Most organisations accept that AI belongs somewhere in their operations. The difficulty is the first step, and the most common instinct about what that step should be is the wrong one.
The instinct is to begin with the most visible, highest-value use case: the client proposal that drafts itself, the week-long analysis compressed into minutes, the demonstration that impresses a board. It looks like the fastest route to proving value. In practice it is the fastest route to eroding confidence in the technology.
Begin where a wrong answer is cheap
High-value work is precisely where a confident, fluent, incorrect output does the most damage. Current AI systems are very good at producing plausible material and considerably less reliable at being correct. The gap between the two is hardest to detect in exactly the high-stakes documents where it matters most. Leading with that kind of work builds an efficient mechanism for producing expensive mistakes at speed.
The better starting point is the opposite: a task that is mechanical, low in stakes, and immediately verifiable. Month-end reconciliation is a representative example. The figures either reconcile or they do not, and the result is visible in seconds. No reputation rides on it, and no specialist is needed to judge whether the output is correct.
That verifiability is the value. A task that can be checked instantly teaches an organisation what the tools do well and where they fail, without anyone having to trust the output on faith.
Adoption stalls on capacity, not willingness
In most organisations a small number of people experiment with AI on their own initiative while the majority do not. It is easy to read the second group as resistant. More often the constraint is capacity.
Experimentation requires slack: unstructured time to attempt something, watch it fail, and adjust. Most roles are already full. Procuring a tool and announcing it is straightforward; making sure anyone has the time to actually use it is not. Usage, not procurement, is where these initiatives quietly stall.
Time and permission are the real budget
The binding constraint on AI adoption is rarely the licence fee or the choice of model. It is protected time, and explicit permission to spend that time on work that may not succeed.
An organisation that genuinely wants to find where AI helps has to allocate the hours and make clear that spending them on an uncertain experiment is the assignment, not a distraction from the real job. Without that, even a well-chosen tool sits unused.
In practice
- Pick one mechanical, checkable task: reconciliation, data tidying, routine formatting. Work where success is self-evident and failure costs nothing.
- Give it to one person and protect their time, so the experiment is their actual responsibility for a defined period rather than something squeezed around a full workload.
- Treat the findings as the deliverable. Where the tools saved real effort, and where they failed confidently, is worth more than the task itself.
- Escalate only once they have proven reliable on low-stakes work. Then point them at work that carries consequences.
Proving the tools on inconsequential work is what earns the judgement to identify which larger initiatives are worth committing to, and which would have quietly damaged credibility had they gone first.
The next constraint, once a task is chosen, is almost always the data beneath it. That is where most of the real work turns out to be, covered in getting your data AI-ready. For organisations that would rather not make the first move alone, it is also the work our AI transformation practice exists to do.
