For a 12-person business, a realistic AI roadmap spends month one on groundwork (accounts, rules and a time baseline), then adds one workflow roughly every six to eight weeks, each piloted with a person checking before it goes live, with reviews at months six and twelve. In the example below, new tool costs stay at about $30 a month.
The business is an illustrative driving school with twelve people: the owner, who still teaches two days a week, an office manager, a bookings assistant and nine instructors. What follows is its first year, month by month: what it did, what it cost in money and hours, what came of it, and what slipped. There's a blank template at the end to copy for your own business.
The starting point
Before month one, the school's week looked like this. Around 120 enquiries a month arrived by phone, website form, email and social messages, and the median wait for a first reply was about six hours, longer at weekends. About 20 lessons a week were cancelled, and freed slots were refilled only about a third of the time. Instructors kept lesson notes in paper diaries, if at all, so the office couldn't answer parents' "how is she getting on?" calls. Chasing unpaid lesson-package balances took the office manager about five hours a month.
Existing tools: Google Workspace Business Starter for all twelve people (about $7 per user a month, so roughly $84 a month, already paid), a driving-school booking and diary app, accounting software and a website with an enquiry form. Nobody had a paid AI tool, though two instructors used free chat apps on their phones.
The year at a glance
| Month | Focus | New monthly tool cost | Owner / office manager hours per week |
|---|---|---|---|
| 1 | Groundwork: rules, accounts, baseline | $0 | 2 / 2 |
| 2 | Workflow 1: enquiry reply drafts | $0 | 2 / 3 |
| 3 | Workflow 2: cancellations, in shadow mode | $29.99 | 1 / 3 |
| 4 | Workflow 2 live | $29.99 | 1 / 2 |
| 5 | Workflow 3: lesson notes pilot, two instructors | $29.99 | 2 / 1 |
| 6 | Six-month review | $29.99 | 2 / 1 |
| 7 | Lesson notes for all instructors | $29.99 | 2 / 1 |
| 8 | Workflow 4: package balance reminders | $29.99 | 1 / 2 |
| 9 | Out-of-hours chat tested, then postponed | $29.99 | 1 / 2 |
| 10 | Workflow 5: diary-gap offers | $29.99 | 1 / 2 |
| 11 | New-instructor onboarding pack | $29.99 | 1 / 1 |
| 12 | Year-end review and year-two plan | $29.99 | 2 / 1 |
Months 1 and 2: groundwork, then enquiries
Month 1 produced no visible results, on purpose. The owner wrote a one-page AI rule: pupils' licence details, medical declarations and payment information never go into any AI tool; everything customer-facing is checked by a person before it goes out; company accounts only. The two instructors using free chat apps moved to the Gemini app on their existing Workspace accounts, which the Business Starter plan includes along with Gemini in Gmail. The office kept a two-week time log and counted enquiries, reply times, cancellations and unpaid balances. That baseline mattered all year, for reasons covered in why a baseline comes before any AI tool.
The rule fitted on half a page and went up on the office noticeboard and into the instructors' group chat:
HOW WE USE AI AT [SCHOOL NAME]
1. Company accounts only. Use the Gemini app and Gemini in Gmail
on your school account. No personal chat apps for school work.
2. Never put in: licence numbers, medical declarations, eyesight
or health information, card or bank details, test booking
references.
3. Pupils' first names and lesson progress are fine in the
company account.
4. A person reads everything before a pupil or parent sees it.
5. Not sure? Ask [office manager] before you paste.Line 3 was the one the instructors asked about most. Without it, some assumed they couldn't mention a pupil at all, which would have made the month 5 lesson notes impossible. Saying plainly what is allowed turned out to be as useful as the list of what isn't.
The baseline itself was a single sheet with five numbers on it: enquiries per month (about 120), median first-reply time (about six hours), cancellations per week (about 20), share of freed slots refilled (about one in three) and hours a month chasing balances (about five). Every later review compared against that sheet and nothing else, which kept the arguments short.
