AI strategy & agents
Work out where AI genuinely earns its place in your organisation, then build the parts that do. Grounded in your own data, not a generic pilot.
The problem this solves
Two failure modes dominate. The first is paralysis: everyone agrees AI matters, nobody owns it, and the conversation loops for a year. The second is pilot sprawl: a dozen experiments, all technically interesting, none connected to a process anyone depends on, and no route from demo to production.
Both come from starting with the technology. The useful starting point is where your people lose time, where knowledge is hard to reach, and where decisions are slow because the data is hard to get at. Some of those have a good AI answer. Many have a better non-AI answer, and saying so is part of the job.
What I do
AI readiness assessment
An honest picture of where you are: what data exists and in what state, what tooling you already pay for, what skills sit in the team, and what policy or governance constraints apply. Most organisations are further along in some areas and further behind in others than they assume.
Use case discovery and prioritisation
Structured sessions across the business to surface candidate use cases, then a consistent assessment of each: value, feasibility, data dependencies, risk and effort. The output is a prioritised shortlist with a clear recommendation on what to do first and, just as usefully, what to leave alone.
Semantic layers
An AI assistant is only as good as its understanding of your business. A semantic layer gives it a reliable, governed view: what a customer is, how revenue is defined, which measures are certified and which are not. This is the piece most often skipped, and the reason most business data assistants give confidently wrong answers.
Agent design and build
Agents that let people ask questions of business data or internal knowledge in plain English and get answers they can trust, with the underlying source visible. Scoped narrowly on purpose: a well-defined agent that reliably answers one class of question beats a general assistant that is unreliable at everything.
Governance and guardrails
Practical rules on what data may be used, how outputs are verified, where a human must stay in the loop, and how usage is monitored. Sized to your organisation and your regulatory position, and written so people can actually follow it.
Enablement
Training that raises the floor across a team, focused on what these tools are good and bad at, how to prompt effectively, and how to recognise a plausible-sounding wrong answer. Often the highest return activity in the whole engagement.
How an engagement runs
Scoping call
Free, around 45 minutes. What is prompting the question, what has been tried, and what would success look like in six months.
Assessment
Interviews, a review of data and tooling, and a written readiness picture with a prioritised use case shortlist.
Pick one and prove it
Build the highest-value case properly, with real data and real users, rather than five demos at once.
Productionise
Guardrails, monitoring, documentation and the operational bits that turn a prototype into something the business can rely on.
Scale or stop
An honest review against the measures we agreed. If it did not pay off, we stop and say why. That is a valid outcome.
What you get
- A readiness assessment covering data, tooling, skills and governance
- A prioritised use case shortlist with value, effort and risk for each
- A working implementation of the first case, in your environment
- A semantic layer where business data is involved, documented and reusable
- Guardrails and policy written for the people who have to follow them
- Team training so capability stays with you
Questions I get asked
Are you going to tell us to build our own model?
Almost certainly not. For nearly every organisation the sensible route is established models via existing platforms, with your value coming from the data and context you connect to them. Training your own model is rarely the right first move.
Our data is not clean enough for AI.
Sometimes true, often used as a reason to do nothing. Some use cases need clean, governed data and should wait. Others, particularly knowledge retrieval, work fine on messy input. Part of the assessment is telling you which is which so you can start somewhere real.
Can you help without touching sensitive data?
Yes. Plenty of the work is strategy, prioritisation, governance design and enablement, none of which requires access to production data. Where a build does need access, it happens inside your environment under your controls, with whatever agreements your legal team requires.
What if the honest answer is that we should not bother?
Then that is what the report says, with reasoning. A clear no, arrived at quickly, is cheaper than an eighteen-month programme that quietly fizzles.
Somebody has to own the AI question.
If nobody in your organisation currently does, that is usually where to start. Tell me what is being asked and by whom.