AI agent development for a business process
One process, taken end to end by an agent that knows when to stop and hand the work to a person.
Most AI agent projects fail in the same place. The demo answers ten questions beautifully, then meets a real customer who asks something slightly off, and nobody knows what the agent did or why. The model is rarely the problem. The problem is that nobody defined where the agent's authority ends.
I build agents the other way around. First we pick one process that costs you real money today: inbound requests, support triage, document intake, lead qualification. Then we write down what the agent may decide alone, what it must escalate, and how we will know week to week whether it is getting better or worse.
Who this is for
- A process that repeats dozens or hundreds of times a week and is handled by people reading and re-typing.
- A team of roughly 5 to 200 people, where one person owns the process and can answer questions about it.
- You can show me real examples: past requests, real documents, actual replies your team sent.
If the process runs twice a month, or nobody can say what a correct outcome looks like, an agent is the wrong tool and I will say so before you spend anything.
What you get
- A working agent on your data, deployed where your team already works: web, Telegram, email or your own product.
- Guardrails: what it may answer, what it must refuse, and the exact point where a human takes over.
- An evaluation set built from your real cases, so a change can be measured instead of argued about.
- Operator handoff with full context, so the person who picks it up does not start from zero.
- Logs and a simple dashboard: volume, escalation rate, where it fails.
- A handover document and a walkthrough, so your team can run it without me.
How it runs
One call plus your real examples. We define the process boundary, the success condition and what must never happen.
The agent handles the common path only, on your data, with escalation for everything else. Usually ready in about two weeks.
Shadow mode or a small share of live volume. We watch the evaluation set, not opinions, and fix what actually breaks.
Expand coverage where the numbers justify it, write the handover, train your operator.
Questions people ask first
- How is an AI agent different from automation I already have?
- Automation follows a fixed path you wrote in advance. An agent decides which path to take, which makes it useful for messy input like free-text requests or documents that never look the same twice. If your input is already structured and the rules are stable, plain automation is cheaper and more reliable, and I will recommend that instead.
- What does it cost to run every month?
- Model usage for a typical support or intake agent at moderate volume usually lands between 30 and 200 EUR a month, plus hosting. I size this during scoping so the running cost is a number you see before we build, not a surprise afterwards.
- Do you need our data to leave the company?
- Not necessarily. We can run against an API with no training on your data, keep retrieval inside your own infrastructure, or use a local model where the data cannot leave at all. The choice affects quality and cost, and we make it explicitly at the start.
- What happens when the agent gets something wrong?
- It escalates rather than improvises. Every agent I build has an explicit refusal path and a handoff that carries the full context to a person. Errors that slip through land in the evaluation set, which is how the next version stops repeating them.
- Who owns the code?
- You do. The code, the prompts, the evaluation set and the documentation are yours at handover, and they run without me.
Related reading
All services
Retrieval over your own knowledge, wired into the product you already ship, with answers that cite where they came from.
ReadA week of work that tells you which three things to do with AI, what each costs, and what to leave alone.
Read