Switching on an AI feature inside software you already use takes a day to a week. Automating one workflow properly, from mapping to trusted live running, usually takes two to six weeks. A customer-facing assistant takes six to twelve weeks, and a programme covering several processes three to six months. Most of the time goes on testing, not building.
The calendar time is mostly waiting rather than working. A single automation might need only three or four days of actual build effort, spread across weeks because logins arrive late, staff can only review drafts between other jobs, and a trustworthy test needs enough real cases to pass through. Plan in calendar weeks, budget in hours, and treat any quote that promises a customer-facing assistant "live in a week" as a quote that has left the testing out.
Planning ranges by type of AI project
These ranges are built up from the tasks each kind of project involves, for a business of roughly 2 to 20 people with no in-house developer. Your own estimate should come from the method at the end of this tutorial; use the table to sense-check it.
| Type of project | Hands-on effort | Calendar time | What fills the calendar |
|---|---|---|---|
| Switch on an AI feature in existing software (email drafting, meeting notes, document summaries) | 2 to 6 hours | 1 day to 1 week | Admin settings, a short usage note, trying it on real work |
| Roll out a team chat assistant with shared projects and a usage policy | 1 to 3 days | 2 to 4 weeks | Policy sign-off, setting up shared context, staff trying it between jobs |
| Automate one workflow with an AI step (sort, extract, draft), with human approval | 2 to 5 days | 2 to 6 weeks | Access, examples, test rounds, a period of running alongside the old way |
| Customer-facing assistant (website chat or email replies sent without review) | 5 to 15 days | 6 to 12 weeks | Writing the knowledge it answers from, testing traps, longer supervised running |
| Several connected workflows across a department | Several weeks | 3 to 6 months | Doing the above in sequence, each one proving itself before the next |
The first row is often underrated. Many owners have AI features sitting unused in email, office and accounting software, and switching those on is the quickest win available; AI features already in your software lists where to look. Even here there are small waits outside your control. Google's admin help says changes to Gemini app settings "can take up to 24 hours but typically happen more quickly", and Microsoft's Copilot set-up guidance warns that after a licence is assigned, some apps can take up to 24 hours to show Copilot and may need restarting.
Where the weeks go in a single workflow
A one-workflow project has six phases. Only one of them is building.
- Mapping and baseline (2 to 5 days of calendar). Watching the job being done, timing it and collecting real examples. A couple of hours of effort, spread over a few days because the right person is busy.
- Access and accounts (2 to 10 days). Admin logins, connecting apps, agreeing whose name the accounts sit in. This is the most common single delay.
- Build (2 to 4 days). The part everybody pictures.
- Test rounds (3 to 7 days). Running held-back real cases, fixing, re-running.
- Running alongside the old way (1 to 3 weeks). The AI produces its output while staff carry on as normal and compare. Length depends on volume, as the next section shows.
- Watched go-live (about 2 weeks). Daily checks at first, then weekly.
Phase 6 is not optional. Automations can fail silently: Zapier pauses a Zap automatically once 95% of its runs error over seven days, and Power Automate switches off a flow after 14 days of continuous failure. In the first fortnight, someone should look at the run history every day. The wider picture of what implementation involves beyond these phases is in what AI implementation actually involves for a small business.
What stretches or shrinks the timeline
Volume sets the length of the test period
This is the factor most plans miss. You need enough real cases to trust the result, say 50. A process that runs 10 times a week needs five weeks of side-by-side running to see 50 cases; one that runs 50 times a day needs a day. Low-volume processes are slower to prove, not quicker, even though they look simpler.
The number of systems involved
One or two systems with ready-made connectors (an inbox and a spreadsheet, a form and a helpdesk) are quick. Each extra system typically adds a few days for connecting, permissions and testing. Anything without a proper connector, such as older desktop software, can add weeks or rule the project out.
Whether customers see the output
Internal drafts that a person approves can go live after a short side-by-side period. Anything that reaches customers unreviewed needs more test cases, planted trick questions and a longer supervised period, which roughly doubles the calendar.
The state of your data
One clean price list or product sheet is ready to use. Three spreadsheets that disagree add one to three weeks of tidying, and that work falls on your side.
