Training isn’t building a model
The models come from Anthropic and other providers, ready to read and reason. What they don’t know is your business: which supplier sends what, what counts as a change, when a job is really ready. Training an agent means teaching it that job, and it’s closer to onboarding a new hire than to a research project.
The five things it’s trained with
Rules. What to match, what to flag, when to stop and who decides. We write them down with the people who do the work, because the real rules are rarely the ones in the manual.
Examples. Real emails, documents and calls from your business, not made-up ones. They’re the test set. Every piece of the agent has to pass against them before it goes near live work.
Documents. Your templates, rate cards, checklists and procedures. The agent reads from them, so a scope comes from your rate card and a checklist from your firm, not from whatever the model remembers.
Guardrails. Spend limits, confidence thresholds, and the things it may never do: send under your name without approval, make a clinical, legal or financial call, or write to a live system when a run can’t be trusted.
Review loops. A person approves the output. Every correction goes back into the rules and the examples, and the agent runs beside the old way until its answers match your team’s. Only then does it take over.
What happens later
Once it’s live, the exceptions it meets change, and providers replace their models. Keeping the agent current, including fine-tuning a model where that earns its place, is maintenance rather than a new project. It’s covered by ongoing support or the fractional AI Automation Director retainer.
Related questions
- What can AI agents do for a small business?
- Read messy information, match it to the right job, draft replies and flag what's missing, with people approving anything consequential.
- Who manages an AI agent after it's built?
- You can, since you own it, or Loque Logic can through ongoing support or a fractional AI Automation Director.