AI Implementation Roadmap for Small Businesses: 5 Phases in 90 Days

Five phases: groundwork (days 1 to 15), choosing and designing one workflow (16 to 30), a pilot with a person checking (31 to 60), roll-out and training (61 to 75), then measuring, handing over and planning the next quarter (76 to 90). Expect three to five hours a week from one owner.

The 90 days deliver one workflow properly rather than five badly. A common slip is phase 1 dragging because nobody can get into the admin accounts; the most expensive is jumping from a good demo to everyone using it with no pilot. For most teams under ten, tools cost roughly $50 to $300 a month, and often less, because the first workflow can run on AI already included in your software.

Follow me on Instagram@sagnikteaches

The 90 days on one page

That's roughly 32 to 46 hours from the person leading it, plus a few hours from each member of staff involved. Each phase below follows the same illustrative business, an independent pet shop: two branches and an online shop, the owner, a shop manager and five part-time staff, on Microsoft 365 Business Standard. Its customers send about 120 messages a week by email, website chat and social media, asking about food, sizes, stock and deliveries. Answering them takes staff about 5 minutes each, around 10 hours a week, and replies often wait until the next day.

Connect on LinkedInSagnik Bhattacharya

Phase 1, days 1 to 15: groundwork before anything is built

Phase 1 is unglamorous and decides whether the rest goes smoothly. Six tasks:

Subscribe on YouTube@codingliquids

1. Name a lead and a sponsor. The lead runs the roadmap, three to five hours a week; the sponsor, usually the owner, makes decisions within a day when asked. At the pet shop, the shop manager led and the owner sponsored.
2. Get the accounts in order. Business accounts for email and files, admin access held by the business, and two-step sign-in switched on. Microsoft 365's security defaults, which are free, require every user to register for multifactor sign-in.
3. Write a one-page AI rule: which tools are approved, what never gets pasted into them, and who checks what before it leaves the business.
4. List candidate jobs with rough weekly hours. Two or three conversations and a week of tallies are enough; auditing your workflows for AI opportunities is the fuller method.
5. Start measuring the top two or three candidates for two weeks, so the baseline exists by the time you choose.
6. Tell the team what's happening and why, before rumours do it for you.

The announcement doesn't need to be long. The pet shop's, sent on day 3:

Hi all,

Over the next three months we're going to try AI on one job to
see if it takes some of the weight off. Nothing changes this month
except two things:

1. From today, please only use AI tools through your work account
   (Copilot in Microsoft 365). No customer details in any other
   AI tool. The one-page rule is in the staff folder.
2. For the next two weeks, please tally the customer messages you
   answer (the sheet is by each till) and roughly how long they take.

Nobody's hours or jobs depend on this. We're looking for the
repetitive work that eats your day, so tell [manager] which jobs
you'd most like help with.

[owner]

The pet shop hit one snag: the online shop's admin login was still in the web designer's name, which took four days and two emails to fix. Access problems are a common reason for phase 1 to overrun, which is why they come first.

Where the lead's 6 to 10 hours go in phase 1, roughly: two to four hours on accounts and access, one hour writing the rule, one hour on the announcement and a staff-meeting slot, one to two hours listing candidate jobs, and half an hour setting up the tally sheets. Other things that stretch the phase: the sponsor taking a week to approve the rule, staff forgetting to tally (put the sheet where the work happens, not in a shared folder), and trying to fix every data problem in the business rather than only those the likely candidates depend on.

Exit gate: the business holds admin access to every system involved, the rule has been shared, and baselines are running on the candidate jobs.

Phase 2, days 16 to 30: choose one workflow and design it on paper

With two weeks of numbers in hand, choose one workflow. Put each candidate through the seven tests in choosing your first AI project: weekly with two or more hours, measurable, runs on existing tools, fits one sentence, cheap mistakes, live in six weeks, and a clean exit.

The pet shop's tallies showed customer messages at about 10 hours a week, product descriptions for new online stock at about 3, and supplier reorders at about 2. Messages won on hours, but they're customer-facing, so the design had to keep a person checking every reply. The shop also looked at Shopify Inbox, which now includes a free AI agent that answers customers on its own, and decided against starting there: that agent can use web search as a secondary source, and the owner didn't want unchecked answers about feeding or pet health going out before the shop had tested its own policy questions. Drafts with a human check came first.

Designing on paper means writing down the new steps before touching any settings:

1. A message arrives. Staff open the shared "Customer replies" document of product facts, delivery rules and policies.
2. Staff ask Copilot, included with the shop's Microsoft 365 business plan, to draft a reply using that document.
3. Staff check facts against the product page, edit and send. Anything about a sick or injured animal gets the standard "please speak to your vet" reply written by the owner, never an AI draft.
4. Staff note the message type and a grade in the log.