Waits you don't control
Vendor approvals, set-up steps and verification checks sit outside any provider's plan. If the project uses the WhatsApp Business Platform, for example, Meta's business verification is a step you can't hurry; start it on day one. Automation tools have their own quirks: Make, for instance, won't connect to a consumer @gmail.com account until its owner has created a Google Cloud sign-in client of their own, a small but fiddly job when it comes as a surprise.
Your own availability
A sign-off that waits a week adds a week. So does a holiday nobody planned around, or a busy season. Don't schedule go-live for your peak weeks: a sports shop before its main season, a caterer in wedding season, a food truck during festival season.
Timelines for a food truck, a sports shop and a caterer
A food truck: about one week
A street food operator running two trucks already uses Google Workspace, where Gemini in Gmail is included in its plan. The project: use it to draft replies to catering enquiries and write the weekly social posts announcing locations. Day 1, the owner checks the admin setting and writes a half-page note for staff on what not to paste (card details, customers' health information). Days 2 to 5, they draft replies to ten real enquiries and keep the three prompts that worked. There is no build and no test period, because a person writes and sends every message. One week, about five hours of effort.
A sports equipment shop: six weeks planned, seven actual
This is the worked example. A seven-person shop with an online store gets about 600 product-question emails a month ("is this racket right for a beginner?", "does the size 9 boot come up small?"). The project: an AI step drafts answers from the product data and size guides, and staff approve each draft in the helpdesk.
| Week | What happened | Effort |
|---|---|---|
| 1 | Mapped the process and timed 30 replies (4 minutes each on average); collected 80 real questions, 40 for tuning and 40 held back | Owner 3h, staff 2h |
| 2 | Waited for helpdesk admin access: the admin login belonged to a manager who had left, and recovering it took four working days | Owner 1h |
| 3 | Product export showed about 30% of products had no size information; scope cut to products with full data | Owner 2h |
| 4 | Build, first test: 29 of 40 drafts usable with light edits | Provider 2.5 days |
| 5 | Fixes and second test: 36 of 40 usable, all planted awkward questions routed to staff | Provider 1 day, staff 1h |
| 6 | Side by side: drafts produced while staff answered normally; 150 real questions compared | Staff 15 min a day |
| 7 | Went live with approval on every draft; daily run-history checks | Staff 10 min a day |
Total calendar: seven weeks, one more than planned, all of it lost to the admin login. Total effort: about four and a half provider days and roughly 20 hours from the shop's own people. At 600 questions a month, cutting reply time from 4 minutes to about 1.5 minutes saves roughly 25 hours a month, so the calendar time was worth it; whether it pays back quickly enough is a separate question, answered in how soon AI should pay for itself.
A catering company: ten to twelve weeks
A catering company wants a website assistant that answers questions about menus, availability and dietary options, then collects event details for a human to quote. Two to three weeks go on writing the material it answers from: menus, policies and an allergen matrix the kitchen signs off. Build and first tests take one to two weeks. Testing includes 60 questions, a dozen of them deliberate allergen traps ("is the satay nut-free?"), all of which must be handed to a person rather than answered. At about 15 website chats a week, seeing 45 real conversations in supervised running takes three weeks. Add two weeks of watched go-live and the total lands at ten to twelve weeks, and it should. The allergen rule alone justifies the extra care.
Launch dates that slipped, and what caused each one
Three realistic stalls, each with its fix.
- An e-commerce homeware brand went live in its biggest sale week. Order-status questions tripled, drafts piled up faster than two staff could approve them, and the team switched the workflow off to cope. The fix: move go-live to a quiet fortnight and treat peak weeks as a freeze.
- A subscription box company stalled at the test stage for a month. Every test round failed on the same email type, customers mixing a skip request with a complaint. The fix was to cut scope: route every two-issue email to a person and automate the rest, which passed the next round.
- A furniture maker's project paused for three weeks because the only person who could judge the drafts, the owner, was fitting kitchens on site. The fix: agree review slots in the diary before the build starts, two 30-minute blocks a week.
Stalls like these are the main reason pilots never reach daily use; why AI pilots stall and how to get them live covers the pattern in depth.
Six ways to shorten the calendar without cutting the testing
Most of the time you can save is on your side of the project, and none of it needs you to test less.
- Gather the examples before you hire anyone. Pull 50 to 80 recent real cases and remove names and account numbers. It takes an afternoon and removes the most common week-one delay.
- Run a login check first. List every system the project touches and confirm who holds the admin login today. The sports shop's lost week would have been a ten-minute discovery.
