
|---|---|
| Access takes weeks | Give access before day 1. Make one person responsible for it. |
| No internal owner | Do not start without one. Without an owner, the system dies after the leader leaves. |
| The first problem is too large | Select a smaller part of the process. One team, one type of request. |
| The data is not ready | Select a different first problem, or make data cleanup the first deliverable. |
| Legal review blocks the launch | Start the legal review in week 3, not in week 9. |
| Nobody measured the baseline | Measure it now, even late. Without a baseline, you cannot show a result. |
| The leader writes documents but does not build | Ask for the eval set and the first version by day 45. |
How much of the leader’s time does this need?
This plan needs about one to two days each week from the fractional leader. It also needs about two days each week from the internal owner. If the leader also builds the system alone, it needs more time.
A leader with only a few hours each month cannot do this plan. With a few hours, the leader can do the inventory and the selection. The build needs more time from someone.
What must the weekly update contain?
Ask the fractional leader for a short written update each week. Five lines are sufficient. The update keeps the work visible, and it gives you a record for the exit.
- Done this week. What changed: documents, systems, decisions.
- Numbers. The eval score, the cost per task and the usage, when the system exists.
- Decisions made. Each decision with one sentence of reason. These go into the decision log.
- Decisions needed. What the leader needs from you, with a date.
- Risks. What can delay the plan, and what the leader does about it.
If the update has no numbers after week 6, ask why. If “decisions needed” stays the same for three weeks, the problem is on your side. Fix it.
What happens after day 90?
There are three common paths.
Next system. The first system works. The leader starts the same cycle for the second problem. The second cycle is faster, because the rules and the tools exist.
Reduce the hours. The internal team can now run AI work alone. The leader stays for a short monthly review of cost, incidents and new tools.
Stop. The work is done, or the result was not good enough. Use the exit plan guide to get a clean handover.
Can this plan work for a company with no engineers?
Yes, with changes. A company with no engineers has two options for the build:
- Use an AI tool that the company already pays for, and configure it. Many support, sales and document tools now include AI features. The leader configures and tests them with the eval set.
- Hire an agency for the build, and the leader reviews the work. The agency proposal review guide helps with this.
In both cases, the internal owner does not need to code. The owner needs to understand the process, read the monitoring and run the eval set.
Next step
Copy the four day-90 items into your contract. Book the week-1 interviews before you sign. If you want a fractional leader who builds the first system, read how my fractional AI CTO engagement works. If you have not selected a leader yet, read how to hire a fractional Chief AI Officer.