Then write the instructions and test them on ten real, anonymised messages. Here's the core of the pet shop's instructions and one illustrative result:

You draft replies for an independent pet shop. Use ONLY the facts
in the "Customer replies" document. Friendly, short, no jargon.
If the answer isn't in the document, write [CHECK WITH MANAGER].
Never give feeding amounts, medical or behaviour advice; for any
health question, use the standard vet reply word for word.

Customer: "Do you have the grain-free salmon kibble in the big bag, and can you deliver Saturday?" Draft: "Hi! Yes, we stock the grain-free salmon in 12kg bags. We deliver locally on Tuesdays and Fridays, so the next slot would be Friday. Would that work?"

What the manager fixed: the draft stated that the 12kg bag was in stock, but stock levels weren't in the document, so the AI had filled a gap with a guess. The instructions gained a line ("never state stock levels; say you'll check"), and the retest was clean. Eight of the ten test drafts were usable with light edits.

Finally, write the success criteria before the pilot starts: the main number, a quality bar, a guardrail and the verdict rules. Setting AI pilot success criteria that hold up covers how. The pet shop's main number was median minutes per reply, baseline 5, target 3 or less including the check.

Exit gate: a signed one-page commitment note naming the job, owner, tools and stop date, and the written criteria.

Phase 3, days 31 to 60: pilot with a person checking every output

Two staff use the new process for four weeks on every routine message, logging the grade and time for each. The lead holds a 15-minute review each week, with the same agenda every time:

WEEKLY PILOT CHECK (15 minutes)
1. Numbers: messages handled, median minutes, grades A/B/C/D.
2. The worst draft this week: what went wrong, and why?
3. One change to the instructions or the facts document (only one).
4. Anything blocking next week?
5. Still on track for the verdict on [date]?

"Only one change a week" matters. If you change three things at once and results improve, you won't know which one helped, and if they get worse you won't know which to undo.

The pet shop's illustrative pilot numbers:

Weeks 3 and 4 met the target of 3 minutes or less with the quality bar above 80%, no C or D draft reached a customer, and complaints about replies stayed at zero. The verdict, written and dated: pass, with one condition that the facts document is reviewed monthly.

A bad week handled well looks like this. In week 2, a draft said that a food was suitable for kittens; the facts document only said "cat food", and the product was for adult cats. The checker caught it, graded it C and flagged it at the weekly check. The one change that week was adding each product's age range to the facts document, and the same mistake didn't recur. Nothing reached the customer, and the pilot learned exactly what the log is for: a C grade is information, not a failure, as long as it's caught and acted on.

What goes wrong in phase 3: changes pile up without being logged, the pilot quietly expands to a third person or a new message type, or the weekly check gets skipped in a busy week and never returns. Protect the 15 minutes above almost everything else in the diary.

Exit gate: a written verdict against the criteria: pass, extend once with a named change, redesign, or stop.

Phase 4, days 61 to 75: roll out, train and write it down

A pass means the process works for two enthusiastic people. Roll-out makes it work for everyone who answers messages, including the sceptical ones. Four pieces of work:

• Train with real examples. A 30-minute session per group, using the pilot's own log: three good drafts, the worst draft and why, and a live go on today's messages. Staff learn faster from their own shop's mistakes than from a generic demo.
• Put the instructions and examples where everyone can reach them. A shared document or a shared project in your AI tool works well. Don't build on custom GPTs: OpenAI is retiring them, and they stop running on 11 December 2026.
• Write the runbook: one page covering what the process is, who owns the facts document, what never goes into a draft, and how to handle the job if the tool is down.
• Keep measuring usage. The Microsoft 365 admin centre's Copilot usage report shows active users over 7, 28, 90 or 180 days; a tally of messages answered with a draft works for any tool.

At the pet shop, four of the five part-timers adopted the process within a week. The fifth kept writing replies from scratch because she found checking a draft slower than typing her own. Rather than insisting, the manager paired her with the fastest pilot user for two shifts. It turned out she'd been re-reading every product page in full; a shortcut to the facts document fixed it. Non-adopters often have a specific, fixable reason, so ask before you push.

Phase 4 is also when to shortlist the next workflow, because the groundwork is done and the lead has capacity again. The pet shop's next candidate was product descriptions for new online stock.

Exit gate: everyone who does the job has been trained, the runbook exists and has been followed once by someone other than its author, and usage is being tracked.