- Decide whose name the accounts go in before the build starts, so nothing is set up twice. Accounts registered to the business are the safer default.
- Book review slots in the diary. Two 30-minute blocks a week for whoever judges the output, agreed at the start, keep test rounds moving.
- Start slow approvals on day one. Vendor verification, data-protection sign-off or a partner's approval can run while mapping and building happen.
- Pick a process with enough volume to prove quickly. If two candidate jobs are equally valuable, start with the one that runs daily; its test period will be days, not weeks.
Together these commonly remove a week or two from a single-workflow project, and they make the provider's time go further too, whether you pay fixed or hourly.
Building a timeline with an AI assistant, then correcting it
An AI assistant can turn a task list into a draft plan quickly, provided you give it the facts it can't guess:
I run a [type of business] with [number] staff. I want to
implement this AI project: [one-sentence description].
Tasks I know about: [list]. Constraints: [holidays, busy weeks,
who must approve what, how many times a week the process runs].
Draft a week-by-week plan. For each week, give the task, who does
it, hours of effort, and anything we are waiting on. Include time
to gather real examples, get access to each system, test on
held-back cases, run alongside the current process, and watch the
first two weeks live. Flag every assumption you make.For the sports shop project, a typical first answer (illustrative) read:
"Week 1: requirements and access set-up. Week 2: build the workflow. Week 3: testing and go-live. Assumption: the process runs frequently enough for testing to complete within a week."
Three corrections. It assumed access would be ready in week one, which is the step that slipped in real life; give it two weeks. It folded testing and go-live into one week and left out the side-by-side period entirely, the part that earns staff trust. And it flagged the volume assumption without checking it; at 600 questions a month there are plenty of cases, but for a low-volume process that assumption would be wrong by weeks. A good way to use the assistant is to make it list assumptions, then challenge every one.
Estimating your own project in four numbers
For a single workflow, add these up:
- Build and fix time, in working days, from the provider's estimate or your own (typically 2 to 5).
- Waits you don't control: access, approvals, verification and data tidying, in working days. Ask everyone involved how long their part really takes, then add a few days.
- Test-period length: the number of real cases you want to see (50 is a sensible minimum) divided by how many arrive each week.
- Two weeks of watched live running.
For the sports shop: 4 days of build and fixes, plus about 8 days of waiting, is roughly two and a half working weeks; a little over a week side by side for 150 questions; then two weeks watched. Six to seven weeks, which matches what happened. If your own total comes out under a month, go ahead and book it. If it passes three months for a single workflow, the project is probably several projects, and splitting it will get the first one working sooner. For a phased plan covering several projects, see the five-phase AI implementation roadmap for small businesses.
Timelines: what owners ask next
Can a small business implement AI in a week?
Yes, if the job is switching on and learning an AI feature inside software you use every day, such as drafting help in your email or office suite. A week is too short for anything that runs unattended or talks to customers, because testing on enough real cases and watching the first weeks of live running can't be compressed much.
Does hiring a consultant make implementation faster?
It usually shortens the build and the fixing, because someone experienced has solved similar problems before. It doesn't shorten the waits that sit with you: gathering examples, granting access, reviewing drafts and signing off. Ask any provider which of their timeline's weeks depend on your side, and block that time in your diary before the project starts.
How soon after go-live will we see time savings?
For drafting and sorting jobs, within the first week or two, but the saving is smaller at first because staff check every output closely. It grows as trust builds and review gets lighter. Measure against the baseline you took before starting, at the end of week two and again after six weeks, rather than judging from the first few days.
Further reads
- What a Typical AI Consulting Engagement Looks Like, Week by Week — What each week of a paid engagement usually contains.
- How to Get Staff Buy-In When You Introduce AI — Staff hesitation is the delay no project plan shows.
- Example AI Roadmap for a 12-Person Business, Month by Month — A month-by-month plan for a 12-person business.
- How to Implement AI in a Small Business With No Tech Team — Realistic timing when nobody on staff is technical.
- How to Roll Out Microsoft 365 Copilot to a Small Team — The timeline for a Copilot rollout specifically.
- How to Rescue a Stalled AI Project or Exit It Cleanly — What to do when the timeline has already slipped badly.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft Learn, Set up Microsoft Copilot and assign licences; Google Workspace Admin Help, Turn the Gemini app on or off; Zapier and Power Automate help on paused and switched-off automations.