Implement · Chapter 6
A 90 day AI implementation roadmap
Ninety days is enough to take one task from idea to something running in production. It is not enough to transform a company, and plans that promise otherwise are the ones that end with nothing in production at all. This roadmap covers one task, one owner, and one decision gate where stopping is a planned outcome.
Key takeaways
- 01Ninety days is enough to take one task from idea to production, and not enough to transform a company. Plan for the first.
- 02Days 1 to 14 are measurement only. No tool is chosen until the current task is timed and written down.
- 03There is one gate, at day 45: if the output is not good enough to use, you stop rather than extending.
- 04Embedding the tool where work already happens is the step that decides whether anyone still uses it in month six.
- 05The project ends with a named internal owner and written documentation, or it ends when the person who built it gets busy.
Before day one
Two things must be settled before the clock starts. You need one named task, specific enough to state in a sentence, and one named owner with the authority to change how that work gets done. If either is missing, the 90 days will be spent discovering that, which is an expensive way to learn it.
If you are unsure whether your task qualifies, the readiness assessment scores it out of 20 across the task, the data, the people, and the budget. Below 9 the roadmap will not help, because the constraint is the process rather than the technology.
The five phases
Measure, do not buy
- □Time the task as it is performed today, several times, on real work
- □Count how often it happens per week or month
- □Write the steps down and have the people who do it correct the draft
- □Collect ten examples of good output and a few bad ones
- □Tell the team what you are doing and why
No tool is selected in this phase. If you find yourself comparing vendors in week one, you are building a solution before you have measured the problem.
Trial against real work
- □Pick the smallest tool that could plausibly shorten the task
- □Run it in parallel with the human version, on live work
- □Record how often the output is usable without edits
- □Note which kinds of input it handles badly
- □Keep the old process running untouched throughout
Running in parallel costs more in the short term and is worth it. Switching over before you know the error rate is how confidently wrong output reaches a customer.
The gate
- □Is the output good enough to use with the checking you can afford?
- □Did the time actually drop against the baseline?
- □Do the people doing the work want to keep it?
- □Decide: continue, change approach, or stop
Stopping here is a successful outcome, not a failure. Six weeks spent learning that this task is a poor candidate is far cheaper than six months.
Embed it where the work happens
- □Put the output into the system people already open every day
- □Remove the old way, or it will quietly continue in parallel
- □Write the short version of what to do when the output looks wrong
- □Set the checking routine and name who does it
This is the phase most failed projects skip. A tool in a separate tab competes with habit, and habit usually wins.
Hand over and measure
- □Name the internal owner and confirm they can run it without help
- □Document the setup in a form someone else could follow
- □Re-measure hours consumed and compare to the day one baseline
- □Decide where the recovered capacity goes
- □Choose whether there is a second task worth doing
If only one person can operate what was built, you have a dependency rather than a capability. Handover is part of the project, not an optional extra.
Why the gate matters more than the tool
MIT's Project NANDA found that 95 percent of generative AI pilots delivered no measurable profit-and-loss impact. Read the failures closely and a shape repeats: the pilot never formally ended. It was not cancelled and it did not reach production. It simply thinned out until the licence lapsed, and because nothing was ever concluded, nothing was learned that improved the next attempt.
A dated gate solves this cheaply. At day 45 you either have output good enough to use or you do not, and you decide in front of the numbers rather than in front of a sunk cost. Teams that build a stopping point into the plan stop earlier and start again sooner, which over a year produces more working systems than teams that treat stopping as an admission of failure.
It also changes how the project is approved. An owner asked to fund an open-ended AI initiative is being asked to trust a category. An owner asked to fund six weeks with a defined decision point is being asked to fund an experiment, which is a much easier thing to say yes to.
Before you start, check the privacy questions
If the task touches customer or employee information, read AI risk, privacy, and policy before day one rather than at day 60. PIPEDA obligations attach to personal information regardless of which tool processes it, and finding that out during the trial usually means restarting with a different setup.
