Every vendor in the go-to-market stack has shipped something described as an agent this year, and most of it is workflow automation with new branding. Underneath the language, though, a small number of genuinely useful patterns have emerged — and they share a property that the marketing collateral tends to obscure.
The property that separates useful from theatrical
The useful ones operate on your data, take an action a human would otherwise take, and produce output a human reviews before it reaches a customer. The theatrical ones generate text. That's the whole distinction. An 'agent' that drafts an email is a text box. An agent that reads every support ticket from last week, clusters the recurring issues, and drafts a note to product with links to the evidence is doing a job somebody wasn't doing.
What's actually working in client accounts
Three patterns have earned their place. Research and enrichment ahead of a human-sent outreach, where the agent assembles context and a person decides what to do with it. Continuous monitoring of unstructured material — calls, tickets, reviews — with alerts when something changes. And internal question-answering over your own documentation, which quietly removes a large amount of interruption from a small team.
- Pre-call research assembled automatically, reviewed by the person making the call
- Monitoring of calls and tickets for emerging themes, with the source quotes attached
- Internal knowledge retrieval — 'what did we agree about pricing for education?'
- Routine reporting assembly, with a human interpreting rather than compiling
Keep a human on anything customer-facing
The failure mode is predictable and expensive: an agent sends something wrong to a real customer at scale, and by the time you notice, it's in several hundred inboxes. The trust cost far exceeds the time saved. My rule in client engagements is straightforward — an agent can prepare anything and send nothing. That constraint removes most of the risk and almost none of the value, because the preparation was always the slow part.
Evaluate on the task, not the demo
Vendor demos are constructed to work. Take one recurring task your team genuinely dislikes, run it manually for a fortnight and record the time, then run it with the tool for a fortnight and record it again. If the saving isn't obvious without measuring carefully, there isn't one. This test is unfashionable and it will disqualify most of what you're being pitched.
The pattern that works is unglamorous: agents that read a lot, prepare well, and never speak to a customer. Almost everything being sold as transformative is the opposite arrangement.
Want to apply this to your business?
Book a free 30-minute strategy call and we'll diagnose your specific growth blockers together.
Book a Free Call