The owner also scored the candidate processes for readiness. Enquiry replies and cancellations scored well enough to go first; lesson notes didn't yet, because nobody agreed what a note should contain. Telling whether a process is ready to automate has the scorecard they used.
Month 2 built the first workflow. The office manager wrote short pages covering prices, packages, the areas each instructor covers, intensive courses, what happens on a first lesson and the cancellation policy. Those pages fed saved Gmail templates and Gemini's drafting in Gmail. Every draft was checked and sent by a person. By the end of the month, the median first reply during office hours was under an hour, and office time on enquiries had roughly halved. Answering pupil and parent enquiries with AI at a driving school covers this workflow in more depth.
A typical parent enquiry showed why the checking step stayed. The message: "My daughter is about to start learning. Roughly how many lessons will she need, and do you do intensive courses?" Gemini's draft (illustrative):
Thanks for getting in touch. Most learners need around 40 to 45 hours of lessons before they're ready for their test, though it varies. We do offer intensive courses, which pack lessons into one or two weeks. Would you like me to book an assessment lesson so the instructor can give you a better idea?
The intensive-course line and the assessment offer came from the school's pages. The "40 to 45 hours" figure didn't: it was a general number the model supplied, and the school's policy was not to estimate before an assessment lesson. The office manager cut the sentence and added a line to the knowledge pages: "We don't estimate how many lessons someone will need until after their first assessment lesson. Say so if asked." It was the most common edit in the first fortnight and had almost disappeared by the end of the month.
Months 3 and 4: cancellations and the waiting list
The school replaced five ways of cancelling with one short form, linked from every booking confirmation. In month 3 it set up Zapier Professional at $29.99 a month on monthly billing, which includes 750 tasks. Each submitted form now sets off an automation that applies the 48-hour notice rule, proposes the diary update, and asks an AI step to write a message offering the newly free lesson to waiting-list pupils whose availability matches.
For three weeks it ran in shadow mode: the automation drafted, the office handled cancellations as before, and they compared. Piloting AI in shadow mode covers the method. Over those three weeks the office logged 62 cancellations. The automation's fee decision matched the office's on 58; the four differences were all illness cases where the office had waived the fee under the "once per course" rule, which the automation hadn't been told about. The offer drafts were the weak part. For a freed Tuesday 10am lesson, the AI step picked a waiting-list pupil whose availability read "weekday mornings mostly, not Tuesdays (college)" and drafted a cheerful offer anyway. It had matched "weekday mornings" and skimmed past the exception. That was the problem the tick-boxes described further down eventually fixed. It went live in month 4 with the office approving every offer before it's sent.
The task count needed watching. Around 85 cancellations a month, each using two action steps plus an AI step, came to roughly 425 tasks on the three-task model tier they first picked (Zapier's help centre says an AI step uses one, three or five tasks per run, depending on the tier). That left headroom within 750 for month 8's reminders, but not much, so the office switched the AI step to the cheaper one-task tier after checking its drafts were just as good.
Months 5 and 6: lesson notes and the six-month review
Month 5 piloted lesson notes with two instructors. After each lesson, the instructor records a short voice note in the Gemini app, which turns it into a structured note: skills covered, what went well, the focus for next time. The instructor checks it and pastes it into the pupil's record in the booking app. The rule was explicit: no medical details and no comments about anyone other than the pupil.
A 40-second voice note such as "Good lesson, roundabouts much better, still late on the mirror check before signalling off, parallel park two out of three, she said she's been anxious and her doctor's changed her tablets, next time faster roads" came back as (illustrative):
Skills covered: roundabouts, parallel parking. Went well: roundabout approach and lane choice much improved; parallel park successful 2 of 3. Focus next time: mirror check before signalling to exit; first session on faster roads. Other: pupil reports anxiety; recently changed medication.
The first three lines were exactly what the office wanted. The fourth was the rule being broken, not by the AI but by the instructor talking naturally, and the AI faithfully kept it. Two changes followed: the saved prompt gained the line "Leave out anything about health, medication or personal life, even if I mention it", and the instructors' checklist before pasting into the booking app became "three headings, nothing else".
