AI/ML Consulting
Somebody has asked for an AI strategy, and the list of proposed use cases keeps growing. We work out which of them are worth building, which are not, and what has to be true first.
A pilot has been running since the spring and still has no owner, no budget line and no decision attached to it. Meanwhile the list of proposed use cases grows every time someone reads an article. The question worth answering is which decisions in your organisation are slow or expensive enough to be worth changing, and whether AI is what changes them.
Start from decisions, not from a platform.
Technology first
Pick a platform or a model provider, stand up a few pilots, and see which ones get traction. It builds real skill fast and gets people using the tools.
- Skills arrive fast, because people are using the tools weekly
- Approval is easy: a pilot is cheap and reversible
- Ranking is by enthusiasm, since every pilot demos well
- The pilot that works still has no owner in the business
Decision first
Start from the decisions that are slow, expensive or inconsistent today, and work back to whether AI changes any of them. Some answers turn out not to be AI.
- Each candidate carries the decision it changes and who owns it
- Ranked on value and on whether the data can be reached
- Fewer pilots run, and the ones that run have a sponsor
- Slower to show something on screen, because ranking comes first
Most stalled pilots did not fail on the model
Three questions decide whether anything gets past a demo. Does the data exist, in one place, with someone who can grant access to it? Will anyone change how they work because of the output? And when the system is wrong in front of a customer, whose name is on the decision? The first is an engineering problem and usually the smallest of the three.
The other two are organisational, and they are where pilots die. A model that scores well changes nothing if the branch manager keeps overriding it, or if nobody can say who signs off a wrong answer. So the readiness work asks for names against each: the person who will change a process, and the one accountable when the output is wrong. If neither exists yet, that gap is the first thing to fix.
How an advisory engagement runs
Most of this work is interviewing and arithmetic rather than workshops. We talk to the people who own the decisions and the people who own the data, then cost the options. It runs to a fixed artefact list and a date.
- 01
Frame the question
The question on the slide is rarely the question. Cost pressure, a competitor's announcement and real operational pain lead to different strategies, and they cannot all be served at once. You get a one-page statement of what this work will answer.
- 02
Find the decisions
We interview the people who run operations, not only those who sponsored the work, looking for decisions made often, made slowly, or made differently depending on who is on shift. You get the list, with how each one is handled today.
- 03
Test feasibility
For the short list we go and look: where the data lives, who can grant access, and how far it is from usable. Some candidates die here, which is the cheapest place to die, and you see the evidence behind each verdict.
- 04
Cost the options
Build, buy or leave alone, priced with the running cost as well as the build: inference, the people who review outputs, and the monitoring nobody budgets for. Options come with their assumptions attached, so you can argue with a number rather than accept it.
- 05
The recommendation
The last session is a rehearsal rather than a presentation: we take the sceptical questions first, so the paper holds up when we are not there. What you keep is a ranked portfolio, a costed first phase and the reasoning behind both.
What the engagement produces
| Output | What it answers | Who uses it |
|---|---|---|
| Ranked use-case portfolio | Which candidates are worth funding, in order | Sponsor and board |
| Readiness assessment | Whether the data and platform can carry it | CTO and data owners |
| Build-versus-buy view | What to buy, what to build, what to skip | Engineering and procurement |
| Risk and governance position | Who is accountable when the output is wrong | Risk owner and sponsor |
| Costed first phase | What the first project costs to build and run | Finance and delivery lead |
Before you commission this
Then we will say so, and say what to do instead. For a good number of organisations the honest first move is a reporting layer people trust, or a single owned customer record, or fixing the process that the model would have been asked to paper over. That advice is worth the fee, and it is cheaper than a year of pilots.
Not sure which one you need?
Describe the problem in a paragraph and we will tell you which service applies.