AI development & agents
We build agents that perform tasks, not demos that impress. They run in your own Microsoft tenant, against rules you own and can change without anyone writing code.
What we do
Mapping the task. We look at what is done by hand today, how often, and what happens when it goes wrong.
Rules the customer owns. Customer profile, exclusions and score weights live in configuration, not in code.
Runs with an audit log. Every run is logged, so you can see what the agent did and why.
Escalation when in doubt. The agent hands a case to a person when the evidence is thin. That is designed in, not an afterthought.
How we work
Half an hour to understand the task and the data behind it.
We show the concept holds before you open your own tenant.
UAT, then production in your tenant, with CI/CD and operations.
What a delivery contains
The agent that performs the task, not a demo showing that it could
The rules it works to, in configuration you can change yourselves
Access that inherits the permissions of the user it acts for
Connections to the sources it reads and the system it writes to
A human approval step wherever the decision has consequences
A full log per run, so it is possible to see why something happened
Operation and monitoring once it is in production
What it costs
An agent costs in two ways, and they behave differently. Building it is driven by how many sources it must read, how messy those sources are, and how much judgement the task demands. Running it is driven by how often it runs and how much it must read each time, a recurring cost, and a predictable one once measured. We set a ceiling per run before anything goes into production, so the bill cannot surprise you.
We start with one task and measure what it actually costs to run before widening it. Here is how that went at a maritime customer, or ask for an estimate.
How long it takes
| Scope | Time |
|---|---|
| First working version to look at | days |
| One task, one source, in production | 3-6 weeks |
| With integrations and an approval step | 6-10 weeks |
| Several agents under shared governance | 3 months and up |
What takes time is rarely the model. It is agreeing what the agent may do without asking, and getting the sources it reads into good enough shape that the answers can be trusted.
When this fits, and when it does not
Fits when
- The task is done today, by hand, and repeats often
- There is a source to read from and a system to write to
- It is possible to say what a good answer looks like
- Somebody internally can review the output at first
Does not fit when
- The task happens once a quarter, the arithmetic does not work
- Nobody can judge whether the answer is right
- A wrong answer is costly and cannot be caught by an approval step
- The data it would read does not exist in one place yet
Questions we get
Does our data leave the environment?
No. The agent runs in your own Microsoft tenant, on your data, and the data is not used to train models. Access follows the roles you already have in Entra ID.
What if it gets something wrong?
Where a decision has consequences we build in a human approval step rather than letting the model decide alone. Everything is logged per run, so it is possible to see after the fact why something happened.
Is this a pilot, or something that actually runs?
An agent we built does real work at a maritime customer every day, in the customer's own environment. We build for production, not for a demo.
Who owns the rules the agent follows?
You do. The rules sit in configuration rather than code, so you can change them without ordering a change from us.
Shall we talk this through?
Tell us briefly what you need and you get a concrete answer, and an estimate you can plan around.