AI agents
Assistants that read your real data before they answer, with the unglamorous guardrails that keep them useful.
An AI agent is only worth having if it is grounded in something true. A chatbot that invents an answer about your pricing costs you more than it saves. So every agent I build starts from your data: your documentation, your database, your ticket history.
The interesting part is not the model, it is what happens around it. Retrieval that finds the right context. A confidence threshold that decides when to answer and when to hand the conversation to a person. An audit log so you can see why it did what it did.
What I build
- Support agents that answer from your own documentation and escalate what they cannot handle.
- Classification and routing: incoming email or tickets sorted by category, priority and language, then sent to the right queue.
- Internal assistants that query your database and return a real number instead of a plausible one.
- Document processing that extracts structured fields from invoices, contracts or forms and writes them to your system.
- Discord and Slack bots that live where the team already works.
The guardrails
- Structured output, so the next system gets fields and not prose.
- Confidence thresholds, so low certainty triggers a human handoff instead of a guess.
- Deterministic rules for what a model must never decide alone: urgency, escalation, business hours.
- An audit trail with the input, the decision and the reason behind it.
- Synthetic data in every public demo. Your data stays in your systems.
Questions
about this
Which model do you use?
Whatever fits the job and the budget, usually OpenAI or Claude. If the data cannot leave your infrastructure, a self-hosted open model is an option, with an honest conversation about the quality difference first.
How do you stop it from making things up?
By not asking it to know things. The agent retrieves the relevant source, answers from that, and says it does not know when the source has no answer. Low confidence routes to a person.
What does it cost to run?
Token cost depends on volume and model. I estimate it before we build and design the flow so the expensive model only runs on the steps that need it.
Can it run inside n8n?
Usually it does. The workflow handles intake, retrieval, tool calls and routing, and the model handles the part that needs language.
Is my data used for training?
Not by me, and the business API tiers I use do not train on your traffic by default. I confirm the exact terms and the processing region for the provider we pick before anything goes live.
Want this for your business?