What it takes to put an AI agent to work

A language model can write an appointment reminder, summarise an email thread or hold a short phone call in an afternoon. Putting one to work in a business every day takes a system around it: a way for work to arrive, limits the software enforces, a record the team already uses, and a hand-off to a person for anything that needs judgment.

We run seven client agents in daily use: Remi confirms clinic appointments, Maverick follows up insurance leads, Blue drafts for law firms, Proctor books and routes jobs for a waste-services company, Apollo supports a research consultancy, Foreman plans production for a manufacturer, and Herald runs outreach for a sales team. The settings could hardly be more different. The same eight rules came up in each, and each is shown below in the agent where it does the most work.

What does an AI agent need around it to be dependable?

Four parts, designed together:

1. A way for work to arrive. A text, a call, an email, a new enquiry, an uploaded file. Customers keep communicating the way they already do.
2. The agent. It understands the request and does the task: writes the message, places the call, drafts the document, prepares the booking.
3. Limits the software enforces. Contact hours, daily limits, opt-outs, approvals. Because they live in code, they apply every time.
4. A record the team already uses. The result is saved where the team works: the schedule, the CRM, the matter file.

Around all four: anything that needs judgment goes to a person, with the full history attached.

Remi shows all four. A patient’s appointment arrives from the clinic’s schedule. The agent texts, then calls. The clinic’s calling hours and the rule that one answer ends the outreach are enforced by the software. The answer is written back to the clinic’s booking system. Cancellations, reschedules, questions and silent patients go to the front desk. At an ENT clinic in Ontario, that system confirmed 92% of 901 confirmations over 30 days (the study).

Who should decide whether an AI agent sends something?

The system, not the agent. Language models are good at reading and writing. Whether something should happen, and when, is better decided by rules everyone can read.

In Herald, the agent drafts every email. Whether an email is sent depends on rules the team set: the recipient has not opted out, it is within working hours and not a weekend or public holiday, the day’s limit for that mailbox has not been reached, and a person has approved the sequence. In Blue, one agent drafts and a different agent reviews, and the software enforces that only the drafter can write to a document and only the reviewer can give the verdict.

The agent is free to be fluent. The limits are simple, visible and the same every time, which is what lets a team rely on it.

Does every step need AI?

No. Use the simplest tool that is reliable for each step. When a patient replies “Yes” to Remi, a few lines of ordinary code record the confirmation with certainty; the language model is kept for replies like “I think so, but where do I park?” When a booking reaches Proctor in a fixed format, such as a standard form, ordinary code reads it, and the agent is kept for the free-text messages.

The simplest reliable tool makes the system faster, cheaper to run and easier to check.

How does an AI agent know when the job is done?

It is told, in rules. Remi stops the moment a patient answers, and that patient is not contacted about the visit again. Maverick honours a request to stop on every channel, and it stays in place. Herald ends the automated sequence the moment a person replies, whatever the reply says.

An agent that stops at the right moment is one customers do not mind hearing from.

Where should an AI agent get names, numbers and prices?

From the business’s own systems, never from the model. Remi identifies which appointment a message concerns, and the patient’s contact details are looked up from the clinic’s records. Proctor checks every customer and site reference against the company’s data before a booking is saved, and the price is always the customer’s agreed rate. Foreman stays in step with the manufacturer’s ERP in both directions: items, customers and stock flow in, and work orders and completed production flow back.

What the agent says stays in step with what the business has on file, and the records remain the single source of truth.

What happens when something breaks?

The work waits, and nothing is guessed. If a service Herald depends on goes down, work pauses where it is and resumes afterwards, and unsubscribe requests are still honoured in the meantime. Proctor regularly confirms that every truck is visible on the live map the dispatcher relies on. If an Apollo agent asks the other for something and gets no reply, the request is sent again and then raised on the project where a person will see it. Blue stops any edit that would remove a large part of a document’s text, formatting or formulas before it happens.

Designing for these moments up front is what lets a team stop watching the system and use it.

What should a hand-off to a person include?

Everything the person needs to start where the agent stopped. In Remi, each appointment has a timeline of every text, email, call and reply, so whoever picks it up sees the whole story in a few seconds. In Maverick, an interested customer is connected live to a licensed agent, and a callback goes to the agent who has been working that lead. In Apollo, an agent asking for a decision always sets out what happened, why it matters, what it has already done, and what it needs.

Should you study your own data before building an agent?

Yes. A business’s own records hold the history of its work, and reading it before building is one of the most valuable things we do. Before we built Maverick, we read 1,283 sold insurance leads end to end and compared them with 1,471 lost ones. Speed to lead barely differed between the two. What differed was persistence: the median sold lead was still being worked 21.2 days after it arrived, and 1,252 lost leads were never reached once. Maverick’s priorities changed because of it. The full findings are published.

What does it add up to?

Dependable AI is built from familiar parts: clear design, limits in code, good records, and a person for anything that needs judgment. None of it is mysterious. With those in place, an agent takes on the routine work and the team gets its day back for the work that needs them. The seven agents are on our solutions page.