Technology partner
bilz.ai
An AI back office for European restaurants that tracks food cost, supplier invoices, recipes, stock and profit. We worked with it as technology partner.
Service
AI agents and automation
AI agents that read documents, answer customers and assemble reports inside your existing tools, with an evaluation set, an action log and a person in the loop.
The problem
Most teams have a few jobs too fiddly to script and too repetitive for a person: reading documents into a system, triaging messages, assembling reports. They sit between tools, so no vendor has automated them.
Language models are good at this work. The hard part is connecting the model to your data safely, measuring whether it is right and keeping it right after the first month.
What we deliver
A working agent on one workflow, connected to your email, CRM, accounting, POS or database.
An evaluation set from your own past cases, so accuracy is a number.
Scoped tools, approval steps for irreversible actions and a full action log.
Monitoring and a runbook.
Source code and documentation in your repository, with no lock-in.
How we work
- Pick the workflow. We map what a person does today and what a correct result looks like.
- Build the evaluation set. Twenty to two hundred past cases with known answers.
- Run in shadow mode. The agent works beside the person for a week or two. Permissions widen as the numbers allow.
- Hand over or keep running. You get the code, the evaluation set and the runbook, or we operate the agent for a monthly fee.
Where this comes from
We have built agents of this kind for the products we worked with. bilz.ai, which we worked with as technology partner, reads restaurant supplier invoices without anyone typing. CardioMood, which we worked with, turns wearable data into alerts a coach can act on.
Questions
What does an AI agent do for a small business?
It takes a job a person does by hand across several tools and does it the same way every time, with a record of each action. Examples are reading supplier invoices into an inventory system, answering routine support email from your documentation and drafting the weekly report from the CRM. The agent handles the normal cases and passes the unusual ones to a person.
Which models do you use?
Whichever scores best on your data. We have shipped on Claude, GPT and open-weight models, and we choose by accuracy on your cases, cost per task and latency. The code is written so the model can be swapped later.
How do you keep an agent from doing something wrong?
Narrow tools, explicit permissions and a human approval step for anything irreversible. Every action is logged with its inputs and the model's reasoning. An agent starts in read-only or draft-only mode and gets wider permissions as the error rate on recent cases allows.
How long does it take?
A first working agent on one workflow takes two to four weeks, including the time to connect your systems and run it in shadow mode against real work. Widening it to more cases comes after that, in short increments.
What happens after launch?
Agents degrade when the business or the model changes. We leave you with the evaluation set, monitoring and a short runbook. We can also keep running the agent for a monthly fee.
Where we have done this
Technology partner
CardioMood
A physiological early-warning system for executive coaching with continuous medical-grade stress, recovery and performance data. We built the platform and apps.