Phase 5, days 76 to 90: measure, hand over and plan the next quarter

The last phase turns a project into normal operations. Five tasks:

1. Compare with the baseline, using the same measures as phase 1. The pet shop's median reply time held at 3 minutes across all staff, against 5 at the start.
2. Count the full cost: tools, the lead's hours, training time and the weekly upkeep of the facts document.
3. Name the ongoing owner and put a 15-minute monthly check in their diary.
4. Diary every renewal date and notice period for the tools involved.
5. Choose the next workflow and draft its 90 days, which will be shorter, since phase 1 is largely done.

The pet shop's quarter in numbers: about 120 messages a week at 2 minutes saved each is roughly 4 hours a week once everyone was using it, or around 200 hours a year at that rate. The tool cost nothing extra, because Copilot Chat is included with Microsoft 365 business plans; the paid Copilot licence, which adds reasoning across emails, meetings and files together, was noted as something to test later, not bought. The real cost was about 40 hours of the manager's time and roughly 10 hours of staff training and pilot logging. Replies now usually go out the same day, which the owner valued as much as the hours.

Write the review on one page and date it. The pet shop's, filled in:

QUARTER REVIEW - AI ROADMAP, DAYS 1-90
Workflow: drafting replies to routine customer messages.
Result: median reply time 5 min to 3 min (all staff, weeks 11-12);
  same-day replies now usual; no complaints about replies.
Cost: $0 extra in tools; ~40 hours lead time; ~10 hours staff
  training and logging.
Kept: Copilot drafting, facts document, vet-reply rule.
Changed during pilot: stock levels never stated; product age
  ranges added; returns wording added.
Owner from now: shop manager, 15-min check first Monday monthly.
Renewals: Microsoft 365 annual renewal diaried.
Next 90 days: product descriptions for new online stock.
Signed: [owner], [manager]   Date: ________

At 90 days after roll-out, run a fuller check using the 90-day tool review, which catches the drift that sets in once nobody is watching closely.

Exit gate: the quarter review is signed, an owner is named, renewals are in the diary and the next workflow is chosen.

Where outside help fits into the 90 days

Every phase here can be run in-house by a non-technical lead with the time. If you do bring in help, whether a consultant, a freelancer or your IT support company, it tends to earn its fee at specific points rather than across the whole quarter:

• Phase 2, choosing and designing. Someone who has seen many first workflows can spot the trap in yours quickly, such as a candidate that's customer-facing in a way you hadn't considered, or data that won't export.
• Building automations that connect several systems, or anything involving an API, where a mistake can quietly run for weeks.
• A second opinion at the phase 3 verdict, if the result is borderline and you're tempted to extend twice.

Whoever you use, the same conditions protect you: the accounts stay in the business's name, the runbook and instructions are deliverables, the success criteria are agreed before the pilot, and nothing requires the helper to keep it running. Ask for a written scope per phase rather than an open-ended engagement across all five, so each phase can be judged on its own exit gate.

What the 90 days cost: three budgets

Tool costs vary more with your existing software than with your size. Three illustrative budgets, using list prices:

Two points on these. First, the fuller budget's Copilot Business licences need an annual commitment, so buy them at the end of phase 3, when the pilot has passed, not on day 1. Second, the lead's time is the largest cost in every budget. At an illustrative $30 an hour, 40 hours is $1,200 of time over the quarter, more than a full year of the typical budget's tools (about $960). Protecting that time is the best spending decision in the roadmap.

Compressing the roadmap for a one- or two-person business

A solo business doesn't need five formal phases, but it does need the same order. An illustrative music teacher working alone compressed the 90 days into about 45:

• Days 1 to 5: business accounts checked, a note-to-self AI rule (first names only for pupils), a week of timing the candidate jobs.
• Days 6 to 12: chose weekly practice summaries for parents, wrote the instructions, tested ten cases.
• Days 13 to 40: used it for every pupil for four weeks, logging time and edits in a simple sheet.
• Days 41 to 45: compared with the baseline, kept the process, and wrote a half-page of notes for next term.

What didn't shrink: the baseline, the ten-case test and the four-week pilot. Those three are what make the result trustworthy, and they take the same time whether you have one pupil list or a hundred staff.

When 90 days is the wrong length

Ninety days suits one workflow on software you already run. Plan longer when:

• the workflow needs systems to be connected that don't talk to each other today, especially an older system with no export;
• the data is sensitive, such as health or children's information, and your data-protection adviser needs to review the set-up;
• an outside supplier is building something custom, which adds their lead times and a handover;
• you're rolling out to more than one site with different ways of working.

Plan shorter when you're only switching on a feature inside a tool you already use for one person's work. For a year-long view with several workflows in sequence, an example AI roadmap for a 12-person business shows how the quarters fit together.

