Data & AI

There is a consistent failure pattern that has almost nothing to do with model quality. A pilot is built quickly on whatever data is reachable, it demonstrates something genuinely impressive, and it receives enthusiastic sponsorship. Then it approaches production and encounters governance for the first time.

The questions that arrive at that boundary are not unreasonable: what is the provenance of this data, does the model surface information a given user is not entitled to see, can we reproduce a specific output from six weeks ago, what happens when the source system changes shape. Each is answerable in a well-governed platform and unanswerable in an ad-hoc extract.

So the pilot stops. Not because anyone decided against it, but because the work required to answer those questions was never scoped, and it is substantially larger than the pilot itself. The organisation concludes that AI is harder than expected, when what it actually discovered is that its data foundation was not production-grade.

The order that works is unglamorous: governed platform first, then AI applied to specific decisions where the value is measurable. It is slower to the first demo and much faster to the first thing in production, which is the only measure that eventually counts.

Three departments reporting three different revenue figures is not a reporting problem. It is usually three defensible definitions - different treatment of returns, different recognition timing, different currency conversion date - that were each correct for their original purpose and have never been reconciled with each other.

The cost is subtle and large. Meetings spend their first portion establishing whose number is being discussed. Decisions get deferred pending reconciliation. Analysts spend the majority of their week assembling and validating data rather than interpreting it, which is both expensive and the opposite of why they were hired.

Fixing it is more organisational than technical. Someone has to own each metric definition and have the authority to make it binding. The platform work - a semantic layer, one modelled definition, lineage back to source - is comparatively straightforward once that ownership exists, and impossible without it.

This is why data ownership is established before modelling in our sequence. A technically excellent platform serving definitions nobody has agreed to simply relocates the argument, and the argument will resurface in the first board pack that contains two numbers for the same thing.

Lineage is the ability to take any figure in any report and trace it backwards through every transformation to the source records that produced it. It sounds like an audit nicety and it is actually the foundation of trust: the reason people believe a number is that someone once traced one and found it correct.

It also determines how fast you can respond when something goes wrong. A source system changes a field's meaning without telling anyone - which happens routinely - and downstream figures shift. With lineage, the impact analysis is a query. Without it, the impact analysis is a fortnight of archaeology conducted while executives wait.

Data quality rules belong at ingest, not at reporting. Validating on arrival means bad data is caught with the context needed to fix it, near the team that owns the source. Validating at the reporting layer means discovering problems weeks later, in aggregate, with no way to attribute them - and quietly correcting them in the report, which is how organisations end up with numbers that are right for reasons nobody has documented.

For AI specifically, lineage does additional work. When a model or an assistant produces an answer, being able to show the underlying records is what makes the answer usable in a regulated context. An unattributable output is a suggestion; an attributable one is evidence.

The core risk with enterprise generative AI is not the model inventing things, though it will. It is that a well-behaved assistant respecting permissions faithfully reveals exactly how permissive those permissions actually are. Ask it a question and it will helpfully surface the salary file that was over-shared in 2022, because it was entitled to.

Which means the prerequisite is the permissions and classification work, not the AI configuration. If sensitive content is not labelled and access is not scoped, deploying an assistant across it converts a latent exposure into an accessible one. We assess that before enabling anything, and it is frequently the reason we recommend delaying a rollout.

The controls that then matter are specific: enterprise deployment with tenant data boundaries so prompts and content do not leave your control or train shared models; access inherited from your identity model rather than granted to the agent as a whole; logging of prompts and outputs for both audit and quality review; and human review positioned wherever a decision carries real consequence.

Finally, scope it to decisions. General-purpose enthusiasm produces general-purpose disappointment. An assistant applied to a specific, high-frequency, well-bounded decision - with a measurable baseline for how that decision is made today - either demonstrates value quickly or fails quickly, and both outcomes are more useful than an open-ended pilot that runs for a year.