Buying AI licences is easy but getting people to change how they work, use the tools safely and create enough value to justify the cost is much harder.
The answer is to treat AI adoption as a business change programme, not a software deployment. Start with real work, combine broad training with targeted improvements, and measure the results before scaling.
That approach works with Microsoft Copilot, Claude, ChatGPT Enterprise or a combination of tools.
A lot of businesses ask: “What can we do with AI?”
But a better question is: “What is slow, repetitive, difficult or frustrating today?”
That could be preparing a board pack, responding to due diligence questionnaires, reviewing contracts, summarising research or finding knowledge buried across email and documents.
Once the problem is clear, decide whether the answer is an AI assistant, reusable prompt, agent, automation, process change or something else. Sometimes AI is only part of the solution.
This matters commercially. A catalogue of AI ideas creates activity. A prioritised list of business problems creates an investment plan.
The strongest AI rollout programmes contain two connected tracks.
The two tracks reinforce each other. Training surfaces valuable problems. Targeted projects produce examples that make wider training more credible. People use AI when they can see how it improves work they recognise.
Uncertainty stops adoption. If staff do not know which tools are approved, what information they can use or when human review is required, many will avoid AI. Others will experiment without support.
The policy does not need to answer every future question. It should give people practical guidance on:
Data and permissions need attention. AI can make overshared information much easier to find. Deal with the highest-risk areas first instead of delaying the programme while waiting for perfect data governance.
Generic demonstrations create interest, but rarely create lasting habits. People need to practise with tasks, files and decisions that resemble their day-to-day work.
A useful model is a short foundation session for everyone, followed by smaller sessions organised around working patterns. Finance may focus on spreadsheets and reporting. Legal and compliance may work on document review and contracts. Marketing may concentrate on content and presentations.
Before each session, speak to someone who understands the work and select two or three genuine use cases. Solve them in advance, build them into the exercises and show how to check the output. This creates repeatable ways of working.
Training should cover judgement as well as features. AI can produce a polished answer that is wrong. Staff need to know when to challenge it, how to verify it and when not to use it.
One training session is not an adoption plan.
Give participants one or two weeks to apply what they learned, then bring them back for a short follow-up. Ask what they used, what worked, what failed and what value it created. The commitment to return creates a useful forcing function. It also produces evidence, better prompts and new use cases.
Support this with champions, although “change champions” may be better than “AI champions”. They surface problems, share working practices and help colleagues without assuming that every problem needs AI.
Weekly team meetings can also help. A simple question such as “What value did AI add this week?” is more useful than asking whether people used it.
Usage data matters. If someone has a paid licence and does not use it, the licence should be reassigned. But logins and prompt counts do not prove a return.
Measure a small number of practical outcomes for each prioritised use case:
For time-based benefits, use a conservative calculation: time saved, multiplied by task frequency, adoption rate and an appropriate employment cost. Then compare the annual benefit with licences, training, implementation and ongoing support.
Not every benefit converts neatly into cash. Time saved matters if it increases capacity, improves service or removes cost. Quality, resilience and risk reduction may be just as important, but still need evidence.
In the first 30 days, set the baseline. Confirm approved tools, review policy and data risks, analyse current licence usage, identify business pain points and select a small number of priority use cases.
In days 31 to 60, run foundation training and function-specific workshops. Test priority workflows with representative users. Give each shared prompt, agent or automation an owner, and record expected benefits before the pilot starts.
In days 61 to 90, hold follow-up sessions, collect evidence, publish the approaches that worked and stop or redesign those that did not. Reassign unused licences. Use the results to decide where wider licensing, deeper automation or further training is justified.
The aim is not to complete AI adoption in 90 days. It is to create enough safe, repeatable use and credible evidence to make the next investment decision.
AI pays for itself when it becomes part of useful work. Start with business problems, train people in context, reinforce the new habits and measure what changed. That is how adoption becomes value rather than another line in the software budget.
Trying to pinpoint where to start can be difficult. If you have any questions about where your firm is on the AI adoption journey, contact us using the form below. We can help with everything from use case development to implementation and policy creation.
Or keep up with our bi-weekly LinkedIn Newsletter all to do with practical Copilot and M365 tips for everyday work.