AI solutions
Models, agents and automation — built with a defined failure path, not just a working demo.
The challenge
Most AI proposals arrive as a demo that works on five examples and fails on the sixth, with no plan for the cases it gets wrong.
In an institution, the cases it gets wrong are the ones that matter.
How we approach it
Define the failure path first
What happens when it is wrong, before what happens when it is right.
Build custom only where it earns it
Otherwise integrate what already exists.
Keep a human route
Low confidence escalates to a person, always.
Measure on held-back data
Accuracy claims come from data the system has not seen.
The same four phases, whatever we are building.
- 012–3 weeks
Discovery
We map your current process, users and constraints, then define scope in writing.
- 022–4 weeks
Design
Interface and data design, reviewed with the people who will use the system daily.
- 036 weeks+
Build
Delivery in two-week increments, each one testable, with progress visible throughout.
- 04Ongoing
Launch & support
Deployment, training and documentation, then optional support at a level you choose.
What you get
- Held-back examples with agreed success criteria.
- The model or integration, with its prompts and rules versioned.
- Confidence thresholds and the human path they trigger.
- What it did, how often it escalated, where it drifted.
- What it costs to run at your actual volume.
What changes
Honest accuracy
Measured on data the system has not seen.
Safe failure
Uncertainty routes to a person rather than guessing.
Known cost
Running cost modelled before you commit.
Often looked at together.
Business & workflow automation
Automating the re-typing, chasing and monthly reporting that quietly consumes a team.
Data platforms & dashboards
Consolidated pipelines, agreed metric definitions, and views built around decisions.
Software development
Custom web and backend systems, built to be maintained by someone else.
About this work specifically.
General questions about scope, ownership and timelines are answered on the resources page.
Both, honestly. We build custom models and agent workflows where they earn their cost, and we integrate established providers where that is the sensible choice. We will tell you which one your problem needs.
Partly, and we will be specific about where the limits are. Handling of these languages is improving but uneven, particularly on scanned material — we test against your actual documents before promising a result.
Wherever you require. We can deploy so that documents never leave your infrastructure, which costs more to run and is sometimes the only acceptable answer. We will price both.
Tell us what you need built.
We reply within two working days with honest scope and next steps — including when we are not the right fit.