Consulting
AI agents that survive past the demo
Copilot Studio, Power Automate and Azure AI Foundry, inside your own tenant. Constrained to your documents, tested on your real content, and documented so your team can change it without us.
What we build
Four kinds of work, and they are genuinely different problems.
An agent that reads your real documents, follows a real process, and produces output somebody can act on. Scoped to a defined set of content, with topics and escalation paths that make sense to the people using it.
Multi-pass extraction that turns messy source material into structured, checkable output. The pattern behind compliance-ready documents that run on a schedule rather than when somebody remembers.
When Copilot Studio is not the right container: retrieval over your own content, custom orchestration, and models called from your own application rather than a chat window.
Where the interface is the product, on your site or inside Teams. Built so the answer set is controlled, because an agent that invents an answer to a customer is worse than no agent.
Most of the value is not the model
This is the part that decides whether an agent is still running in six months.
The unglamorous half
Work out which part of a messy real process genuinely needs a model and which part just needs good engineering, then build both. Most of the value is in the second part.
Constrained, not creative
Agents are bounded to your content and made to say they do not know rather than guess. A confident wrong answer to a customer costs more than no answer.
Tested on your content
One that works on sample data and fails on yours is not finished. Testing happens against the real thing before handover.
Your tenant, your data
Built inside your Microsoft tenant under your existing agreements. We do not host your data and do not need a copy of it.
Running costs, estimated up front
Consumption is yours and paid to Microsoft directly. You get an estimate before committing, not a surprise on the first invoice.
An honest no
A lot of what arrives as an AI project is a process problem or an ordinary integration. Both are cheaper and more reliable without a model. We say which one you have.
Constrained, not creative.
A confident wrong answer to your customer costs more than no answer at all.
How a project runs
The problem, not the technology
What the process is today, who does it, what it costs in time, and what "good" would look like. This is also where we work out whether you need a model at all.
Scope, fixed figure, estimated running cost
A written scope with a price and an estimate of Microsoft consumption. If the honest answer is that this should not be built, that comes now.
Build and test against real content
Constrained to your documents, with the failure cases tested deliberately, including what it does when it does not know.
Handover your team can maintain
The agent, the topics, the source content and the documentation to change any of it. Not a black box.
Questions about AI work
How is this priced?
Per piece of work, against a written scope, with a fixed figure before anything starts. Not a monthly retainer and not hourly. Running costs after handover are yours, paid to Microsoft directly, and we will estimate them before you commit.
Will the agent make things up?
That is the whole engineering problem, and it is why most agents die after the demo. An agent is constrained to your documents, made to say it does not know rather than guess, and tested against real content before handover. If a use case cannot be made safe, we will tell you rather than ship it.
What does it run on?
Microsoft Copilot Studio, Power Automate and Azure AI Foundry, inside your own Microsoft tenant. Your data stays in your tenant under your existing agreements. We do not host your data and we do not need a copy of it.
Do we need an AI project at all?
Often not. A large share of what people bring to an AI conversation is a process problem or an ordinary integration, and both are cheaper and more reliable to solve with plain engineering. We will say which one you have before quoting.
What happens after handover?
You get the agent, the topics, the source documents it uses and documentation on how to change them. Your team can edit it. Support is available and never assumed.
How long does it take?
A focused agent over a defined document set is typically two to six weeks, most of which is content, access and testing rather than building. The estimate comes with the quote.
Bring a process, not a technology
Tell us what actually happens today and we will tell you whether this needs an agent, an integration, or neither.