What it takes to put an AI agent to work

It has become remarkably easy to see what an AI agent can do. In an afternoon, a modern language model can write a friendly appointment reminder, summarise a long email thread, or hold a short phone conversation.

Turning that ability into something a business can rely on every day is a different and very rewarding kind of work. We have now built agents for clinics, insurance agencies, law firms, field-service companies, research teams and sales teams. The settings could hardly be more different, and yet the same ideas have proved useful in each. This article shares them.

Think of the agent as one part of a system

The most helpful shift in thinking is to see the agent as one part of a larger design. Every system we have built has four parts that work together:

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.
The agent. It understands the request and does the task: writes the message, makes the call, drafts the document, prepares the booking.
Built-in limits. Contact hours, daily limits, opt-outs, approvals. These are part of the software, so they apply every time.
A reliable record. The result is saved where the team already works, such as the schedule, the CRM or the matter file.

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

When these parts are designed together, the agent can do its work confidently, and the people around it always know what to expect.

Let the agent write, and let the system decide

Language models are very good at reading and writing. Decisions about whether something should happen, and when, are better made by clear rules that everyone can read.

In our outreach system, for example, the agent drafts every email. Whether an email is actually sent depends on rules the team has set: the recipient has not opted out, it is within working hours, the day’s limit has not been reached, and a person has approved the sequence. In our legal system, one agent drafts a document and a different agent reviews it, and the software ensures those roles stay separate.

This division brings out the best in both. The agent is free to be helpful and fluent. The limits are simple, visible and consistent, which is what gives a team the confidence to rely on it.

Use the simplest tool for each step

Not every step needs artificial intelligence. When a patient replies “Yes” to a confirmation message, a few lines of ordinary code can record that with complete certainty. The agent’s understanding is saved for the replies that benefit from it, such as “I think so, but can you tell me where to park?”

Choosing the simplest reliable tool for each step makes a system faster, more economical and easier to trust. It is one of the quiet pleasures of this work.

Know when the job is done

A thoughtful agent knows when to step back. Once a patient has confirmed, the reminders stop. When a customer asks not to be contacted, that request is honoured on every channel, and stays in place. When someone replies with a question, the automated follow-up pauses so that a person can answer.

People notice this. An agent that finishes politely, at the right moment, is one that customers are comfortable hearing from.

Keep facts in the record

Names, phone numbers, addresses, prices and identifiers should always come from a business’s own systems. In our clinic system, the agent identifies which appointment a message concerns, and the patient’s contact details are looked up from the clinic’s records. In our booking system, every customer and site reference is checked against the company’s data before anything is saved.

This keeps what the agent says in step with what the business has on file, and it means the records remain the single source of truth.

Plan for the unexpected

Technology has off days. A well-designed system treats them calmly. If an agent cannot read a reply, the message is passed to a person to review. If a service is briefly unavailable, work waits where it is until it can continue. Nothing is guessed at, and nothing is lost.

Designing for these moments from the start is what allows a team to stop watching over the system and simply use it.

Design the hand-off

Every one of our systems passes work to people, and that hand-off deserves as much care as the automation itself.

In the clinic, each appointment has its own timeline showing every message, call and reply, so whoever picks it up can see the whole story at a glance. In the insurance agency, an interested customer is connected live to a licensed agent, who can see everything that came before. In the research workspace, an agent asking for a decision always sets out what has happened, why it matters, what it has already done, and what it needs.

A good hand-off means the person can begin where the agent left off.

Learn from what has already happened

Most businesses hold a rich history of their own work: conversations, outcomes, what led to a happy customer. Studying that history before building is one of the most valuable things we do. It often reveals which habits lead to good results, and those findings shape how the agent behaves.

An encouraging conclusion

The most reassuring thing we have learned is that dependable AI is built from familiar ingredients: clear design, sensible limits, good records and respect for the people involved. None of it is mysterious.

With those in place, an agent can take on a great deal of routine work, and give a team more of their day for the work that needs them.