How do you actually make AI work in a business like ours?
Outcome first. Then the data it needs. Then the smallest build that proves it.
We agree what success means, get the data that one job needs, and build the smallest thing that proves it.
What comes first?
The outcome. Everything else follows from it.
Every engagement runs this chain, in this order. The technology, if there is any, is decided fourth.
Business problem
What is slow, expensive or manual, in the words of the people doing the work.
Measurable outcome
What better means as a number, agreed before anything is built. Ideas ranked for value, data readiness and risk.
Data required
The minimum data that outcome needs, and whether it can be trusted for that job.
Technique
A model, a pipeline, a statistical analysis, or no AI at all. Chosen from the problem backwards.
Evaluation
Tested against your real cases, on the measure agreed in step two.
Deployment
Live in your estate, at the level of autonomy the evidence supports.
Why do most AI projects fail?
Six common reasons.
- 01
Planned for the business you wish you had
We build for the one you have, with the data you have today. - 02
Automating a process that was already broken
We map how the work really runs before we automate it. - 03
No agreed way to measure success
We agree the measure before we start. - 04
Only one side of the business in the room
Business, marketing and technology, together from day one. - 05
A demo instead of a test
Tested on your real cases before anyone sees it. - 06
Handed over with nothing to hand over
You get documentation, monitoring and a team that can run it.
What about the data?
Enough trustworthy data for the job. Not perfect data for everything.
We make the data that one job needs trustworthy, then widen. We do not clean everything first.
What does ready mean?
Six questions we ask about your data before we build anything.
Where did it come from, and who owns it?
Data recorded for one purpose often misleads when used for another.
Is it recent enough for the decision?
A weekly export is no use to a system that has to answer today.
Who is allowed to see it?
Permissions have to survive the journey into the system, not get lost on the way.
Have you ever recorded the thing you want to predict?
You cannot forecast churn if nobody has ever written down who churned.
Does it include the awkward cases, or only the easy ones?
The difficult cases are the ones the system will be judged on.
Does it actually describe the decision you are making?
Data about a neighbouring decision looks similar and gives the wrong answer.
Our prototype
Campaign Intelligence
Campaign Intelligence is our working prototype of what the answer looks like once marketing data is joined: one model you can question in plain language, with the numbers behind the answer.
It was built on synthetic data to prove the output. Joining your real systems, and hardening that for production, is the larger part of the work. It is done per client, not bought off a shelf.
Six sources, never designed to talk to each other
Why did email revenue drop last month?
Sends were down 18 per cent after the Black Friday list suppression was left on. Open rate actually improved, so the drop is volume, not engagement.
Prototype output, shown on synthetic data. Joining your real sources is the engineering work.
How do you build it?
The thin line.
Define success. Build the smallest end-to-end thing that proves it, on the data you have today. Evaluate it against real cases. Iterate. Then widen: another source, another process, more autonomy, each on top of something already running. Live in weeks, not months, because the first line is narrow on purpose.
The three rejected paths on the left are the analogue intelligence. Anyone can build the line. Knowing which line to build, and which three to leave alone, is the part you are paying for.
Why choose us?
Anyone can build the line.
Knowing which line to build, and which ones to leave alone, is what you pay us for.
Where do I start?
Tell us what is slow, expensive or manual.
We will tell you honestly whether AI is the answer, what form it would take, and what a first build looks like. If the honest answer is not yet, we will say that too.