Frequently asked questions
How long does it take to implement AI in a small business?+
One narrow task with an off-the-shelf tool fits comfortably in 90 days from first measurement to running in production. Anything requiring system integration, data cleanup, or changes to how several teams work runs to two or three quarters. The 90 day shape below assumes one task and one owner, which is the version most likely to finish.
What should happen in the first two weeks?+
Measurement and nothing else. Time how long the task takes, count how often it happens, write down the steps as they are actually performed, and collect examples of both good and bad output. Choosing a tool in week one is the most common way to spend three months proving that the wrong task was automated well.
Why is there a decision gate at day 45?+
Because without a fixed date to stop, projects continue by inertia rather than by merit. At day 45 you have run the tool against real work and know its error rate. If the output is not good enough to use, extending rarely fixes it, and the honest move is to stop with six weeks spent instead of six months. A gate makes stopping a planned outcome rather than an admission of failure.
What does embedding the tool actually mean?+
It means the output arrives where the work already happens rather than in a separate application someone must remember to open. A draft that appears in the inbox alongside the request beats a draft sitting in another tab. This step is skipped more than any other and it is the strongest predictor of whether the thing is still in use six months later.
Who should own an AI project in a small business?+
One named person who does the task or manages the people who do, with enough authority to change how the work is done. Ownership by committee produces a project nobody drives. The owner does not need technical skills; they need process knowledge and the standing to say the new way is now the way.
What if the output is wrong some of the time?+
All of it is wrong some of the time, so the useful question is how often and how visibly. Measure the error rate during the trial against real work. If a person can spot and fix errors in seconds the tool is usable at a fairly high error rate. If errors are subtle and expensive to miss, even a low rate makes the task a poor candidate.
Should I tell staff before or after the project starts?+
Before, in week one. Staff who first hear about an AI project through a rumour assume it is about cutting jobs and stop helping, which removes the process knowledge the project depends on. The people doing the task know where the time goes better than anyone. Bringing them in early buys accurate requirements and a team that wants it to work.
How do I know if the project succeeded?+
Compare hours consumed after against the baseline you measured in the first two weeks, including checking time, and confirm the recovered capacity went somewhere you named in advance. Adoption numbers and logins are activity rather than results. If you cannot state the before figure, the project cannot be assessed and will struggle at its next budget review.
What happens after the first 90 days?+
Either the task runs in production with a named owner and you choose the next one, or you stopped at the gate and you apply what you learned to a better candidate. Both are legitimate. What should not happen is a fourth month of the same trial, since that is the pattern behind pilots that neither succeed nor conclude.
Can I run two AI projects at the same time?+
Not for the first one. MIT found 95 percent of generative AI pilots produced no measurable profit impact, and breadth is strongly associated with that outcome: several half-finished pilots and nothing embedded in daily work. One task carried fully into production teaches more and produces a result you can point to when asking for the next budget.
What is the most common reason a 90 day project fails?+
Nobody changed the process. The tool works, the output is fine, and the team keeps doing it the old way because the old way is habitual and nothing forced the change. This is why the roadmap spends weeks on embedding and handover rather than on tool selection, which is the part people expect to be difficult.
Do I need to buy anything in the first month?+
Usually not. The first two weeks are measurement, and the initial trial can often run on a free tier or a single paid seat. Committing to a company-wide licence before knowing the error rate is premature. The exception is anything touching personal information, where the free consumer tier is the wrong place to start for privacy reasons.
Sources and references
- Statistics Canada: Analysis on artificial intelligence use by businesses in Canada, Q2 2026 (published 11 June 2026)
- Bank of Canada: Canadian businesses' use of AI: What the evidence shows (August 2026)
- MIT Project NANDA: The GenAI Divide: State of AI in Business 2025 (PDF)
- Office of the Privacy Commissioner of Canada: Principles for responsible, trustworthy and privacy-protective generative AI technologies
- Office of the Privacy Commissioner of Canada: PIPEDA in brief
- Parliament of Canada: Bill C-27, Digital Charter Implementation Act, 2022 (died on the Order Paper, 6 January 2025)
The MIT Project NANDA finding was verified against the published report on 3 September 2026. The 90 day structure is this practice's own delivery framework, offered as a working method rather than as validated research.