The month 6 review compared everything with the month 1 baseline. Enquiry replies had held their improvement. The share of freed slots refilled had risen from about a third to about half. The two pilot instructors were writing notes for almost every lesson, in about two minutes each, and the office could now answer progress calls without ringing the instructor. The decision was to roll notes out to everyone, with no new tools needed. The review also caught one quiet failure: the school had raised its hourly rate in month 5, and two of the saved Gmail templates still quoted the old price. Nobody had complained, because the office usually spotted it while checking, but it was exactly the kind of drift a scheduled review is for. The fix was a "templates that mention money" list, checked whenever prices change. The review format came from reviewing an AI tool after 90 days, applied to the whole programme rather than one tool.
Months 7 to 9: roll-out, reminders and the chatbot that waited
Month 7 brought all nine instructors onto lesson notes, after a one-hour session where the pilot pair showed the others how they did it. Adoption was uneven: by month 9, two instructors were still writing in paper diaries. The owner stopped pushing and asked the office to type up their notes weekly instead, which was a realistic compromise rather than a failure.
The compromise had a price worth knowing. Two instructors at about 25 lessons a week each is 50 paper notes, and at a minute or so to read each scrawl and type three headings, the bookings assistant spent close to an hour every Monday on it. That's cheaper than losing two good instructors over a note format, and it gave the owner a figure to weigh when the notes came up again at the year-end review.
Month 8 automated package-balance reminders. Most of it is a plain rule (balance owed, days since the last lesson), with AI used only to draft the three reminder messages once, which the owner edited and approved. Office time on chasing fell from about five hours a month to under two.
The logic, written down before anything was built, was short enough to check by eye:
| Condition | Action |
|---|---|
| Balance owed, last lesson 7+ days ago, next lesson booked | No reminder; the instructor mentions it at the next lesson |
| Balance owed, last lesson 14+ days ago, nothing booked | Reminder 1 (friendly) |
| Still owed 14 days after reminder 1 | Reminder 2 (clear, with a payment link) |
| Still owed 14 days after reminder 2 | Flag to the owner; no automatic third message |
Every line is a plain rule. The only AI involvement was drafting the two messages once. The owner deliberately kept the final step with a person, because by then the pupil has usually stopped learning for a reason worth knowing.
Month 9 was meant to launch an out-of-hours chat assistant on the website. Before choosing a tool, the office tested the idea by asking an AI assistant 25 real weekend enquiries using only the knowledge pages. It answered 17 correctly. The failures were mostly questions about which instructor covers which area and pricing for pupils switching from another school, neither of which the pages covered clearly. Two of the failures show the pattern. Asked "I've had ten hours with another school, can I go straight onto one of your intensive courses?", the assistant answered "Yes, you can join any of our intensive courses", when the school requires an assessment lesson first. Asked "Does anyone cover the villages north of town?", it named an instructor who had stopped covering that area in the spring, because an old page still said so. Both were errors in the pages rather than the AI, and both would have gone straight to a pupil at 9pm on a Saturday with nobody checking. The area page showed how. It read "Instructor C covers the north side and the villages; Instructor D does most of the east", a sentence written three years earlier and never dated. Its replacement was a small table, one row per instructor, listing the areas covered, the days they teach there and a "last checked" date, with the office manager checking it whenever an instructor's diary changes. The owner postponed the chat assistant until the pages could pass a stricter test. No money spent, and a clear list of what to fix.
Months 10 to 12: diary gaps, onboarding and the year-end review
Month 10 tackled unfilled instructor time. Each Monday, the office exports the week's diary and asks AI to list the gaps by instructor and area, then drafts offers to waiting-list pupils who match. It reuses the month 3 automation's approval step.
