Most businesses don't need one all-powerful AI assistant. What they need is a small team of bounded agents. Each one has a narrow job, works from the same customer record and hands sensitive decisions to a person.
That's how I design AI for service businesses now. One big "AI employee" with access to everything sounds efficient. In practice it's hard to test, hard to trust and hard to fix when something goes wrong. Several small agents with clear jobs are easier on all three counts.
The catch is that the operating model has to come before the prompts. Here's what that looks like.
Pick one repeated process with a visible problem. Missed lead follow-up. Appointment juggling. Chasing documents. Service triage. "Where's my order?" enquiries.
Map the trigger, the trusted record, the allowed actions, the exceptions, the approvals and how you'll measure it. Then choose the technology. Often a plain workflow covers most of it, like the five HubSpot workflows that replace 20 hours of manual work a week. Agents fill the gaps where judgement is needed.
My rule is simple: one visible problem, one credible first step. An agent should remove one bounded piece of friction. It shouldn't become a vague digital staff member with access nobody can explain.
You won't need all seven on day one. Most businesses start with one or two. But this is the shape a mature setup tends to take.
Gathers approved public or internal context, flags missing information and prepares a structured brief. It doesn't guess facts, and it doesn't change a record without a defined rule.
Checks submitted information, spots missing fields, links the submission to the right customer record and prepares the next admin step.
Applies your documented criteria, assigns an owner and sends edge cases to a person. People approve any consequential eligibility, commercial or clinical decision.
Coordinates availability, sends approved reminders and records attendance or cancellations. When bookings clash or look odd, it raises them with a person and leaves the call to them.
Pulls together documents, history and approved knowledge for the team member doing the work. In healthcare it can organise and prepare information. It must not diagnose, interpret results or prescribe.
Watches the gaps between systems: failed notifications, bounced emails, missing documents, stalled stages and integrations that didn't finish. It builds a visible queue with an owner and a due date.
Summarises throughput, exceptions, ageing work and outcomes from the one record. Every claim links back to the underlying record, so a manager can check it in seconds.
One record. HubSpot, or another trusted system, holds the customer, company, deal, service and activity history. Agents read and write through governed fields and actions. No private spreadsheets on the side.
A self-service portal. Customers see their status, documents, bookings, payments and next steps in one signed-in place. That cuts down "just checking in" calls, and it gives agents the right context about who they're dealing with.
Bounded agents. Every agent has a narrow purpose, limited tools, documented data access, run limits and clear exception rules.
People approve. Sending outreach, changing commercial terms, acting on sensitive personal information and making clinical or professional judgements stay with people.
None of this works if the team won't use the system. Keep the human side simple. The 3-click rule for CRM adoption applies to approval queues too.
There's a regulatory reason to get this right as well. From 10 December 2026, organisations covered by the Privacy Act that use personal information in automated decisions that could significantly affect someone's rights or interests will need to explain this in their privacy policy. The OAIC has consulted on guidance for the new obligation. If you know exactly what each agent decides and what data it uses, that disclosure is much easier to write.
The best design is the smallest governed team that improves a real process. Long instruction sets and clever prompts matter far less than a clean record, clear boundaries and a person on the important calls. I learned plenty of that the hard way, and wrote about it in the honest truth about replacing a hire with an AI workflow.
For Australian telehealth providers, the same model applies with a stricter boundary. Agents can organise, extract, route, notify and prepare. Named clinicians keep clinical decisions.
If you want to map your first agent team around one high-friction process, see how we build AI agents and tell us where the friction is.