Adopting AI without depending on a consultancy: how to build capability in-house

The colleague who prepares reports with AI is away. Someone else finds the instructions, produces the next report and checks it against the source material. That is a useful test of in-house capability: the work continues, and someone in the business knows what to do when a result is wrong.

A small team can start with one task, an approved tool and time to practise. Outside help remains an option. Agree what your employees will learn during the engagement and which services you expect to keep buying afterwards.

What does building AI capability mean for a small business?

Building AI capability means teaching your team to choose a suitable task, prepare the information, give clear instructions and judge the result. A practical first target is for two colleagues to complete and explain the same piece of work without relying on the person who ran the original experiment.

Skills are a documented obstacle. In 2025, 70.3% of EU enterprises that had considered using AI but were not using it cited a lack of relevant expertise, according to Table 7 of Eurostat's 2026 report. The survey covers enterprises with at least ten people employed in the sectors studied; that percentage concerns the group that had considered adoption.

Choose work you already know how to assess: turn notes into a procedure, summarise public documents or draft a customer email. Before opening a chat, write down what the finished work must contain and which mistakes would make you reject it. For help choosing a starting task, see the guide to generative AI for business.

Someone who knows the job needs time to lead the trial. Name a second colleague to practise with them and take over when they are unavailable. You can expand to another task once that handover works.

What does the AI Act require for AI literacy?

Article 4 of the AI Act, amended in July 2026, requires providers and deployers to support the development of AI literacy among staff and others operating AI on their behalf. It takes their knowledge, experience and working context into account, without requiring a guaranteed level for each individual.

For a small business, an initial three-hour session gives colleagues time to practise one familiar job together. Treat the schedule below as a suggested starting point. Completing it does not certify compliance or prepare everyone for every AI use.

Use public or fictional material for this first session. Adapt later AI training for employees to the work they do: a communications colleague needs to check claims and tone, while someone handling spreadsheets must check units, formulas and missing data. Make room for follow-up questions when people apply the method to everyday work.

Build a prompt library that survives a handover

A shared prompt library should hold the instruction, its purpose, an example and the checks needed before using the output. Start in a document or folder the business already manages, with access for the people doing the work and an owner for each entry.

A collection of unexplained prompts gives the next person more to decipher. Pick three recurring tasks and create a short record for each:

• Job and boundaries: what the prompt produces and when to avoid using it.
• Required information: sources, fields and approved data categories.
• Instruction and example: reusable wording with a fictional or authorised test case.
• Tool and model tested: their names, the test date and known problems.
• Review and upkeep: who approves the output, what they check and who updates the record.

The prompt-writing guide explains how to specify context, constraints and format. The Gemini, Claude and DeepSeek libraries provide starting instructions to adapt and test.

This practice prompt uses fictional warehouse notes:

1Write a one-page goods-receiving procedure for a new warehouse colleague.
2Use only these fictional notes:
3- When a delivery arrives, compare the package count with the delivery note.
4- If there is visible damage, take a photo and tell the warehouse supervisor.
5- Save the delivery note in the folder for that day's receipts.
6
7Use numbered steps, followed by questions that remain unresolved.
8Keep the conditions in the notes. Do not invent deadlines, contacts
9or obligations. Turn missing information into a question.

The reviewer checks all three source facts and looks for additions. The notes do not say whether a damaged delivery can be accepted, for example. The draft should leave that decision open. Save this check alongside the prompt. Repeat the test when changing models: keeping the wording makes another tool easier to try, but does not guarantee equivalent results.

Three things never to outsource completely

Keep responsibility for your data policy, working instructions and professional judgement inside the business. Specialists can help with each, while a named colleague remains responsible for understanding and maintaining the agreed approach.

1. Your data policy. Management needs to understand which information may be used, for what purpose and with which service. An adviser can assess contracts and risks; employees still need usable rules and someone to consult about exceptions.
2. Your prompts and test criteria. If you commission them, agree delivery of the instructions, examples and rights needed to use and edit them. Colleagues should be able to find a procedure after the engagement ends.
3. Your judgement. Your business decides whether an answer is correct and appropriate for a customer. Assign someone to check prices, claims, calculations or recommendations before use. The supplier accepting its own work does not replace your review.

Test these arrangements by asking a second colleague to complete a task using the record. Note where they need help. Those gaps identify an explanation, permission or practice session still needed.

A 30-day plan for your first repeatable task

Use the first month to develop one repeatable task with an owner and a colleague who can cover for them. The intended result is a procedure both can execute and review; any time savings must be measured during the trial.

Include a difficult example: a missing date, conflicting notes or a number with the wrong unit. If the system fills those gaps without flagging them, change the procedure or restrict its use.

Count the time spent preparing sources and correcting work, not just generating the first draft. If reviewing the output takes longer than the usual method, stop using AI for that task or try a different one. Buying a tool does not oblige you to extend its use.

When is outside expertise worth paying for?

Bring in a specialist when the task calls for skills your team lacks, such as systems integration, security assessment or a regulated use case. Practical training can also be worth buying when nobody has the time to prepare and support the first exercises.

Automatically moving information from business software, searching internal files with their access permissions intact, or deploying an assistant that takes action on customer records requires design, testing and maintenance. These are separate projects from drafting a document in a chat. Ask for the work and ongoing support to be priced, and retain an internal contact who understands the process.

In the enterprise: make knowledge transfer part of delivery

Some vendors assign engineers to work directly with a customer's team, often called forward-deployed engineers or FDEs. Cohere describes this approach and argues for developing the customer's capabilities during the engagement. Turn that aim into agreed deliverables: joint working sessions, usable documentation, tests your staff can repeat and access rights for the agreed components. Include a handover rehearsal and an exit plan covering owners, export formats and expected support. Some dependence on the model or platform may remain; spell out what your team will be able to maintain and what will still require the vendor.

A workspace for practising with useful outputs

ilisai brings available models from several providers into one workspace for drafting, research with cited sources and downloadable document generation. Each colleague uses their own account and credit balance. Your team can choose a model for each exercise while keeping its prompt library in the company's document repository; connecting business systems and maintaining custom applications requires other tools and, where appropriate, specialists. Check the plans and credits when setting a spending limit for the trial.