Services
AI assistants and agents
Most "AI projects" fail for a boring reason: nobody defined what the model is allowed to do, or what happens when it is wrong.
There is a large gap between a demo that answers a question impressively and a system your business can rely on daily. The gap is not the model. It is the surrounding engineering: what data it can see, what it is allowed to do, when a human must check the output, and how you find out when it starts getting things wrong.
We build assistants that sit inside a defined workflow with a narrow job — extract these fields, classify this enquiry, draft this reply from these sources — and hand off to a person at the point where judgement is required. Narrow scope is what makes them reliable enough to trust.
Where it does not fit: if the task needs judgement across incomplete information, or a wrong answer would be expensive and hard to detect, keep a person in the loop for all of it. If nobody can define what a correct answer looks like, there is nothing to measure, and a system nobody can evaluate should not be making decisions.
Signs this is worth fixing
- Staff spend hours reading documents and typing out the same fields
- The same customer questions are answered from scratch every time
- Knowledge lives in one person’s head or a shared drive nobody searches
- Enquiries arrive after hours and wait until morning
- An AI trial impressed everyone and then went nowhere
- Nobody can say what the AI costs per month or when it is wrong
What we automate
Document extraction
Pulling structured fields out of invoices, contracts, applications and forms, with a confidence score and a review queue for anything uncertain.
Classification and routing
Sorting incoming enquiries, tickets or documents by type, urgency or department so they reach the right person already labelled.
Drafting from your own sources
Replies and documents drafted using your actual policies, pricing and procedures — retrieved from your documents rather than invented by the model.
Internal knowledge search
A question in plain language returns the relevant passage from your own material, with the source shown so it can be checked.
Human-in-the-loop by design
Anything above an agreed risk level waits for a person to approve. The assistant proposes; your team disposes.
How it works
- 01
Pick a narrow, measurable job
We choose one task with a clear definition of correct, so success can be measured rather than argued about.
- 02
Assemble and clean the sources
The assistant is only as good as the material it reads. We gather the real documents and remove the outdated versions that would poison the answers.
- 03
Build with evaluation, not vibes
We build a test set of real cases with known correct answers, and measure every change against it before it goes near a customer.
- 04
Deploy with guardrails and cost limits
Scoped permissions, review thresholds, logging, and a monthly cost ceiling. We hand over the evaluation set so you can keep testing.
What you get
- A defined task scope and definition of "correct"
- The assistant, connected to your systems and sources
- An evaluation set of real cases with expected answers
- Human review thresholds and escalation rules
- Cost monitoring and a monthly ceiling
- Documentation of prompts, sources and limits
Systems we work with
We build on mainstream model providers and keep the integration in your existing stack. Where data may not leave your environment, we discuss deployment options and the trade-offs honestly before committing to an approach.
Questions people ask
Will it make things up?
Any language model can. That is why we constrain it to retrieved sources, require citations where possible, score confidence, and route uncertain answers to a person. The goal is a system that fails safely and visibly, not one that never fails.
What does it cost to run?
Model usage is metered. For a narrow task the monthly cost is usually modest, but it depends on volume and document size. We set a ceiling and monitor actual spend so there are no surprises.
Is our data safe?
We scope access to only what the task needs, avoid sending more than necessary, and agree which provider and region is acceptable. Where data cannot leave your environment, we will say so rather than quietly working around it.
Can it talk to customers directly?
It can, with limits. Most clients start with the assistant drafting and a person sending. Fully automated responses work for narrow, low-risk questions; anything involving money, commitments or complaints goes to a human.