An assessment of where AI genuinely helps your business, what it would cost, and what to ignore — grounded in your processes rather than in product capability.
Most businesses do not need a strategy document
They need answers to two or three specific questions, and the questions are usually the same ones.
Is Microsoft 365 Copilot worth buying for our people? Is there a repetitive process here that could be automated? What are our rules for staff pasting company information into AI tools they found themselves?
That is a short, concrete piece of work with a decision at the end of it. Businesses that commission a comprehensive AI strategy before answering any of them tend to end up with a well-formatted document and no change to how anyone works.
Start with where the time goes
The useful starting point is not what AI can do. It is where your business currently loses time, produces errors, or cannot find information.
Once that list exists, matching it against current capability is straightforward, and most items get discarded quickly — either because AI would be unreliable at them, or because a much simpler automation would do the job better and cheaper. That discarding is most of the value.
The shape of a good candidate
The tasks where current AI genuinely earns its cost share a pattern: the person receiving the output can tell whether it is right, and an occasional error is recoverable.
Summarising a long email thread fits. The reader knows the context and would notice a wrong summary. Drafting a first version of a routine document fits, because someone edits it anyway. Finding relevant material across a large document set fits, because you open what it finds.
Calculating a figure that goes into an invoice does not fit. Neither does anything where a confident, plausible, wrong answer would pass unnoticed into something that matters.
Be clear-eyed about fabrication
The failure mode that catches businesses out is not that AI is sometimes wrong. It is that it is wrong in exactly the same confident, articulate register as when it is right.
A tool that failed obviously would be easy to manage. One that produces a fluent paragraph containing an invented figure, a misattributed quote or a regulation that does not exist requires the person using it to verify rather than trust — and that is a behavioural change, not a configuration setting.
Any deployment plan that does not address it is incomplete. That is what AI policy and governance covers.
Buy before you build
Custom development should be the last option, not the first. It is expensive, it needs maintaining as the models underneath it change, and it very often reproduces something an existing product already does.
The legitimate cases exist — a process specific enough to your business that nothing off the shelf addresses it, and valuable enough to justify building and maintaining. See custom AI solutions for when that is genuinely true. It is a much smaller category than the current enthusiasm suggests.
We will tell you to do nothing if that is right
For some businesses, at the moment, the honest assessment is that nothing available justifies the cost, the change effort and the verification burden.
That is a legitimate outcome of a strategy engagement and we will say it. The alternative — recommending a deployment because you paid us to recommend something — costs you more than the fee.