The first 90 days decide whether AI becomes a real capability in your company or an expensive experiment nobody mentions again. The second outcome is common: Gartner predicts at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs, or unclear business value (Gartner, 2024). A plan is how you stay on the right side of that number. Written for an owner, not a tech department: what to do in each 30-day block, a worked example, and the traps we see most.
Picture the 90 days as a path with three milestones. Find it, build it small, then decide.
timeline title The first 90 days of AI Days 1 to 30 : Find the right first project Days 30 to 60 : Build it small and prove it Days 60 to 90 : Measure, learn, and decide
Days 1 to 30: find the right first project
Most failed AI projects were lost in the first month, when someone picked the wrong thing to build. The right first project is the one that clears four gates, in order.
flowchart TD
A(["Candidate process"]) --> B{"Does it eat real hours or money every week?"}
B -- No --> X(["Not worth being first"])
B -- Yes --> C{"Can one team and one process contain it?"}
C -- No --> X
C -- Yes --> D{"Can a wrong answer be caught and fixed cheaply?"}
D -- No --> X
D -- Yes --> E{"Does the data already exist in a system you control?"}
E -- No --> Y(["Fix the data first"])
E -- Yes --> F(["Your first 90-day project"])
- Walk the business, not the vendor deck. Spend two weeks with the people doing the work. Which emails get the same answer every day? Which report gets rebuilt by hand each month? You will have ten candidates.
- Run every candidate through the gates. Support inbox triage, supplier invoice extraction, and first-draft proposals usually pass. Pricing, hiring, and anything customer-facing with no human check usually do not.
- Get honest about the data. If the invoices are scanned PDFs in someone’s inbox and the customer history lives in three spreadsheets, sorting that out is your starting work, not a step you skip. Data quality is the most-cited reason AI projects stall, per the same Gartner release. Our AI readiness check covers what “ready” means.
- Write the success number before you build anything. “Cut first response on routine support emails to under 30 minutes” beats “improve service.” Then measure the baseline now, for a week, by hand if you must; without it you prove nothing on day 90.
What you hold at day 30: a one-page scope note with the process, an owner from the team that does the work (not IT), the success number, the baseline, the data source, and what the AI may and may not do. If you cannot fill that page, you have not found your project.
Days 30 to 60: build it small and prove it
Month two is where the temptation to grow the project hits hardest. Resist it. This phase exists to learn how AI behaves on your data.
- Build on your real data, not a tidy sample. Real invoices have stamps and handwriting; real customer emails mix Arabic and English in one sentence. A demo on clean data tells you nothing. We wrote up the discipline in how to run an AI proof of concept that leads somewhere.
- Keep the boundary you set on day 30. One process, one team. When someone says “could it also”, park the idea for project two and say no.
- Put a human on every output at first. For the first two weeks live, a person reviews everything before it goes anywhere and keeps a review log: what the AI did, whether the reviewer changed it, and why. Once the log shows what it gets right, the human moves to exceptions only.
- Sort the rules out now, not on day 89. If customer or employee data touches the system, PDPL applies. It took effect on 14 September 2023, full compliance was required by 14 September 2024, and SDAIA is the regulator (Morgan Lewis, 2024). In practice that means a one-page data note: what personal data goes in, where it is processed (in-Kingdom or not, which decides whether the cross-border transfer rules apply), who sees the outputs, and how long anything is kept. Settle it with your compliance person before real data flows.
- Write down what the AI may and may not do. It may draft; it may not send. It may flag an invoice; it may not approve a payment.
Days 60 to 90: measure, learn, and decide
- Measure against the day-30 number, and only that number. Did first response time fall? By how much? What did it cost to run, including the reviewer’s hours? The math is in how to measure ROI on AI.
- Read the review log for patterns. A handful of input types usually cause most errors. Some are fixed with better instructions; some need a permanent rule sending that case to a person.
- Make one of three calls. Expand it, fix it (another 30 days on the specific failures), or stop it. All three are good outcomes; stop is the one owners find hardest and the one that protects the budget.
- Write the one-page decision memo. Result against target, cost to build, monthly cost to run, verdict, and what the next project would be. It becomes the template for every project after.
A worked example: 90 days on a support inbox
Say a distributor gets 1,500 customer emails a month, about half of them order status, delivery windows, and invoice copies. Two people handle the inbox, and first response averages five hours in busy weeks. The numbers are made up to show the shape.
The day-30 scope note says: routine categories under 30 minutes, humans on everything else. From day 30 to 60, an assistant drafts replies for the routine categories from the order system, and a person reviews every draft. The log shows order-status drafts are right about nine times in ten, and delivery-window drafts much worse because the logistics data behind them is stale. The fix is a rule: delivery windows always go to a person.
By day 90 the routine categories are answered in around 20 minutes, and the two people spend their time on complaints and larger accounts. Run cost: model usage plus roughly ten hours a month of spot checks. The decision memo says expand, with the logistics data feed as project two. What made it work: narrow scope, a measured baseline, a log showing where it fails, and a decision with a number behind it.
Where the first 90 days go wrong
- The demo that never meets real data. A polished vendor demo in week two, and you do not see it fail until day 70 on your own documents. Insist on your data in the demo.
- Scope that grows quietly. Three processes by day 45, none finished by day 90. The most common way a 90-day plan becomes a nine-month one.
- Treating compliance as a day-89 checkbox. Real customer data in a foreign-hosted tool with no data note. Fixing that on day 35 is a one-page conversation; fixing it after the fact is expensive.
- Mistaking usage for value. McKinsey found that while a majority of organizations use generative AI, only ~39% see enterprise-level EBIT impact, and only ~5 to 6% are “AI high performers” (McKinsey, 2024). The day-30 target and the day-90 memo exist to close that gap.
What day 90 should actually look like
Not a transformed company. If someone promised you that, treat it as a warning sign. Success is one working AI capability delivering measurable value against the day-30 number, plus a scope note, a data note, a review log, a decision memo, and a team that has done this once. In our experience the second project takes closer to 60 days than 90, because the paperwork exists and the people know what “done” means.
How to start on Monday
The pattern is boring and it works. Pick the one process that annoys you most and clears all four gates. Measure it for a week, by hand. Pilot it with a person reviewing every output. Measure the net result, not the gross, and make the call. Then scale the pattern, not the tool: the next project reuses the same documents with a new process filled in. For a second pair of eyes on the first pick, that is what our AI proof of concept service is for.
Want a clear plan for your first 90 days with AI? SDCG helps owners go from “we should do something with AI” to a working, valuable first project: scoped, built, measured, and ready to build on. We’re independent, so the plan is built around your business, not a tool we’re selling. Book a free 30-minute review.