Service
Custom AI agent development services that do real work
A demo that answers questions about a PDF takes an afternoon. Something your team trusts with a customer, a ledger, or a contract takes retrieval that actually retrieves, permissions that actually restrict, and an evaluation set that proves it behaves. That gap is the work.
What we build
Three shapes that cover most of the demand
Almost every request we get is one of these, or a combination. Naming the shape early settles scope, cost, and risk in one conversation. If your organisation already lives in Microsoft 365, the same three shapes are buildable on the platform you own — see hiring a Microsoft Copilot Studio consultant for what that changes about grounding, channels, and licensing. If what you need is Claude rolled out to a team rather than a bespoke build, see Claude AI implementation consulting.
Grounded knowledge assistants
Ask a question in plain language, get an answer drawn from your policies, contracts, tickets, or manuals — with citations back to the source document.
Tool-using agents
Assistants with a defined set of actions: create the record, draft the reply, raise the ticket, update the status — each scoped and logged.
Document processing
Structured data pulled from unstructured input — invoices, forms, statements, contracts — validated before anything downstream trusts it.
The second shape needs somewhere for those actions to land. Rather than wiring each one into a specific assistant, we usually put them behind a protocol server the agent calls — the connector layer covered in MCP server development, which handles authentication, per-tool permissions and the audit log, and survives a change of model or vendor untouched.
One channel changes the engineering more than it changes the agent: the telephone. Latency the caller can hear, interruption part-way through a sentence, and details that have to be spelled back are all phone problems rather than model problems, and they are what AI voice agent development services deal with — answering the line, booking by phone, and writing the call into the CRM.
Method
Why most internal assistants disappoint
The model is rarely the problem. Projects fail on retrieval quality, undefined success criteria, and permissions nobody thought through until security asked.
| Failure mode | What it looks like | How we address it |
|---|---|---|
| Weak retrieval | Confident answers assembled from the wrong document | Chunking and indexing tuned against a real question set, with citations shown to the user |
| No evaluation | Quality judged by whoever last tried it | A written test set of real questions with expected answers, re-run on every change |
| Over-broad permissions | One service account with access to everything | Per-action scopes, least privilege, and human approval on consequential steps |
| Stale knowledge | Answers from a document superseded months ago | Scheduled re-indexing tied to the source systems, with freshness visible in the answer |
| No fallback | The assistant guesses rather than admitting it does not know | Explicit low-confidence handling that routes to a person and records the gap |
On governance: for clients whose security review needs a recognised reference, we map controls to the NIST AI Risk Management Framework. It gives both sides a shared vocabulary for what the system does with data, where a human stays in the loop, and how failures are detected.
Questions
What clients ask before starting a build
What is the difference between a chatbot and an agent?
A chatbot answers questions. An agent takes actions — it can look something up, decide what to do next, and then write to a system, send a message, or open a ticket. The distinction matters commercially: answering saves someone a search, while acting removes a task from the queue entirely. Most projects should start as the former and earn the right to become the latter.
How do you stop an AI assistant inventing answers?
By never asking the model to recall facts from memory. Answers are grounded in retrieved passages from your own documents, and the assistant cites the source it used so anyone can verify it. When retrieval finds nothing relevant, the correct behaviour is to say so and hand off to a person, and we test for that explicitly rather than assuming it.
What controls exist before an agent can change something?
Each action an agent can take is defined individually, scoped to the minimum permission it needs, and logged. Anything consequential — sending external email, issuing credit, changing a record of account — sits behind explicit human approval by default. The agent proposes, a person confirms, and the whole exchange is recorded.
Which model provider do you build on?
Whichever suits the task, and the integration is written so the provider can be changed later. Model quality and pricing move quarterly, so being locked to one vendor is a liability. What stays stable is the retrieval layer, the tool definitions, and the evaluation set — those are the durable assets, not the model behind them.
Free consultation
Tell us what is costing you the most time
Describe the process. You get a straight read on whether it is worth automating, roughly what it would take, and what we would tackle first — on a call that costs nothing.
- Under an hour, free, and no obligation
- Fixed price per stage, agreed before any work starts
- You own everything we build, documented and in your accounts
- If you do not need us, we will say so on that call
The form is the fastest way to reach us — we reply from a real person, usually within one business day.