The team says the model needs more training. Before you go there, look at what the agent is actually working with.
By default, an agent draws on everything it was ever trained on. No bounds usually means no focus. You’ve handed it the entire massive messy world and asked it to do one infinitesimally small specific job.
Think about a new hire who gets no onboarding, just “Handle customer issues.” They’ll do something. Whether it’s the right thing, done well, is mostly luck. Give the same person the background, the playbook, the goals and the guardrails, and you get a completely different, much better outcome.
Agents work the same way. The more clearly you define what the agent should think about (business goals, allowed actions, relevant knowledge, decision principles), the fewer possible responses it has to choose from. Fewer possibilities means fewer wrong answers.
The inputs that matter: work roles and goals, business documentation, process guidance, brand standards, KPIs. Most organizations already have these. They just haven’t put them deliberately in front of the agent.
Context is often the fix: information, specificity, clarity and constraints. How to use it well is the start of success. Feedback on usage powers the flywheel of wisdom.
How to start:
- Write down the agent’s job the way you’d explain it to a new hire: the role, the goal and what success looks like.
- Gather what you already have: roles and goals, business documentation, process guidance, brand standards, KPIs.
- Put what’s relevant to that job in front of the agent, and leave out what isn’t.
- Spell out its allowed actions and your escalation rules.
- Then test it against that definition, in whole conversations.
- Define success and measure against it for the job you assign.
- Collect feedback on how it’s used, and feed it back into the agent’s context. Keep the flywheel turning.
Is your agent working from a path to success, or from a world of noise and confusion?