DATA AND AI
What data platform do you actually need?
Most teams buy a platform sized for a company ten times larger, then struggle to staff it.
SERVICE 02
Numbers people stop arguing about, then models worth building.
Data platforms, analytics and machine learning — in that order, because almost every AI project that fails does so for reasons that were visible in the data layer first.
THE PROBLEM
Two dashboards disagree and the meeting becomes about the dashboards
When metrics are computed in several places by several people, they drift. Once they have drifted, every discussion that depends on them turns into a discussion about them, and decisions stall.
The temptation is to buy a tool. Tools do not settle definitions. The work is deciding what a number means, computing it in one place, and making that the only place anyone looks.
WHO IT IS FOR
Who data and ai is designed for
Teams whose reporting is manual
Someone spends two days a month assembling a deck from spreadsheets.
Teams with a trust problem
The numbers exist, and nobody quite believes them.
Teams considering AI
There is pressure to do something with AI and no clear read on which ideas are real.
Teams with data and no platform
Everything is in production databases and queries against them are starting to hurt.
WHAT IT INCLUDES
What this work may cover
Not every engagement covers all of it. The shape depends on what you are deciding.
Source mapping
Where every number comes from, and why two versions of it disagree.
Warehouse and modelling
A platform sized for your organisation rather than for a conference talk.
Metric definitions
Written down, computed once, used everywhere — the unglamorous part that does the work.
Analytics and dashboards
Few, maintained, and tied to decisions somebody actually makes.
ML and LLM implementation
Where there is a real case: evaluation first, then the feature, then the monitoring.
THE FRAMEWORK
How the work is sequenced
- 01
Trace
Follow the disputed numbers back to source and find where they fork.
- 02
Define
Agree what each metric means, with the people who argue about it in the room.
- 03
Build the layer
One modelling layer, versioned and tested like the rest of your code.
- 04
Evaluate before you build
For anything ML or LLM: decide how you will measure it working, before writing it.
- 05
Hand over
Your analysts extend the models without us, or the project has not finished.
ILLUSTRATIVE EXAMPLE
A company wanting an AI assistant over its internal documentation
A constructed example to show the shape of the work, not a client case study.
- 1
Establish what question people are actually asking, and how often — often far narrower than imagined.
- 2
Check the documentation is correct and current. Retrieval over stale content produces confident, wrong answers.
- 3
Build an evaluation set of real questions with known good answers before writing any feature.
- 4
Ship the narrow version to a small group and measure against that set.
- 5
Decide on the evidence whether to widen it, change the approach, or stop.
RELATED READING
Worth reading alongside this
DATA AND AI
AI features that survive contact with users
The demo is the easy part. Everything that makes an AI feature usable happens around the model.
WORKING TOGETHER
How to brief a consultancy
The brief determines the outcome more than the firm does. Bring the problem, not the solution.
QUESTIONS
Common questions about this work
Should we be doing something with AI?
Sometimes. The question worth answering first is which specific task would change if it worked, and how you would know. If that cannot be answered concretely, more platform work usually returns more than a model would.
Do you build custom models?
Rarely, and usually it is the wrong call. Most useful work is retrieval, evaluation, prompt and data engineering around existing models. Training from scratch is a large bet that needs a specific reason.
Which warehouse should we use?
It depends far more on your data volume, your team and your existing cloud than on the vendors involved. For most mid-sized organisations several options would work and the choice is not the risky part.
Settle the numbers first
Most data and AI work gets easier once everyone agrees what the metrics mean.