How to implement AI in a business: a 7-step roadmap
Updated:
To implement AI in a business, start with one process, not a company-wide programme. Pick work that repeats and takes real time, check that its data and systems can be reached, and measure how it runs today. Then run a small pilot on real cases with a person checking every result, compare it with the baseline, and roll it out only if the numbers hold. The seven steps below work for a company of ten people or two hundred.
What should an AI implementation strategy include?
For a smaller company, the strategy is a short list, not a slide deck. It answers five things:
- Which processes, in what order. Ranked by time spent and how easy a mistake is to catch.
- What success means. Hours saved, errors avoided or faster replies, measured against today.
- Who owns it. One person who decides and checks the results.
- Which data may go where. What may be sent to an AI provider and what stays inside.
- What the AI may do alone. Where it acts on its own and where a person approves first.
Write it after you have looked at the processes, not before. The AI readiness checklist gives you the answers to fill it in.
The AI implementation roadmap, step by step
- Pick one process. Choose work that happens often, follows a usual path and has input that arrives digitally: copying orders into invoicing, sorting incoming emails, pulling fields out of PDFs. If every case is identical, a plain rule may do the job without AI.
- Measure how it runs today. Count how often it happens, how long each case takes and how often something goes wrong. Two weeks of tallying is enough. Without this baseline you can’t tell whether the AI helped.
- Check the data, the systems and the rules. Find out where the input comes from, where the result has to go, and whether both systems have an API or an export. If personal data is involved, read GDPR and AI before you choose a provider.
- Decide how to build it. An AI tool you buy, an automation platform with an AI step, or a custom integration. The table below compares them.
- Run a pilot on real cases. Use real emails, invoices or tickets, not demo data. Let the AI do the work while the person who does it today approves every result. Run it long enough to see the usual exceptions.
- Compare and decide. Put the pilot’s time, errors and corrections next to the baseline. Go ahead, adjust and test again, or stop. Stopping is a valid result.
- Roll out and hand it over. Reduce approvals to the cases that need them, log what the AI does, and name the person who watches the results and reports problems. Then go back to step one with the next process.
Should you buy an AI tool or build something?
Most companies end up using more than one. Choose per process.
| Fits when | Watch out for | |
|---|---|---|
| An AI tool or AI feature you buy | The work happens inside one product, such as writing in an AI assistant or the AI features of your helpdesk or CRM | Data stays inside that product, per-user fees add up, and you can only change what the vendor lets you |
| An automation platform with an AI step | Data moves between a few common tools with ready-made connectors, and the steps are simple | Exceptions and error handling get messy as the flow grows, and you pay per run or per task |
| A custom integration | Several systems are involved, there are many exceptions, or a mistake is costly enough that you need full control and logs | Higher cost up front, and someone has to host and maintain it |
Buy before you build when a tool already does the job. Build when the tools force you to change how the work is done, or when the process runs across systems no single tool connects.
What does an AI implementation plan look like?
One page per process. An example for supplier invoices:
| Field | Example |
|---|---|
| Process | Supplier invoices arrive by email and are typed into the accounting system |
| Owner | The person in finance who does it today |
| Baseline | Invoices a month, minutes per invoice, typing errors found later |
| Target | The AI fills in the fields, a person approves, and errors don’t rise |
| Data and systems | Mailbox, accounting system API, which provider processes the invoices and where |
| AI alone / needs approval | Reads and fills in alone; a person approves anything over a set amount or from a new supplier |
| Pilot and decision | Start date, end date, and the date you compare the numbers and decide |
How long does it take, and what does it cost?
It depends on the process more than on the AI. The cost of AI implementation comes from the same things in almost every project:
- How many systems are connected, and whether they have an API.
- How many steps and exceptions the process has.
- Where a person has to approve, because each approval needs a screen or a notification.
- Running costs after launch: AI usage fees, which most providers bill by volume (see OpenAI’s API pricing for one example), plus hosting, monitoring and fixes when a connected system changes.
The running costs are the part most often left out of the first estimate. Ask for them in writing before you start, whoever builds it.
Why do AI implementations stall?
- Starting too broad. “Use AI in customer service” never finishes. “Draft replies to delivery questions” does.
- Buying the tool first. The process then bends to the tool, or the licences go unused.
- No baseline. Without the numbers from before, nobody can show it helped, and the project loses support.
- Piloting on demo data. Real input is messier, and the problems only appear after launch.
- No plan for mistakes. The AI will get things wrong. Decide in advance who notices and what happens next.
- No owner. When the person who started it moves on, a system nobody watches slowly breaks.
Can you do this yourself?
Yes, for a first process that runs inside one tool you already use. Steps one to three need no outside help, and the checklist and the plan above cover them. An outside view helps when several systems are involved, personal data goes in, or nobody in the company has built an integration before. vSites does this as AI consulting: we go through your processes and give you a written, fixed-price plan for what is worth handing to AI. If you go ahead, we also build the AI automation, so the people who wrote the plan are the ones who make it work. If AI isn’t the answer, we say so.