The first Monday's list (illustrative) read: "Instructor A: Tue 13:00-14:00, Wed 09:00-11:00. Instructor B: Thu 12:00-14:00, Fri 15:00-17:00." Three of those four gaps were real. Thursday 12:00-14:00 was Instructor B's regular lunch and school run, which the booking app stored as blocked time but the export showed as an empty cell. The office added a "Blocked" value to the export's status column and a line to the prompt ("treat any slot marked Blocked as unavailable") before any offers went out. In its first month the workflow filled about a dozen extra lesson hours that would otherwise have sat empty.
Month 11 used the knowledge pages and a year of lesson-note examples to build an onboarding pack for a new instructor joining in the spring, drafted with AI and edited by the office manager. It took a day instead of the week the owner had set aside.
The pack had six sections: how the booking app works, the cancellation and fee rules, the lesson-note format with three real examples (pupils' names removed), the areas each instructor covers, how to handle a parent asking about progress, and the AI rule from month 1. AI drafted the first four from the existing pages in minutes. The office manager rewrote the fifth almost completely, because the draft was polite but vague, and the owner wrote the pack's last page personally: what to do on your first day.
Month 12 was the year-end review: keep, fix or cancel for every workflow, and a plan for year two. All five workflows were kept. Two got a "fix": lesson notes (two instructors still on paper) and diary-gap offers (only run when the office remembered on a Monday). The review sheet itself was one table, filled in at a 90-minute meeting:
| Workflow | Verdict | Evidence against the baseline | Action for year two |
|---|---|---|---|
| 1. Enquiry reply drafts | Keep | Median first reply under 1 hour, from about 6 | Review templates whenever prices change |
| 2. Cancellations and offers | Keep | Refilled slots about 2 in 3, from 1 in 3 | None; runs itself |
| 3. Lesson notes | Fix | Most lessons noted; two instructors on paper | Instructors redesign the note format together |
| 4. Balance reminders | Keep | Chasing under 2 hours a month, from about 5 | None |
| 5. Diary-gap offers | Fix | About a dozen hours filled in month 1, fewer since | Run it from a scheduled automation, not memory |
Writing the evidence column forced one honest line: the diary-gap figures had tailed off after the first month, not because the workflow failed but because nobody owned the Monday run. The year-two list had three items in order: rewrite the knowledge pages so the chat test passes 23 of 25, re-test and then launch the out-of-hours assistant, and trial AI help with the weekly instructor rota. Anything else went on the parked list.
What the year cost and what came of it
| Measure | Before | After 12 months |
|---|---|---|
| New tool spend | n/a | About $300 for the year (Zapier, months 3 to 12) |
| Owner and office manager time on the programme | n/a | About 170 hours in total across the year |
| Median first reply to enquiries (office hours) | About 6 hours | Under 1 hour |
| Freed lesson slots refilled | About 1 in 3 | About 2 in 3 |
| Lessons with a written note | Few | Most, including the typed-up paper notes |
| Office time chasing balances | About 5 hours a month | Under 2 hours a month |
The biggest cost was time, not software, which is typical. The biggest return was the refilled slots: each one is a paid lesson that would otherwise have been lost. The rough sum: of about 20 cancellations a week, refilling two in three instead of one in three means six or seven more lessons taught each week. At an illustrative $45 a lesson, that's roughly $300 a week of lessons that used to vanish, against about $300 of new tools for the whole year. The 170 hours matter more than the subscription, and even those look modest next to that figure, although that rate was only reached in the second half of the year.
What slipped, and what the owner would do differently
- The cancellation shadow run took three weeks, not two, because waiting-list pupils had given their availability in free text ("weekday mornings mostly") and the AI step kept offering slots they couldn't make. Three tick-boxes on the waiting-list form fixed it.
- Instructor adoption was slower than planned. The owner's view: involve instructors in designing the note format in month 1, not month 5.
- The chat assistant was postponed, which the owner counts as the roadmap working, not failing. A test that costs an afternoon beats a tool that answers customers wrongly.
Year two starts with the knowledge-page fixes, then the out-of-hours assistant, re-tested first. For the five-phase version of this kind of plan compressed into 90 days, see the AI implementation roadmap for small businesses.
