What AI agents actually do in a commercial team
Stripping the hype from agent architecture: the four roles that produce measurable commercial value, and the boundaries that keep them safe.
Daniel Krishnan
Director of AI Systems
A working definition
An agent, commercially useful, is a bounded process that can read context, take a limited set of actions and escalate what falls outside its scope. That definition is deliberately unexciting. It is also the version that survives contact with a real business.
The four roles that pay for themselves
Across our deployments, four agent roles account for most realised value.
- Responder: answers within seconds, at any hour, within approved boundaries.
- Qualifier: scores fit and intent, then routes to a named owner.
- Producer: drafts content, quotations or summaries from structured knowledge.
- Analyst: reconciles data across systems and reports exceptions.
Boundaries are the product
The engineering that matters is not the model choice. It is the specification of what the agent may do unsupervised, what it must escalate, and what evidence it must log. Businesses that treat this as documentation rather than design get incidents.
Measuring agents honestly
Track exception rate and time-to-escalation alongside volume. An agent handling 90% of enquiries with a 15% error rate is worse than one handling 60% with a 1% error rate, and only the second builds organisational trust.
Keep reading
Related articles.
From Brand to Benchmark
Ready to Build an Intelligent Growth System?
Stop buying disconnected services. Start building a scalable AI-powered business with one team accountable for brand, systems and revenue.
30-minute working session. You leave with a growth system map, whether we work together or not.