The claim doing the rounds is that AI agents will replace whole departments. The reality we see in production is narrower and more useful: agents are absorbing functions, not job titles. A function is a bounded loop of work with a clear trigger and a clear finish — answer the call, qualify the lead, chase the unpaid invoice. Where the loop is that tight, agents genuinely run it end to end today. Where it is not, they still need a human at the edges.
What "agent" actually means
The word is doing a lot of work in marketing copy, so it is worth being precise. The difference is not intelligence, it is who decides the next step.
| Who decides the next step | Breaks when | |
|---|---|---|
| Chatbot | You do — it answers, you act | The answer needs an action taken |
| Workflow automation | The developer did, in advance | Reality goes off the mapped path |
| AI agent | The agent does, at runtime, using tools | The goal is ambiguous or the tools are missing |
That middle row is why so much "agentic" work is really just automation with a model in the loop — which is fine, and often the right build. A workflow automation that never surprises you is worth more than an agent that occasionally invents a step. The honest question is not which is more advanced, it is which one the job needs.
Functions agents genuinely run today
- Front desk. Answering every call, qualifying the caller, booking into a real calendar and writing the record to the CRM. This is the most reliably solved function we deploy — see the AI receptionist for the shape of it.
- Sales development. Sourcing accounts that match a profile, enriching them, sequencing outreach across channels and handling the first replies. Our own Scalaro stack runs this as a set of cooperating agents rather than one.
- Quote and estimate production. Reading a messy inbound enquiry, extracting a structured spec, pricing it against rules and issuing the quote — the core of the FABRIOZA build, which took quoting from days to inside the hour.
- Follow-up that never forgets. The unglamorous one, and often the highest-value: chasing the quote, the no-show, the unpaid invoice, on schedule, forever.
- Intake and triage. Reading what arrived, deciding what it is, routing it to the right place with the context attached.
Notice what these share. Each has an unambiguous success condition — the meeting is booked, the quote is sent, the record exists. That is the property that makes a function automatable, far more than how clever the model is.
What is still marketing
We would rather lose a deal than oversell this, so plainly:
- "Replace your whole sales team." Agents book meetings well. Closing a complex deal involves reading a room, negotiating and being trusted, and they do not do that.
- Fully autonomous multi-agent "companies." Impressive in demos, fragile in production. Every extra agent multiplies the ways a run can drift, and nobody wants to debug a conversation between two bots at 2am.
- "It learns your business on its own." It learns what you write down and give it access to. The work of encoding your rules does not disappear; it moves to the start of the project.
- Anything irreversible without a checkpoint. Sending money, signing, deleting, promising a price outside the rules. Not a capability limit — a judgment one.
The pattern that actually works
The builds that survive contact with real customers all look the same: the agent owns the loop, and a human owns the exceptions. Concretely — the agent handles the routine call and warm-transfers the unusual one; it drafts the quote and a person approves anything outside the pricing rules; it chases the invoice and escalates the dispute. You are not choosing between a human and an agent. You are deciding which slice of a function is routine enough to hand over.
That framing also makes the ROI legible. You are not paying to "add AI" — you are moving a measurable share of a function off payroll, and the ROI calculator will tell you what that share is worth in hours and full-time-employee equivalents before you commit.
How to pick the first function
Pick the loop where the cost of missing it is obvious and the success condition is unarguable. Missed calls are the classic first build because the loss is countable — you can put a number on the enquiries that went to voicemail last month, which means you can also tell whether the agent worked.
Avoid starting with the most complex function to prove the technology. Start with the one where you will know within a fortnight whether it earned its keep, then use that result to fund the next one.
