An agent differs from a chatbot in one respect that decides the whole project: it doesn’t just answer, it acts. It reads a record, makes a decision and changes something. That makes the interesting question not which model to use, but what it is allowed to do without asking — and how you’d know afterwards what it did.
The three shapes this work takes
- Support. Incoming tickets triaged, classified and resolved where they’re routine — checking the account, looking up the order, issuing the standard fix — and escalated with context attached where they aren’t. The gain is in the eighty percent that were always going to get the same answer.
- Commerce. Product discovery, order status, returns and reorders handled inside the store against live catalogue and inventory data. This is where an online store stops being a catalogue and starts answering questions at 11pm.
- Back office. Form routing, document classification, approval chains, and moving data between two systems nobody ever intended to integrate. Unglamorous, and usually the highest return of the three because the work is pure repetition.
Why the data layer comes first
An agent is only as good as what it can read. A model given stale, contradictory or incomplete data will make confident wrong decisions faster than a person would — and it will make them at volume.
So the warehouse is usually the first phase, not an afterthought: ingestion from the systems you already run, a schema that survives new sources being added, and enough history to query rather than reconstruct. That layer pays for itself in reporting even if the agent is never built, which is a reasonable way to de-risk the whole programme.
What we insist on before anything runs unsupervised
The failure mode here isn’t a model being wrong. It’s a model being wrong in a way nobody notices for six weeks.
- A documented boundary. What the agent may do alone, what needs a human signature. Anything touching money, a citizen record or an irreversible action keeps an approval step unless you explicitly decide otherwise.
- An audit trail with reasoning. Every action logged with why it was taken. “The system did it” is not something a department can put in front of an auditor.
- Evaluation against real history. Behaviour measured on actual past cases, including the ones that went wrong, before it touches a live queue.
- A rollback. A defined way to switch it off and return to the manual process without losing the work in flight.
Where an agent turns out to be the wrong answer — too little volume, too many exceptions, a process that changes every quarter — we’ll say so during scoping. That’s the same test we apply to everything else on this service.
What's included
How we scope this
Agentic support systems
Triage, classification and resolution of routine tickets end to end — reading the account, checking the order, taking the action — with a defined point where a person takes over.
Agentic commerce
Agents that handle product discovery, order status, returns and reorders inside a store, working against real catalogue and inventory data rather than a scripted FAQ.
Workflow automation
Multi-step back-office processes — form routing, document classification, approval chains, moving data between systems that were never designed to talk.
Data warehousing
Ingestion from the systems you already run into a warehouse your reporting and your agents both read from, with a schema that survives new sources being added.
Approval boundaries and audit trails
What an agent may do alone, what needs a human signature, and a log of every action with the reasoning behind it — because "the system did it" is not an answer anyone can act on.
Evaluation before production
Agent behaviour measured against real historical cases, including the ones that went wrong, before it touches a live queue.
How we work
The process, in order
- 01
Map the process
The actual steps, the exceptions and who currently signs off — written down before any of it is automated, because most processes are not documented anywhere accurate.
- 02
Build the data layer
The agent needs to read reliable, current data to be worth anything. Where that doesn't exist yet, the warehouse comes first and the agent second.
- 03
Pilot on a narrow slice
One queue, one process, with a person reviewing every action. Measured against how the process performs today, not against a demo.
- 04
Widen with guardrails
Autonomy extended only where the pilot earned it, with monitoring, escalation and a documented rollback.
FAQ