Where to start with AI? Choose one business process

Enquiries keep arriving in the inbox. A colleague opens a message, reads the attachment, copies details into a system and works out what the customer forgot to mention. That specific stretch of work is worth examining when a business considers AI. “Let's introduce artificial intelligence” is too broad a brief to let you verify the outcome.

A first project needs a clear beginning, an end and someone who can recognise a correct result. The following is our practical approach to choosing that project, not a promise of a particular saving.

Name the task before choosing the tool

Write one sentence: “From an incoming service enquiry, we will prepare a summary of the details and draft questions about missing information for a member of staff.” That identifies both input and output. “Automate the sales department” does not say where the system should stop.

Check the simpler option too. Moving a message based on an exact sender address may only need an email rule. If customers keep omitting the same detail, a required form field might help. The UK government's AI Playbook also notes that established technology can sometimes solve a problem more easily than AI.

AI is worth testing where differently worded text needs to be processed. Even there, do not assume that a well-written answer contains the correct details.

Compare candidates using work you understand

List recurring tasks and answer these questions for each. Support the answer with an actual example, not an impression from a supplier's presentation.

• Does it happen often enough? Check the actual number of cases and staff time. An occasional task may not justify setting up and maintaining a solution.
• Do we have usable input? Can you obtain the messages or documents and establish which data may be processed? If not, resolve access and source material first.
• Can we check the result? A staff member must be able to compare the draft with its source and spot omitted information.
• Can an error be stopped? An internal draft awaiting approval carries a different risk from automatically sending a price to a customer.
• Who will use it? Involve the person doing the task today. Find out whether the new process adds yet more copying between windows.

The best first candidate need not be the company's biggest problem. Choose a part of the workflow where you can obtain a trustworthy trial result without rebuilding an entire department.

Example: prepare an enquiry, not a repair commitment

Consider a hypothetical service company. It wants to extract the equipment, the reported problem and the customer's requested date from a message. If the equipment model is missing, the system should draft a follow-up question. Each completed field should point to the relevant part of the message; missing details should be flagged, not invented.

A staff member compares the draft with the original, corrects it and only then sends the reply. A requested date must not become a confirmed appointment. Pricing, binding orders and unchecked writes to the service system are outside this first trial.

This is a possible implementation design, not a description of an automatically enabled Informio feature. Connecting an inbox or an internal system needs separate assessment and agreement.

Keep the trial brief to one page

Before preparing the solution, write down the following with the staff member:

1. Scope: which enquiries are included and which will still follow the existing process.
2. Inputs: where source material comes from, who will provide access and which data is excluded from the trial.
3. Output: the exact fields, how missing information is marked and where the person will see the original.
4. Review: who approves the draft and how a case returns to manual handling if something goes wrong.
5. Decision: what you will count as a benefit and which errors will stop the trial.

Include an incomplete enquiry, a conflict between the message and its attachment, and two devices mentioned in one message. Set some cases aside for the final check rather than using them to refine the instructions.

Measure completed work, not the speed of the first draft

Record the time taken by the existing process and by the new one, including reading, corrections and transfer to the next system. Track what staff corrected: a wrong model, an omitted request or an invented date. The number of drafts generated does not tell you whether work became easier.

The NIST AI RMF recommends documenting test sets and metrics and evaluating behaviour under conditions similar to deployment. In your trial, compare similar types of cases and retain the reasons a draft was rejected.

If someone has to rebuild the entire result, include that case too. If attachments are the only problem, the finding may support a narrower scope instead of a full rollout. Include usage, maintenance and review costs in the assessment as well.

Expand what has proved useful in practice

At the end of the trial, decide whether to keep the limited scope, change the brief and test again, or not deploy the solution. Discovering that a better form is enough is a useful result too. Add another process only when you understand how the first works and where it falls short.

To get started in your business, bring a description of one task and a sample without sensitive data to an initial consultation with Informio. Together we can assess a suitable implementation scope. If your main need is answering website visitors' questions, you can also try Informio's AI chat separately.