Data & intelligence

AI and generative AI

Use-case assessment, retrieval systems, evaluation and human review.

Capability in business terms

Why this capability matters

Most AI projects fail before a model is chosen, because nobody agreed what a correct answer looks like or what the system is allowed to see.

What we typically scope

Use-case assessment, retrieval systems, evaluation harnesses, agentic workflows, human review design, governance.

How we approach the work

Define the decision first, the evaluation second, the model last. Every accepted output is traceable to a person who is accountable for it.

Work we take on

Example engagement — illustrative. Typical scopes of work, not named client case studies.

The assistant that confidently invents policy

Why it happens
An internal support assistant answers HR and policy questions from a document set, and occasionally states a policy that does not exist.
How we work it
Build a retrieval evaluation set from real historical questions, measure grounding and citation accuracy, add answer refusal thresholds, surface source passages inline, log every low-confidence response for review.
What you end up with
A measured hallucination rate with a baseline, a refusal path instead of a fabrication, and a weekly review queue with an owner.

Choosing between three vendors and a build

Why it happens
Leadership has three AI vendor proposals and an internal team claiming it can build the same thing.
How we work it
Define the use case's acceptance criteria first, run the same evaluation set against each option, cost the total including retraining and evaluation maintenance, assess lock-in and data-exit terms.
What you end up with
A scored comparison against your own criteria, a written recommendation with the trade-offs stated, and the evaluation set retained as a durable asset.

An agent that needs permission to act

Why it happens
A workflow agent is being asked to take real actions — raise tickets, update records, send messages — and nobody has defined its limits.
How we work it
Classify actions by reversibility, gate irreversible actions behind human approval, scope credentials per tool, add structured error handling and an audit trail of every tool call.
What you end up with
An agent with a written permission model, a human-in-the-loop path for irreversible actions, and a log an auditor can read.

Governance before the regulator asks

Why it happens
AI use has spread across teams without a register, and a customer's security questionnaire has just asked which models process their data.
How we work it
Inventory every AI system and its data flows, classify by risk, define permitted models and data boundaries, write the AI use policy, establish model and vendor review.
What you end up with
A maintained AI register, a published use policy, and a defensible answer to the questionnaire.

Document extraction that has to be right

Why it happens
Invoices, contracts or claims arrive as PDFs and are keyed in by hand, with a backlog that grows faster than headcount.
How we work it
Build a labelled ground-truth set, measure field-level accuracy, set confidence thresholds per field, design the exception queue, instrument drift detection.
What you end up with
A per-field accuracy baseline, a defined automation threshold, and an exception path that stays manageable as volume grows.

Related business solutions

Ways to engage

How we approach the work

Understand the estate

Map dependencies, constraints, decision owners and the current operating model.

Make the decision explicit

Record options, risks, costs and acceptance criteria before implementation.

Make the handover usable

Document changes, operating instructions and the remaining risks.

Delivery is subject to specialist fit and availability.

Start with clarity.

What would make your business easier to run?

Tell us the operational challenge. We will discuss whether a short conversation, a scoped assessment or a delivery engagement is the right next step.