Resource · Agentic Workflows & Machine Learning
The AI-Readiness audit.
A practical read of whether your organisation can actually carry AI, before the budget is committed. Six dimensions, one honest score, about ten minutes. From the engineers who built the cloud backend for NASA’s Perseverance rover.
At the end, you walk away with
- A readiness score out of 18, and whether AI can carry the roadmap
- A dimension-by-dimension read, from the use case to the data underneath
- The blocking dimensions to fix before any model work starts
- The option of a written read from us, free, within two working days
- A one-click way to send the whole result to a colleague
Scored on this page. Your answers go nowhere unless you ask for the written read.
Why this exists
Most AI initiatives don’t fail on the model.
They fail on the layer underneath.
The board asks for AI. The value-creation plan assumes it. The engineering team is told to deliver it. And then it stalls, not because the model was wrong, but because the data underneath couldn’t carry it.
It’s a pattern we see again and again. The data architecture hasn’t been meaningfully touched since the last big system change. It worked well enough at the time. It quietly stopped being good enough somewhere along the way. And nobody noticed, because nothing broke, until the AI work needed clean, joinable, trustworthy data, and there wasn’t any.
The cost is rarely the model. It’s the six to twelve months spent discovering that the foundation can’t hold it.
The question underneath the question
Do you actually need AI?
Run the audit, fix what it surfaces, and a harder question appears, one worth sitting with before any budget is committed: how much of the AI roadmap genuinely needs a model at all?
Most AI mandates aren’t really about AI. They’re about outcomes, better reporting, a clean view of the customer, faster answers, less manual work. A surprising amount of that is a data-engineering problem wearing an AI label.
| Often briefed as "AI" | Usually a data engagement |
|---|---|
| "AI for better reporting" | → A clean, joined data model |
| "AI to understand our customers" | → One deduplicated source of truth |
| "AI to automate this manual step" | → A proper pipeline, often rules, not a model |
| "AI to answer board questions fast" | → Data that is trustworthy and queryable |
Real model territory is the genuinely probabilistic and unstructured: extracting meaning from documents, prediction, classification at scale, natural-language interfaces. That work is real and worth doing, but it’s a subset of most roadmaps, and gets cheaper and more reliable once the foundation underneath it is solid.
For each item on the roadmap, ask one question: model problem, or data problem? Most honest lists get shorter.
The pattern, in practice
What happens when the foundation gets fixed.
Four engagements where the work underneath made the work on top possible.
Global compliance advisory
ComplianceCompliance operations were running on Excel and email. We rebuilt them into a modern services platform with proper case management in four months. The board’s automation questions became answerable.
Consilient
Financial Crime DetectionDetection was slow to deploy and environment-bound. We re-engineered it from Apache Spark to Polars: deployment dropped from weeks to about a day, and it now runs in any banking environment.
MMIT
Data ExtractionManual data collection was the bottleneck. We replaced it with autonomous, scalable extraction pipelines on Databricks, so the data the business depended on could finally be trusted.
Pixlise, NASA
Cloud EngineeringWe built the cloud backend for the Perseverance rover’s PIXL instrument, cutting science data processing from 30 hours to 10 minutes. Runner-up, NASA Software of the Year 2023.
Frequently asked
Questions we get
How long does the audit actually take?
About ten minutes and one honest conversation with whoever knows the systems best, or two weeks and a senior engineering team if you want us to run it for you. The six-dimension structure is the same either way; the difference is who finds the answers and how rigorously they’re tested.
Who is this for, engineering leaders, or operating partners?
Both. Operating partners and deal teams use it to size up whether a portfolio company’s AI thesis is real or aspirational. CTOs and engineering leaders use it to size up their own stack before committing to an AI roadmap. The questions are the same; only the vantage point changes.
What if we score low?
You stop the AI roadmap and fix the foundation first. That sounds dramatic but it’s usually two to four months of focused work, not a multi-year rebuild, and skipping it costs three to four times as much when the model can’t get the data it needs. We’d rather tell you that than sell you a model you don’t need.
Do all six dimensions need to be 3 before you start AI work?
No, but the dimensions scoring 0 or 1 are non-negotiable. A single blocking dimension is enough to stall an AI roadmap, however strong the rest. The framework is binary in that sense, get the floor up to a 2 across the board, then optimise toward 3 where it matters most for your use case.
How much of an AI roadmap is genuinely a model problem?
Less than most teams assume. A surprising amount of what gets briefed as AI, better reporting, a clean customer view, faster answers, less manual work, is a data-engineering problem wearing an AI label. Real model territory is the genuinely probabilistic and unstructured: extracting meaning from documents, prediction, classification at scale, natural-language interfaces. Fix the foundation first and part of the roadmap simply resolves itself.
Can you run this audit on a specific company?
Yes, that’s a two-week engagement ending in a plain-English map of the gaps and what it costs to close them. Not a 60-slide deck. Tom Barber and the team run it directly; no layers, no steering committees, no juniors doing the analysis.
Want us to run this audit on a specific company?
Two weeks, ending in a plain-English map of the gaps and what it costs to close them. Yours, or one you’re assessing. Not a 60-slide deck.