A medium-sized company could do something quite sensible tomorrow: buy ChatGPT licences for its employees, give people some basic guidance and tell them to experiment. I would probably do it. What I would not do is mistake that for the company’s AI strategy.
Giving people access to a capable general-purpose AI tool is a bit like giving everyone spreadsheets or internet access. You cannot predict all the useful things they will discover. A finance analyst may automate part of a monthly reporting process, a salesperson may research accounts faster, someone in HR may reduce the time spent drafting job descriptions, and a developer may use it to understand unfamiliar code. Those gains are often small individually, but they can accumulate across an organisation, and trying to design all of them centrally would be hopeless.
There are good reasons, then, to make AI widely available. It creates familiarity, reduces the risk that AI becomes something done by a specialist innovation team while everyone else carries on as before, and it reveals where the real opportunities are. Employees usually know far more about the irritating, repetitive and unnecessarily difficult parts of their jobs than an AI steering committee ever will.
The problem comes when those local gains are confused with business value.
If everybody saves twenty minutes here and forty minutes there, the company may end the year with thousands of claimed hours of productivity and almost no detectable improvement in its financial performance. This is already one of the central problems with AI adoption. Organisations measure what is easy to count, so they report licences activated, prompts submitted, employees trained and applications created. Then somebody on the board asks whether revenue grew faster, costs fell, customers stayed longer or products reached the market sooner, and the connection is often surprisingly hard to make.
A sensible AI programme therefore needs two things happening at the same time. The first is broad experimentation. Give employees good tools, establish clear security and data rules, train people enough that they understand what the technology can and cannot do, and encourage them to use it on real work. There is no need to demand a business case every time someone wants to save half an hour preparing for a meeting.
The second activity needs much more discipline. Choose a small number of business outcomes that already matter and deliberately use AI to try to change them.
Suppose a company wants to increase revenue from existing customers. That is a much better place to start than a list of AI technologies. Ask what is preventing that revenue from growing faster. Perhaps new functionality takes nine months to reach customers. Perhaps salespeople spend too much time preparing proposals. Maybe account managers cannot identify which customers are most likely to buy an additional product, or onboarding takes so long that customers reach useful adoption months later than they should.
Now there is something concrete to work on.
Take the software example. Imagine a product feature currently takes nine months from approval to customer availability, and the company believes AI-assisted development could reduce that to six. The interesting measure is not lines of code written or even developer productivity in isolation. The interesting measure is whether those three months change anything commercially. Customers can use the product earlier, some may adopt it earlier, and some may therefore pay for it earlier.
The value chain is fairly simple: faster development leads to earlier availability, which may lead to earlier adoption and earlier revenue. That is now a hypothesis worth testing.
Before changing anything, establish the existing baseline. Measure how long comparable work currently takes, run a controlled experiment with a real product team, and observe what changes. Then keep following the chain towards the business outcome. If development becomes 30 per cent faster but product revenue does not move, that is not necessarily a failed experiment. It may reveal that engineering was never the real constraint. Product decisions may still take six weeks, security approval may take another month, sales may not know how to sell the new capability, or customers may simply not want it.
That is useful information because AI often makes one part of an organisation faster and exposes the next bottleneck.
This is why I would be cautious about launching a large collection of centrally sponsored AI projects. A medium-sized business does not need fifty pilots. It needs a handful of experiments where success or failure can actually be observed.
For each one, write down the argument before spending serious money. What business measure are we trying to change? What operational measure should move first? Why do we believe changing that measure will influence the business result? What happens today, and what would probably happen if we did nothing? Most importantly, what evidence would convince us that the experiment worked?
That can fit on one page and is much more useful than a thirty-page AI strategy.
There is another reason to separate broad employee adoption from strategic AI investment: they have different economics. Giving hundreds of employees ChatGPT may cost a meaningful amount of money, but it is relatively reversible. If people use it and find value, continue. If they do not, reduce the licences.
Changing a core operating process is different. Rebuilding product development around AI, automating customer support or introducing AI into underwriting, compliance or financial operations changes how the company works. Those decisions deserve proper measurement, controls and management attention.
The mistake is treating these two activities as though they are the same.
The better model is to let employees explore widely while the company invests narrowly. A central AI team, if there is one, should spend less time dreaming up use cases and more time finding places where local experimentation has uncovered something economically interesting.
Imagine that forty salespeople independently start using AI to prepare for customer meetings. That is worth paying attention to. The next step is not necessarily to buy a new enterprise sales platform. It is to study what is happening. Does preparation time fall? Do salespeople conduct more meetings? Does opportunity creation increase? Does conversion improve? Does revenue per salesperson move?
If the causal chain survives examination, industrialise it. If it does not, leave it as a useful personal productivity tool and move on.
This distinction also helps with one of the most awkward claims being made about AI productivity. Employees will quite reasonably report that AI saves them hours each week, and vendors will publish studies showing impressive efficiency gains. Some of those numbers will be true, but an hour saved is not automatically an hour of economic value.
If an employee finishes the same work at 4:30 rather than 5:00, that may be a genuine improvement in their working life and worth having for that reason alone. But unless the organisation does something with the released capacity, it should not quietly turn thirty minutes into a financial return in a spreadsheet.
Management has to decide what it wants to do with the productivity. It might allow people to handle more customers, build more product, reduce external spending, avoid hiring that would otherwise have been necessary, improve service or shorten a process that delays revenue. Those are measurable consequences, and that is where the business case becomes real.
For most medium-sized organisations, I would start with exactly this combination. Give knowledge workers access to AI where the economics make sense, let them experiment, and accept that some of the benefit will remain informal and difficult to measure. At the same time, choose perhaps three strategically important problems and attack them with much greater discipline.
After a year, I would rather see a company with widespread everyday AI use and three demonstrably valuable changes to how the business operates than one with seventy-five AI projects and an impressive slide showing that 82 per cent of employees have used a chatbot.
Adoption matters, but what ultimately matters more is whether the business has actually changed because of it.
