01 · Automations & AI agents
Getting an AI to do something once is easy. Getting it to do the right thing every day, at a predictable cost, and to tell you when it isn't sure, is the actual job. I design and build automations and AI agents for teams that need them to be dependable, not impressive for a week.
Sound familiar?
Someone on your team spends hours a week moving information between tools by hand.
Requests and issues arrive in chat and get lost, or nobody knows who owns them.
You've tried an AI tool, and it worked in the demo but not on real data.
An automation you already have breaks quietly, or costs more than anyone expected.
What you get
We walk through the process as it really works, not as the diagram says. Most savings come from fixing the process; automation locks it in.
Anything predictable is plain code: free, fast and testable. The AI is only used where judgement is needed, and only woken up when there's something new to judge. That's how one agent's running cost dropped from £490 to £39 a month.
Outages, missed messages and API limits are designed for from day one. If something goes wrong, the system catches up on its own and never quietly drops work.
Plain-English specs, a runbook, and a handover session. Every rule is written down with the reason it exists, so your team isn't dependent on me.
Proof
Problems reported in chat were getting lost, and sorting each one meant reading 50-message threads. I built an agent that turns every report into a tracked ticket, routes it to the right owner and keeps the person who raised it updated. The team runs it without me.
Weeks of empty results turned out to be two bugs hiding each other. After the fix, the same job finished in two minutes and produced a real, sourced business proposal, with 340 automated tests guarding it.
Questions
Slack, Jira, Asana, CRMs, spreadsheets, data warehouses, and most things with an API. If it can't be reached properly, I'll say so before we start rather than after.
It depends on the data and the model provider, and it's decided up front. Customer data stays out of prompts by default, and anything that must go in is masked or covered by the right agreements.
You get a running-cost estimate before the build and a measured figure after. The design aim is that cost grows much more slowly than volume.
Contact
Thirty minutes to talk through what you’re trying to fix. If I’m not the right fit, I’ll say so, and point you to someone who is.