Signs the roadmap is slipping, and the fix for each

A one-page 90-day plan to fill in

90-DAY AI ROADMAP                        Start date: ________
Lead: ________ (3-5 hrs/week)   Sponsor: ________

PHASE 1  Days 1-15   Groundwork
  [ ] Admin access held for: ______________________
  [ ] Two-step sign-in on all accounts
  [ ] One-page AI rule shared on: ________
  [ ] Candidate jobs + baselines started: ___________
PHASE 2  Days 16-30  Choose and design
  [ ] Workflow chosen (passes 7 tests): _____________
  [ ] New steps written; instructions tested on 10 cases
  [ ] Success criteria signed on: ________
PHASE 3  Days 31-60  Pilot
  [ ] Weekly 15-min check booked (day/time): ________
  [ ] Verdict on: ________   Result: pass / extend / stop
PHASE 4  Days 61-75  Roll-out
  [ ] Training sessions held for: ______________________
  [ ] Runbook written and tested by: ________
  [ ] Usage tracked from: ________
PHASE 5  Days 76-90  Review and hand over
  [ ] Result vs baseline: ______  Full cost: ______
  [ ] Owner named: ________  Monthly check booked
  [ ] Renewals diaried   [ ] Next workflow: ___________

Print it, fill in the names and dates on day 1, and tick it off at each weekly check. A roadmap that lives on the wall gets followed; one that lives in a folder gets remembered at day 80.

Following a 90-day AI roadmap: common questions

Can we pilot two workflows at the same time?

Only if different people own them and neither owner is the same busy person. With one AI lead giving three to five hours a week, two pilots halve the attention each gets, and both tend to drift. It's usually faster overall to run one pilot to a verdict, then start the second in phase 4, when the first is being rolled out and the groundwork is already done.

What if we miss a phase's end date?

Slip the whole roadmap rather than squeezing the next phase. Squeezed pilots are how untested tools reach customers. If a phase overruns by more than a week, find the blocker, which is usually access, time or a decision nobody has made, and fix that specifically. If the same phase slips twice, the workflow may be the wrong choice, so revisit phase 2.

Do we need a consultant to follow this roadmap?

No. Every phase can be run in-house by a non-technical person with the time. Outside help tends to earn its fee at two points: choosing and designing the workflow in phase 2, where experience saves false starts, and building any automation that connects several systems. Whoever you use, keep the accounts in your name and ask for the runbook as a deliverable.

What happens after day 90?

The first workflow moves to a quarterly check by its owner, and the next 90 days start with a shorter phase 1, because the accounts, rules and habits already exist. Most businesses can bring the second workflow live in about 60 days. Review the whole set once a year, dropping anything nobody uses before adding anything new.

Further reads

• How Long Does AI Implementation Take for a Small Business? — Realistic timelines when your project is bigger than one workflow.
• What AI Implementation Actually Involves for a Small Business — The six pieces of work inside any implementation, explained.
• Small Business AI Implementation Checklist: Before, During, After — Tick through every check before, during and after the 90 days.
• How to Write an AI Usage Policy for Your Small Business — Write the one-page rule that phase 1 asks for.
• How to Build a Shared Prompt Library for Your Team — Keep the instructions from phase 4 where the whole team can use them.
• How to Scale AI From One Workflow to the Whole Business — Plan the quarters that follow the first 90 days.
• Hidden Costs of AI Implementation Most Small Businesses Miss — The costs a first roadmap most often leaves out.
• What Happens in a 1:1 AI Implementation Consultation? — What to send beforehand, how a consultant maps your work and scores AI jobs, the tool check, and the one-page plan you should leave with.
• What Is a Fractional Chief AI Officer and Do You Need One? — What a fractional chief AI officer does each month, what providers publish as prices, and cheaper set-ups that suit most small businesses.
• After an AI Consultation: Turn the Advice Into a 30-Day Plan — A one-page AI action plan template, a filled-in 30-day plan for a boutique hotel, the owner's weekly check and how to judge the result at day 30.
• AI Strategy vs AI Implementation Consultant: Which Do You Need? — What an AI strategy consultant and an implementation consultant each hand you, a decision table for eight common situations, and a worked garden centre example.
• What a Typical AI Consulting Engagement Looks Like, Week by Week — A six-week AI consulting engagement explained stage by stage: what the consultant does, what you do, what you should have each Friday, and why weeks slip.
• 10 AI Implementation Mistakes Small Businesses Make — Ten setup, billing and upkeep mistakes that turn a sensible AI project into a year-long cost, each with how it shows up and the fix.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft 365 business plan and Copilot Business pricing pages; Microsoft Learn (Copilot Chat capabilities; Copilot usage report); OpenAI and Anthropic business plan pages; Zapier and Make pricing pages; Shopify help centre (Shopify Inbox AI agent); OpenAI help centre (custom GPT retirement).