Timelines for AI work in a UK SME fall into four bands. Rolling out assistants on business plans with a policy and a session per team takes a week. A first automation, one process, one integration, an exceptions path and a measured threshold, takes four to eight weeks from brief to go-live. A custom system with several integrations, an agent that acts across systems, or a voice channel takes eight to sixteen weeks, delivered in phases with a demo each week. The 90-day measurement then runs after go-live and is part of the project, not an afterthought. What makes projects slip is rarely the AI: it is access to systems that takes a month to grant, data that turns out to need tidying, decisions nobody owns, template and number approvals from third parties, and client sign-off that waits for a meeting. A fixed-price supplier plans for those on day one.
- Assistants and policy
- One week
- First automation
- Four to eight weeks
- Custom system or agent
- Eight to sixteen weeks
The four bands
Four bands cover almost every engagement; the audit sits before any of them.
| Type | Typical duration | What it contains | Price band |
|---|---|---|---|
| Assistant rollout | One week | Business plans, SSO, policy, templates, a session per team | Seats plus training from £1,200 |
| First automation | Four to eight weeks | Baseline, one integration, exceptions path, threshold measured, go-live | £4,000 to £15,000 |
| Custom system or agent | Eight to sixteen weeks | Several integrations, controls, phases with weekly demos, handover | £8,000 to £25,000 |
| Audit before any of the above | Two to three weeks | Interviews, baselines, ranked options with prices, roadmap | £2,500 to £6,000, credited |
RelatedA 30-day AI rollout plan How much AI automation costs in the UK How long a business app takes to build
What the weeks contain
For a first automation: week one, baseline measurement starts and access to systems is requested; weeks one to two, discovery on your fifty samples and the threshold agreed in writing; weeks two to five, the build with a demo each week on real inputs; week five to six, the threshold test on the samples, the exceptions path tried by the person who will work it, ceilings tested; week six to eight, go-live in shadow mode beside the manual process, then live, with the runbook handed over. The 90-day measurement then compares against the baseline. A custom system runs the same shape in phases, each phase ending in something usable.
RelatedHow to measure AI return in 90 days Why most AI pilots never reach production
What makes projects slip
Six causes account for nearly every late AI project we have seen, and none is the model.
- 01
Access. API keys, admin accounts and connector permissions that take a month to grant because the person who can grant them is busy. Request them on day one.
- 02
Data. The system of record turns out to have duplicates, missing fields or two versions of the truth, and the first fortnight becomes tidying. The readiness check before the build finds this.
- 03
Decisions. What counts as an exception, who approves what, which template wording: small decisions that wait for a meeting. Name an owner with authority to decide in a day.
- 04
Third-party approvals. WhatsApp templates, telephone numbers, app store reviews, a supplier’s API access: each has its own clock. Start them in week one.
- 05
Scope drift. The second process added to the first mid-build. Write it down for phase two and finish phase one.
- 06
Sign-off. Go-live waiting for a director’s diary. Book the decision meeting when the project starts.
RelatedIs your data ready for AI? An AI strategy for a UK SME
How a fixed price keeps to the calendar
A fixed price puts the risk of overrun on the supplier, which is why fixed-price suppliers plan access, data and approvals in the first week and hold a decision-maker to a cadence. A day rate makes slippage billable. The published-price market in the UK, surveyed in our cost guide, still largely works on day rates and discovery phases; the question to ask any supplier is what the calendar looks like and who carries the cost if it moves.
What not to compress
The baseline, because without it there is no return to measure. The threshold test on real samples, because it is what makes finished a fact. Shadow mode beside the manual process, because it catches the failures a demo did not. And the 90-day measurement, because the surveys say the projects that skip it are the ones abandoned. Compressing the build is possible; compressing the evidence is not.
RelatedAI adoption audit
Common questions
How long does it take to implement AI in a small business?
A week for assistants and a policy; four to eight weeks for a first automation with one integration; eight to sixteen weeks for a custom system, agent or voice channel; then a 90-day measurement after go-live.
What takes the time in an AI project?
Access to systems, tidying data, decisions that wait for meetings, third-party approvals and sign-off. The model work is rarely the long pole.
Can an AI automation be live in a month?
A simple one with a ready integration and a decisive owner, yes. Most first automations take four to eight weeks because the baseline, the threshold test and shadow mode take real time.
How long should an AI audit take?
Two to three weeks: interviews, baseline measurements, ranked options with prices and running costs, and a roadmap. Longer than that is a sign the audit has become a project.
Does a fixed price make projects faster?
It makes overrun the supplier’s cost, so fixed-price suppliers plan access, data and approvals early and hold the calendar. Day rates make slippage billable.
When is an AI project finished?
When the threshold is met on real inputs, the exceptions path is worked by its owner, the runbook is handed over, and the 90-day measurement is compared against the baseline. Not at go-live.