Key takeaways
- AI implementation for business is a set of five decisions, not a technology build.
- A 30-day deadline improves the outcome, because the failure mode is indecision, not haste.
- Spending limits and admin configuration belong in week one, not after go-live.
- An implementation is finished when a named employee can do a real task without asking anyone for help.
Most businesses do not have an AI problem. They have a decision problem wearing an AI costume. The technology has been ready for a while; what is missing is a chosen provider, an agreed budget ceiling, a configured admin console, and someone telling the team exactly what to do on Monday morning.
This guide sets out what AI implementation for business actually involves, in the order it needs to happen, and how to compress it into 30 days without cutting the parts that protect you. It is written for the person who has to sign the invoice and then explain the result to a board, a partner group, or a finance director.
What AI implementation actually means
The phrase is used loosely enough to be useless, so here is a working definition. An AI implementation is complete when all of the following are true:
- Your organisation has one primary AI provider, chosen on the record, with the reasons written down.
- Licences are bought at the right tier, under a business agreement rather than personal credit cards.
- Specific models are matched to specific jobs, so you are not paying premium rates for trivial work.
- A hard spending ceiling exists at the provider, not just a dashboard someone is meant to check.
- The admin console is configured: data retention, training opt-outs, sharing rules, workspace structure, and who can invite whom.
- Staff have a short written guide that names the tasks they should use AI for and the tasks they should not.
Notice what is absent from that list. There is no custom model, no data science team, no integration project, and no eighteen-month platform programme. For the overwhelming majority of businesses, the first genuinely valuable AI implementation is a governance and enablement exercise on top of a commercial product. The exotic work, if it is ever justified, comes after you have a baseline that people actually use.
The most common starting point we see: a handful of personal subscriptions expensed by different departments, no admin visibility, no retention settings, and nobody able to answer what the organisation is spending in total. That is not a failure. It is evidence of demand, and it is a good reason to move quickly.
Why 30 days beats six months
The instinct with unfamiliar technology is to extend the evaluation. Run a longer pilot. Compare one more vendor. Wait for the next model release. This feels prudent and is usually the opposite, for three reasons.
The comparison never converges. The leading providers leapfrog each other every few months. If your decision rule is "pick the best model", you have written a rule that can never be satisfied, because the answer changes before procurement finishes. A decision rule that does converge is "pick the provider whose admin controls, commercial terms and support model fit our organisation", because those change slowly.
Long evaluations cost more than fast ones. While the committee deliberates, staff carry on using consumer tools with company information in them. You are not avoiding risk during the evaluation period. You are accumulating it, unmanaged, without the benefit of any admin controls.
Deadlines force scope honesty. When the delivery date is fixed, the conversation shifts from what would be nice to what is genuinely required for people to work differently on day 31. That is a healthier conversation, and it produces a narrower, better first release.
A 30-day window is not a stunt. It is roughly the amount of calendar time the work needs once the decision to proceed has been made, and it is short enough that the sponsor who approved it is still in the room when it lands.
The five decisions that make up an AI implementation
1. Which provider
This is a commercial and administrative decision more than a technical one. The questions that matter are whether the provider offers a business or enterprise tier with real admin controls, whether your data is excluded from training by default at that tier, whether single sign-on is available without jumping to the most expensive plan, how billing works, and whether the provider already sits inside a suite your organisation pays for. Model quality is a factor, but it is the factor most likely to change after you sign, so it should not carry the whole decision.
2. Which model does which job
Providers now ship a range of models at very different price points. Using the most capable model for every request is the single most common source of unexplained AI spend. Summarising a meeting note, drafting a routine email and reformatting a spreadsheet do not need the same model as analysing a contract or writing code. Mapping jobs to model tiers at implementation time, and telling staff which is which, routinely takes a large slice off the running cost with no visible loss of quality.
3. What the ceiling is
Decide the maximum monthly spend before you buy anything, then enforce it at the provider with a hard limit rather than an alert. An alert tells you that you have already overspent. A hard limit means the worst case is a service interruption you can resolve in an afternoon, rather than an invoice you have to explain. Set the ceiling deliberately high enough that normal use never touches it, and low enough that a misconfigured automation cannot quietly run for three weeks.
