Pragmatic AI

Where AI can genuinely help today, and where it cannot yet. We look at the processes, data, skills and governance an AI deployment would depend on, and find the judgement line: the points in each process where a named human must own the output.

Readiness is as much about people as technology. Skills decay when people stop doing the work they are asked to check, so we look at that too.

An independent challenge to the business case before you commit. Is the efficiency claim tested or assumed, and what is it being compared against? Does it budget for verification, skills and process redesign?

We also model the scenario most business cases skip: the one where AI volume means you need more people, not fewer, at the points only people can handle.

Governance that works at the speed agents are built: guardrails embedded where creation happens, a gateway for policy, authorisation and audit, entitlements applied when information is retrieved, and escalation paths that give people the means to exercise their accountability.

Our reference architecture for consequential AI sets out the pattern.

The right model for the job, or no model at all. Deterministic software where it already does the job, smaller and local models where they are good enough, and cost and energy metered alongside latency, so that AI absorbs complexity rather than generating more of it.