A 90-Day AI Implementation Roadmap for Small Teams

The two ways an AI rollout usually fails are opposite in shape. Either the business buys something on a Tuesday, switches it on for customers on the Wednesday, and spends the next fortnight apologising for it. Or it disappears into a six-month evaluation and never launches anything at all.

Ninety days sits between those, and it is roughly how long it takes a business with no technical staff and no spare afternoons to get from a decision to a system it trusts. What follows is the AI implementation roadmap I use, in four phases, each ending with a specific deliverable and a decision gate. The gates matter more than the schedule. If you cannot pass one, you repeat the phase rather than pressing on.

This assumes you have already chosen the problem. If you have not, run the scoring test in choosing your first AI project before day one, because the roadmap will faithfully deliver the wrong thing if you point it at the wrong task.

Days 1 to 30: measure and write it down#

No software gets bought this month. That is the hardest instruction in this whole article and the one that separates the projects that work from the ones that quietly stop.

Weeks 1 and 2: baseline#

You are capturing the numbers you will be judged against later. For an enquiry-handling project, four are enough.

• How many enquiries arrive per week, split by phone, web and social.
• How many get a first response within one hour, and how many wait until the next day.
• What proportion turn into quoted work.
• What your average job value is.

Get them from your phone log, your inbox and your booking system rather than from memory. Two weeks of real data beats an impression collected over two years. Most businesses discover something uncomfortable here, usually about weekends, and that discovery is worth the fortnight on its own. The five metrics worth tracking explains how to capture each one properly.

Weeks 3 and 4: write the knowledge document#

This is the asset. In a plain document, in your own words, write your service list with actual prices, your service area including the suburbs you decline, your hours, and the twenty questions customers ask you every week with the answers you would give.

Then write the escalation rules. Which enquiries must reach a human, and how fast. Which words or situations trigger it. What the AI should say when it does not know something.

Resist the temptation to hand over your entire website and call it done. Dumping everything in makes answers worse, not better, because the system cannot tell your 2023 price list from your current one. What to feed an agent covers what helps and what actively degrades the output.

Gate 1: you have a baseline you would defend to your accountant and a knowledge document a new employee could work from. If a new hire could not answer the phone from it, neither can the AI.

Days 31 to 60: build it and run it in the dark#

Now you buy something, and it processes real enquiries without a single customer seeing its output.

Weeks 5 and 6: configuration#

Whatever you have chosen, self-serve or configured, this is where your knowledge document gets loaded, the calendar and CRM connections get made, and the tone gets set. Keep the initial scope narrow. One channel, one type of enquiry. You are trying to get one thing right rather than five things nearly right.

Set up the handoff before you set up anything clever. A customer who gets passed to you having already explained their problem twice is worse off than one who reached voicemail, so the transcript has to travel with them. The mechanics of a clean handoff are worth reading before you configure the rules.

Weeks 7 and 8: shadow mode#

Shadow mode is the part people skip and the part that saves them. The system receives real enquiries and drafts real responses. A person reads every one, corrects it, and sends the corrected version. The customer sees only human-approved output.

Keep a simple log with three columns: what came in, what the AI produced, and what you actually sent. After two weeks you will have a list of every recurring error, and nearly all of them will be knowledge gaps rather than model failures. Fix them in the document, not by arguing with the system.

Two weeks of shadow mode typically removes the bulk of the mistakes that would otherwise have reached a customer. It costs you review time and nothing else.

Gate 2: across the last fifty shadow conversations, you would have sent the AI's draft unedited at least four times in five, and it escalated correctly every time it should have. Below that, extend shadow mode by two weeks rather than launching.

Days 61 to 75: live, but narrow#

The system goes live on the smallest slice that still matters. For most service businesses that means after-hours only, or one channel, or one category of enquiry.

Watch it daily for the first week. Read every conversation. This is not a supervision burden you carry forever, but for seven days it is the job, and it will teach you more about how customers actually ask for things than any amount of planning.

Two behaviours to look for specifically. Confident wrong answers, which mean a knowledge gap you need to close immediately. And unnecessary escalations, where the system passes something to you it could have handled, which usually means the rules are too broad rather than too tight. The escalations are the safer failure, so fix the wrong answers first.

Tell your team what is running and why. Front-desk staff who discover an AI is answering the phone by hearing it happen will not cooperate with it, and their cooperation is what makes month three work.

Gate 3: a week of live operation with no customer complaint traceable to the system, and escalations arriving with enough context that whoever picks them up does not have to start over.

Days 76 to 90: widen and hand over ownership#

Now the scope grows. Add the second channel, extend the hours, allow the next category of enquiry. One change at a time, with a few days between them, so that when something degrades you know which change caused it.

The more important work this fortnight is ownership. Name the person who reviews a sample of conversations each week and keeps the knowledge document current. Fifteen minutes a week is enough, and it has to be in someone's actual role rather than in everyone's general awareness. Systems without an owner drift, and you find out in month seven that it has been quoting a price you stopped charging in March.

Then compare against your baseline. Same four numbers from weeks one and two, measured the same way. This is the only honest test of whether the thing worked, and it is only possible because you did the boring fortnight at the start.

Gate 4: the numbers moved in the right direction, a named person owns the system, and you can state the monthly cost against the monthly gain. Run that comparison properly using the ROI method.

Where an AI implementation roadmap usually stalls#

Failing a gate is information, not defeat. The common failures each have an obvious response.

If shadow mode keeps producing wrong answers, your knowledge document is thin and the fix is another afternoon of writing rather than a different vendor. If escalations arrive without context, the integration is incomplete and the customer will feel it. If the numbers did not move by day 90 but the system is answering well, the bottleneck was somewhere else in the business, most often in how long a quote takes to go out.

That last case is the useful one. A quick response into a slow quoting process just gets people to the queue faster, and no amount of AI fixes it. Better to learn that in ninety days than after a year of subscriptions.

If you are still weighing up which system to run through this roadmap, the overview of AI solutions for small business covers what each option does and costs.

Start with the fortnight of counting. It is unglamorous, it requires no budget, and every good decision later in the ninety days depends on it.