The word "agent" has been melted into a marketing puddle. Vendors call any feature with an LLM in it agentic. Buyers, sensibly, have started to roll their eyes. So before we tell you what Polluxa CRM does, here is the test we use internally — and the test we recommend you use when anyone, including us, claims to have shipped an agent.
The three-question test
For any feature being sold as "agentic," ask three questions. Does it observe a real-world signal without being asked? Does it take an action against a system without a human pushing a button? And does it own a measurable outcome you would otherwise have paid a human to produce?
If the answer to any of those is no, it's a chatbot, an autocomplete or a workflow. Those are fine things. They are not agents.
Every Polluxa agent — the ones we ship and the ones you build — passes all three. Each watches signals continuously. Each takes scoped actions on the platform without a button-press. Each is held accountable to a specific outcome — meetings booked, replies received, deals advanced, support tickets closed, contacts enriched, or whatever outcome your workflow defines.
Why twelve and not one
The dominant industry approach is to build one large, general-purpose assistant that can supposedly do everything. We tried this. It is a bad idea.
Generalist agents fail in three predictable ways. They make confident decisions in domains where they shouldn't have any confidence. They are impossible to evaluate, because their behaviour is too broad to test. And when they break, they break in a way that's hard to debug, because the same model is responsible for everything.
Specialist agents are different. Each one has a narrow scope, a clear KPI, and a deterministic test harness. The SDR Agent only does one thing: find the right person, write the right opener, get the meeting booked. That is enough surface area to be useful. It is small enough surface area to be evaluable.
One generalist agent is a marketing demo. A growing roster of specialists, each accountable to a KPI you defined, is a workforce.
The starter set — and the builder
This is the part most "agentic CRM" pitches get wrong. Vendors ship a fixed roster and tell you it's enough. It isn't. Every business has workflows that aren't on anyone else's roster — a unique re-engagement loop, a particular kind of forecast revision, a niche compliance check — and a closed roster has nothing to say about them.
Polluxa CRM ships a useful starter set of agents on day one. You'll recognize most of them by the role they're named for:
- SDR Agent — researches, qualifies, sequences and books meetings.
- Enrichment Agent — keeps every contact and company record live.
- Outbound Agent — writes and sends across email, LinkedIn, WhatsApp.
- Inbound Agent — qualifies form fills, ads clicks, demo requests within seconds.
- Meeting Agent — handles scheduling, reschedules, no-show recovery.
- Pipeline Agent — flags stalled deals, surfaces next best actions for AEs.
- Forecast Agent — produces probability-weighted forecasts every Monday.
- Insights Agent — answers "show me X" questions in natural language.
- Support Agent — handles tier-1 tickets, escalates the rest with full context.
- Compliance Agent — watches for GDPR/CCPA/DPDP triggers, applies the policy.
- Returns Agent — decides refund-vs-replace at the door.
- Catalog Agent — writes product descriptions, alt text, SEO copy.
Each runs continuously. Each has a policy file you, the admin, control. Each writes its actions back to an audit log that anyone can review.
And then the more important half — the agent builder. Every Polluxa workspace ships with the same builder used to construct the starter set. No-code for the long tail of policy-shaped agents your team can describe in plain language. TypeScript when you need to wire in a private API or a custom scoring model. Same telemetry, same audit log, same handoff protocol as the agents we ship. The starter set is a worked example; the builder is the product.
What customers actually build varies wildly. A jewellery brand built a returns-quality agent that pulls a photo from the customer at receipt and decides resaleable-vs-refurb in seconds. A B2B distributor built a credit-renewal agent that nudges accounts approaching their limit before sales has to chase. A creator-led beauty brand built a shade-launch agent that watches social pre-orders and triggers the regulatory pack the moment a threshold trips. None of those were on our roster. They didn't need to be.
How they hand off to humans
The interesting design problem isn't what an agent can do. It's where the human enters. We use three handoff signals.
Confidence below threshold. Every agent has a confidence score on every action. If it falls below a settable threshold, the action goes to a human queue instead of executing.
Value above threshold. Even at high confidence, certain actions always need a human. A discount above 15%. A contract change. A response to a customer who has paid more than $500k.
Policy exception. Some patterns are pre-flagged: a customer in legal escalation, a regulated geography, a VIP account. Agents see the flag and route to humans automatically.
What the user actually sees
Most days, you don't talk to the agents. You walk into a clean queue. The agent has done overnight prospecting, booked meetings, drafted follow-ups and pulled three deals to your attention because they need judgment you have and it doesn't. You spend your morning on the three deals. The agent handles the long tail.
That's the whole pitch. Software that does the boring 80%, so humans get more of the interesting 20%. It is, in retrospect, the only sensible shape for enterprise software in a world where compute is cheap and intent is expensive.
Want to see the starter set running — and the builder where you'd ship your own? Book a 30-minute walkthrough — we'll run it on a workspace that looks like yours.