Service
AI Agents
For work that varies every time. We scope the job, give the agent only the tools it needs, set what it may do without asking, and test it against the awkward cases rather than the tidy demo.
Problems it addresses
If any of this sounds familiar, it is worth a conversation.
The work varies every time, so the automation you were quoted for could never cover it.
Enquiries sit in a queue until somebody has a clear hour, and the clear hour keeps not arriving.
Answering a routine question means opening three systems, and only two people know where to look.
You have tried an AI tool and could not tell what it was allowed to do, so it quietly stopped being used.
What we deliver
A working outcome, not a slide deck.
Four artefacts from one build. The boundary document matters as much as the agent: it is the thing that lets you answer what it is allowed to do.
Scope and Boundary Document
The agent's job in one paragraph, then three lists: what it may do without asking, what needs a person to approve, and what it must never touch. Agreed before anything is built, because it decides everything that follows.
The Configured Agent
Built against your actual prices, policies and procedures rather than general knowledge, with access to exactly the systems the job needs and no more. Read-only wherever reading is enough.
Audit Trail and Monitoring
A record of what it did, what it decided, what it was unsure about and what it escalated, so you can answer why a customer got that reply. Plus alerting for the failure that matters: an agent looping, or quietly doing nothing.
Handover and Runbook
How to change what it knows, how to widen or narrow what it may do, and how to switch it off. Written for your team, because an agent nobody can adjust becomes an agent nobody trusts.
The engagement
What's involved
Here is what an agent build involves, activity by activity. Most of it is decisions and testing rather than construction.
| Activity | What's involved |
|---|---|
| Task selection | Finding the job worth starting with: varies enough to need an agent, frequent enough to matter, recoverable when wrong |
| Boundary definition | What it may do alone, what needs approval, what it must never touch, agreed in writing |
| Access and tooling | Connecting only the systems the job needs, read-only wherever reading is enough, with its own credentials |
| Knowledge preparation | Getting your prices, policies and procedures into a form the agent can use and you can update |
| Build and iteration | Assembling the agent, its tools and its instructions, then refining against real examples |
| Adversarial testing | The incomplete enquiry, the angry email, the two-questions-in-one, and input designed to redirect it |
| Logging and monitoring | Recording decisions and escalations, and alerting on looping or silent failure |
| Handover and runbook | Documentation your team can use to adjust what it knows, change its limits, or switch it off |
Ongoing operation, if you would rather not run it yourselves, is AI as a Service and is quoted separately.
Delivery process
How an engagement runs.
Pick the job
One narrow task with a clear finish line. We look for work that varies, happens often, and where a wrong answer is recoverable, which is the right place to start rather than the most impressive.
Set the boundary
Which systems it reads, which it can write to, what it may do alone and what waits for approval. This is agreed in writing with you and is the part that makes the rest safe.
Build and try to break it
Built against your real knowledge, then tested on the awkward cases: the incomplete enquiry, the angry email, the request that is really two requests, the input that tries to talk it into something else.
Run, log and tune
Live with a person reviewing what it does, tightening the boundary where it overstepped and widening it where it earned trust. Then handover, or we keep operating it for you.
Pricing approach: discovery is a fixed fee. Build work is quoted as a fixed scope and price once discovery is complete, so you decide with the numbers in front of you. Ongoing support is a monthly retainer you can stop at any time.
Questions
Frequently asked.
How is this different from AI Automation?+
An automation does what you specified, every time, the same way. An agent decides its own steps toward a goal. If your task is identical on every run, an automation is cheaper, more predictable and easier to check, and we will tell you so rather than sell you this. Agents are for work that varies so much that writing the path down was never realistic.
Can it do things without us?+
Only what you have agreed it can, and that list is written down before anything is built. Everything else routes to a person. In practice most agents we build start almost entirely in draft-and-approve mode, and the boundary widens only where it has earned it.
What stops it doing something expensive or stupid?+
Mostly what it cannot reach. An agent with read-only access to your systems and permission to draft has a worst case of a bad draft. The limits do the work, not the cleverness, which is why we spend the early part of the project on access rather than capability.
Could somebody talk it into misbehaving?+
If it reads anything from outside your business, yes, that is a real risk: a message can carry instructions as easily as information, and the agent cannot reliably tell them apart. It is handled by limiting what it is allowed to reach and requiring approval for anything with consequences, and it is one of the things our Security Audit looks at.
Whose AI models do you use, and where does our data go?+
Closed commercial models under business agreements that do not train on your data, running in your own accounts with your own keys. We stay model-neutral and pick what fits the job and your data rules. You own the configuration and the code; the models themselves are licensed from their providers, by you or by us, and we are specific about which.
What happens when it gets something wrong?+
You see it, because everything is logged, and then we tune: narrow the job, fix the knowledge it was working from, or move that case back to a person. Some cases should end up back with a person permanently, and saying so is part of the work rather than an admission.
What does it cost?+
Quoted on scope after we have looked at the actual task, because an agent handling one queue against one system is a different build to one working across four. Book a call and we will tell you what is involved, including when the answer is a simpler automation.
Further reading: What an AI agent actually is · 8 minute read, Where AI should, and should not, make decisions · 5 minute read
Related services
AI Automation
Project-built automations that take repetitive work off your team: designed once, handed over, yours to run.
Custom AI Tools
AI tools built specifically for how your business works, not a generic product you have to work around.
AI as a Service
We run the automations for you, every month, so you stop doing them entirely. Where our project services build a system and hand it over, AI as a Service is the ongoing operator that keeps the work off your team's plate, month after month.
Tell us the job, and we will tell you if it needs an agent.
We will tell you honestly whether this is the right answer for it.
Book a Strategy Call →