Where AI agents pay back first in a mid-sized business

Every leadership team I speak with has the same question about AI: where do we start? Usually it comes after a few experiments that impressed everyone in a demo and changed nothing on the P&L.

The problem is rarely the technology. It is the starting point. Agents pay back fastest where three things are true at once: the work is frequent, the work is rule-shaped but messy, and the cost of a small mistake is low and easy to catch. In a business with somewhere between 20 and 500 people, those conditions show up in the same handful of places again and again.

1. The inbox nobody owns

Every mid-sized business has a shared inbox, a contact form or a WhatsApp number that someone checks "when they get a minute". Leads wait. Customers chase. Supplier questions go round in circles.

An agent that reads every incoming message, works out what it is (new enquiry, order question, complaint, invoice, spam), pulls the relevant record from your CRM or order system, drafts a reply and routes anything unusual to a person is one of the quickest wins available. It does not need to be perfect. It needs to be faster and more consistent than "when someone gets a minute".

What to measure: time to first response, and the share of messages that are resolved without a person touching them.

2. Moving data between systems by hand

Look for the job that involves two browser tabs and a spreadsheet: copying orders from a marketplace into the ERP, re-keying supplier invoices, updating stock levels across channels, exporting a report every Monday so someone can paste it into another report.

This work is boring, error-prone and invisible, which is exactly why it survives. Agents are very good at it, because the rules are already known by the person doing it. Sitting with them for an hour and writing those rules down is usually most of the project.

What to measure: hours per week returned to the team, and the error rate before and after.

3. First-line answers from your own knowledge

Your team answers the same 40 questions every week, from customers, from new staff and from each other. The answers live in PDFs, old emails and one experienced person's head.

An assistant grounded in your own documents (policies, product specs, pricing rules, process notes) can answer most of those questions instantly and point to the source. The payback here comes from fewer interruptions for your most experienced people, and faster onboarding for new ones.

What to measure: questions answered, and how many were escalated to a person.

4. Follow-up that never happens

Quotes that were sent and never chased. Customers who bought once and never heard from you again. Abandoned carts. Renewals that come up without warning.

Follow-up is high-value and low-glamour, so it is the first thing to slip when a team is busy. An agent that watches for these moments and drafts the right next message, for a person to approve or to send automatically within agreed limits, turns forgotten revenue into a routine.

What to measure: conversion on quotes and repeat purchase rate.

Where not to start

Avoid starting with anything that is high-stakes, rare, or politically sensitive: pricing decisions, hiring, legal positions, or a complete "AI transformation" of a core process. Those may come later, once the business has seen agents work and trusts the controls around them.

Also avoid starting where the data is a mess. An agent cannot fix a CRM that nobody keeps up to date; it will just make the mess faster.

How to pick your first agent

Score each candidate from one to five on four questions:

1. Frequency. How many times a week does this happen?
2. Clarity. Could a sensible new hire follow the rules after a day of training?
3. Reversibility. If the agent gets one wrong, how easily is it caught and fixed?
4. Value. What is an hour of this work, or one faster response, worth to the business?

Start with the highest total, not the most exciting idea. Put a person in the loop for the first few weeks, measure against the baseline you took before you started, and only widen the agent's authority once the numbers hold.

The real payback

The return from the first agent is rarely the hours it saves. It is what the business learns about its own processes in the course of building it, and the confidence to tackle the next one. In my experience, the companies that get real value from AI do not run one big programme. They ship one useful agent, then another, and keep a scoreboard.

If you would like help choosing your first one, that is a good use of a private consultation.