A blank roadmap to copy
AI ROADMAP: [business name] Year starting: [month]
Programme owner: [name] Monthly check: [day], 30 minutes
MONTH 1: GROUNDWORK
AI rule written and shared [ ]
Company accounts set up [ ]
Baseline measured (list measures): ______________________
Processes scored for readiness: ______________________
WORKFLOW [n]: ______________________
Owner: ________ Months: ___ to ___ Tool cost: $____/month
Shadow run: ___ weeks Go-live check: ______________________
Success measure vs baseline: ______________________
(repeat for each workflow, one every 6-8 weeks)
MONTH 6 REVIEW: keep / fix / cancel for each workflow
MONTH 12 REVIEW: results vs baseline; year-two list
PARKED IDEAS (and why):
______________________Filled in for the school's second workflow, one entry looks like this:
WORKFLOW 2: Cancellations and waiting-list offers
Owner: office manager Months: 3 to 4 Tool cost: $29.99/month
Shadow run: 3 weeks Go-live check: fee decision matches the
office on 19 of 20 cases; no offer sent
to a pupil who can't make the slot
Success measure vs baseline: freed slots refilled, from about
1 in 3 to at least 1 in 2 by month 6The go-live check is the line people leave vague. Write it as something you can count, so the decision to switch on isn't a matter of mood.
The "parked ideas" line is worth keeping. A roadmap that records what you chose not to do, and why, is much easier to pick up next year than one that only lists successes.
Questions about adapting this roadmap
What if we use Microsoft 365 instead of Google Workspace?
The shape of the year stays the same. Copilot Chat is included with Microsoft 365 business plans and covers the drafting in months 1 and 2. If you want AI inside Outlook for enquiry replies, budget for Microsoft 365 Copilot Business at $21 per user a month on annual billing, but only for the two or three office staff who handle enquiries, not all twelve people.
Our instructors are self-employed. Can they use the company's AI tools?
They can, but decide it deliberately. If they record notes about your pupils, those notes should live in your systems under your rules, so give them access to the company tool for that task rather than letting each use a personal account. Write the arrangement into their agreement, including what happens to the data if they leave.
How do we stop the roadmap slipping?
Give each workflow one named owner and a fixed review date, and keep the monthly progress check to 30 minutes. Expect slippage anyway: the example school lost about three weeks across the year. The roadmap is a sequence, not a deadline, and moving a date is better than launching something that failed its shadow run.
Further reads
- How to Write a One-Page AI Strategy for Your Business — The one-page summary that sits above a roadmap like this.
- How to Scale AI From One Workflow to the Whole Business — How year two builds on the five workflows from year one.
- AI Change Management for Small Teams: A Practical Plan — Handling the instructor adoption problem from months 7 to 9.
- How to Choose and Support an AI Champion in a Small Team — Why the office manager's role mattered more than any tool.
- How to Calculate AI ROI for Your Business (Worked Example) — Turn the year-end results table into a proper return figure.
- How Driving Schools Use AI to Keep Instructor Diaries Full — More on the diary-gap workflow from month 10.
- How to Build a Company Knowledge Base AI Can Answer From — The pages behind months 2, 9 and 11, and how to write them.
- How to Get Started With AI in Your Small Business: First 7 Steps — Seven steps that take about a month, from a five-day time log to a keep-or-drop decision, with a yoga studio's numbers at each stage.
- AI Adoption Stages: Where Is Your Business Now, and What's Next? — Place your business on a five-stage AI adoption scale with a ten-question check, then see the one move that gets you to the next stage.
- How Long Does AI Implementation Take for a Small Business? — Realistic planning ranges for small-business AI projects, where the weeks really go, and a week-by-week timeline for a sports equipment shop.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Workspace plan inclusions, Zapier plan prices and task rules, and Microsoft 365 Copilot prices from the vendors' pricing pages; Zapier help centre on AI by Zapier task pricing. Checked September 2026.