AI implementation step by step

An AI implementation in a small business can start with a task someone does every week: copying data from emails, preparing a report or finding information in company documents. It is worth looking at that work first. How long does it take? What repeats? Where does someone need to make a judgment, and where do they simply move data?

That gives you a small scope to test. It becomes easier to see whether the tool helps and what needs changing.

Find a task you can describe

“We want AI in sales” is not enough to plan a project. You need a description of what a person does. For example: they open a form submission, check the attachment, choose the service type and copy the details into the CRM. If the enquiry is incomplete, they reply asking for more information.

These questions help prepare for the first conversation:

What gets copied from one place to another? From email to spreadsheet, spreadsheet to CRM, PDF to a system?
Who does the task, and how often?
What takes the most time?
What happens when information is missing or an unusual case comes up?

Go through a few recent examples with the person who handled them. You may find that locating the current price list is the main difficulty, while writing the reply takes very little time. That changes the project’s scope from what the initial description suggested.

Some tasks can also be simplified without a language model. If the data is already in fixed form fields, an ordinary connection between tools may be enough to move it into a spreadsheet.

Agree what should change

Start with one part of the work whose output you can check. In the enquiry example, that might mean preparing a draft reply for a salesperson to approve. Automatic sending can wait until you know what kinds of errors appear in the drafts.

Before building anything, record how the work is done now and what would count as a useful result. An example target: reduce quote preparation from 40 to 20 minutes, including review and corrections. That is a target for a test, not a measured result from a completed project.

Time alone is not enough. You also need to check that the quote contains the right prices, service scope and customer details. If drafting takes a minute but corrections take half an hour, the saving will be smaller than the model’s speed suggests.

At this stage, agree the trial budget, who will assess the results and the conditions for continuing. The post on AI implementation costs explains how to separate the build cost from ongoing usage.

Give the team a version to try

A pilot needs examples that resemble everyday work. Include an email with a missing attachment, an upside-down scan and a message saying “the same as last time, just a different size”. These cases reveal missing information and show when the tool should ask for help.

The same set can help you compare AI models on business tasks before choosing one for the implementation.

Agree the scope of data access before connecting the tool. Some features can be tested using materials with customer details removed. Testing on company documents requires access and data handling arrangements suited to the project.

The person trying the working version should have an easy way to flag what went wrong. Specific notes help: “used last year’s price”, “missed the second product”, “asked for the order number unnecessarily”. They identify what to fix and provide an example for checking the change.

An unsuccessful pilot does not establish that the task cannot be automated. The cause might be missing data, poorly configured search or too broad a scope. Work out what failed and assess the cost of another attempt. Sometimes a correction is enough; sometimes it makes sense to stop the project there.

Prepare the tool for everyday use

Before a permanent rollout, there are issues that a successful demonstration can easily hide. What happens if the CRM does not respond? Who gets notified of the error? Could retrying the task create a second quote for the same customer?

Someone inside the company also needs to know where to check the output, how to report a problem and how to return to manual processing. The changes a team can make themselves depend on the tool and their skills. An employee might be able to edit a short piece of text, while changing the connection to the invoicing system requires the implementer.

In my projects, I hand over the elements created for the solution according to the agreed scope. I can also provide ongoing support covering updates and fixes. It helps to establish who is responsible for the integrations and who keeps the price list or instructions used by the tool up to date.

After launch, return to the measurements from the start of the project. Check working time including output review, the number of corrections and running costs. Use those results to choose the next part of the process to develop.