4. Who can do what
Admin configuration is unglamorous and it is where most of the risk actually lives. Decide who can create workspaces, whether staff can share conversations outside the organisation, how long history is retained, whether connectors to your file storage are enabled and for whom, and how accounts are removed when someone leaves. If you have an identity provider, connect single sign-on so offboarding is one action rather than a checklist item somebody forgets.
5. What people are supposed to do with it
A licence is not adoption. The gap between "everyone has access" and "everyone uses it" is closed by naming specific tasks. Not "use AI to be more productive", but "use it to turn your site visit notes into the client update email, using this prompt, and check these three things before you send it". Five concrete tasks written for the actual roles in your business will outperform a two-hour generic training session by a wide margin.
A week-by-week 30-day implementation plan
| Week | Focus | What exists at the end of it |
|---|---|---|
| Week 1 | Discovery and decision | A written record of how teams actually work, a shortlist of two providers, a recommendation with reasons, and an agreed monthly spending ceiling. |
| Week 2 | Purchase and configure | Licences bought at the right tier, admin console configured, retention and sharing settings set, hard spending limit in place, single sign-on connected if you use one. |
| Week 3 | Pilot with real work | A small group using the tool on their genuine daily tasks, with task-specific prompts written down and the rough edges found while the group is still small. |
| Week 4 | Rollout and handover | Wider access enabled, a short staff guide published, an internal owner briefed on the admin console, and a first month cost review scheduled. |
The sequence matters more than the exact dates. Configuration comes before pilot, and pilot comes before wide access. Businesses that reverse this order tend to spend week four undoing habits formed in week two.
What the business needs to provide
An implementation stalls on availability far more often than on complexity. Three things need to exist on your side:
- A decision maker who can approve spend without convening a committee for each step. One person, reachable within a day.
- Access to whoever administers your identity and email, for roughly two hours in total across the month. This is usually the single scheduling bottleneck, so book it in week one.
- Two or three staff who will actually use the tool and can spare an hour a week in the pilot. Choose people with real workloads rather than the most enthusiastic volunteers, because enthusiasm is not evidence.
How to tell the implementation is finished
Vague completion criteria are how projects drift. Use a definition of done that a non-technical sponsor can verify personally:
- You can name the provider and say in one sentence why it was chosen over the alternative.
- You can open the billing page and see a hard limit, not just current usage.
- You can open the admin console and see every user, with no personal accounts holding company information outside it.
- Removing a departing employee's access is one action.
- You can pick a random employee in a covered role and they can complete a named AI-assisted task without asking anybody for help.
- Somebody internal, not your supplier, owns the setup and knows where the settings are.
That last point is worth defending in the contract. An implementation that only works while the consultant is available has not been handed over. It has been rented.
What it costs
There are two separate numbers and conflating them causes most of the confusion in budget conversations. The first is the implementation fee, paid once, for the decision, the configuration and the enablement. The second is your ongoing provider licence cost, paid monthly or annually directly to the provider, which continues whether or not anyone helps you set it up.
Ask any supplier to state both, and to state the implementation fee as a fixed number rather than an hourly rate against an open scope. Hourly billing on a project whose scope is genuinely knowable in advance transfers all the estimation risk to you, which is the wrong way around.
A useful question for any AI supplier: "If this takes twice as long as you expect, who pays for the extra time?" The answer tells you more about the engagement than the proposal document does.
Frequently asked questions
How long should an AI implementation take?
Thirty days from kickoff to live is realistic for a business adopting a commercial AI provider. Timelines beyond that are usually a symptom of an unmade decision rather than genuine technical complexity.
Do we need AI expertise in-house?
No. You need a budget approver, access to whoever runs your identity provider, and a few staff willing to use the tool on real work. Provider selection, model mapping, admin configuration and spending controls can all be done for you and handed over.
We already bought some licences. Have we ruined it?
No, and this is the most common starting point. Existing licences get audited, the ones that are working get kept, and the rest get folded into one governed setup with visibility and a spending ceiling.
What if the provider we choose falls behind?
Choose on admin controls and commercial terms, keep the contract term short, and avoid building deep custom dependencies in the first release. Switching providers is genuinely disruptive only when you have entangled your processes with one vendor's proprietary features.
Ready to make the decision and move
The Clarity Package gives you a provider recommendation and a written plan your board can approve. The Implementation Package delivers the whole rollout, configured, capped and handed over, live in 30 days or less. Both are fixed price, agreed before work starts.