Most businesses that decide to try an AI agent start in the wrong place. They pick the best demo they saw and then look for work to give it. We start the other way round: find where work visibly gets stuck, then check whether the right answer already exists in material you own.
For a business with real operations work, that is usually the phone or the inbox. A first agent should be easy to judge, easy to correct and useful in the week it goes live.
Start where the work gets stuck
Work gets stuck in small, repeated places: the call nobody picks up while everyone is out on a job, the email that sits until someone has a spare hour, the same details written down once and typed again by the next person, the question only one colleague can answer.
At one of the companies we have worked with, most new work began as a phone call, and what was said on that call was written down by one person and typed again by another further along. It did not need a clever agent. It needed the first step caught properly, once.
Where people wait, retype or answer late, you have a candidate. Put each one through five questions.
Five questions for choosing the first agent
- How often does it happen? An agent earns its place on work that arrives every day. Rare tasks are better left with a person.
- Is the right answer already written down? If the answer lives only in one person’s head, write it down first. An agent is only as good as the material you approve.
- What does a wrong answer cost? A wrongly labelled email is fixed in seconds. A wrong promise to a customer is not. Start where a mistake is cheap and easy to see.
- Can a person approve before anything leaves? The safest first agent prepares the work and lets a person send it. One that answers customers directly needs a clear line it does not cross and someone to hand to.
- Does the result land in the system you already use? If the output sits on a new screen nobody opens, the work is still stuck. It should arrive in the mailbox, calendar, schedule or record your team already works from.
A task that does well on all five is a good first agent.
Four good first agents
These are four of the AI agents we build: when each is the right place to begin, when it is not, and what has to be true first.
Inbound call taking
It is the right first agent when calls are missed because the people who would answer are busy with the work itself, or when calls come in after hours and on weekends. Inbound call taking picks up, says at the start that it is an AI assistant, books onto your real schedule, transfers emergencies to a named person and writes a summary and the reason for every call.
It is not the right first agent when most calls need judgement written down nowhere, such as a negotiation.
Before it works, you need the topics it may handle written down and approved, the schedule it books into, and a named person for emergencies and hand-offs. It answers only from material you approved. Outside that, it takes a message or hands off.
Inbox triage and reply drafts
It is the right first agent when the morning goes on email and the few messages that matter are found late. Inbox triage and reply drafts reads incoming mail in Gmail or Outlook, labels it with the categories you set, moves urgent messages to the top and writes reply drafts from your approved answers.
It is not the right first agent when most of your email needs a fresh answer every time.
Before it works, you need the categories, your own definition of urgent and a set of approved answers to draft from. The drafts wait in your drafts folder, and nothing is sent until a person sends it. It is the gentlest start of the four.
Customer support agent
It is the right first agent when customers message on WhatsApp, website chat or text and then wait while your staff are busy. The customer support agent says at the start that it is an AI assistant, answers from the FAQ and documents you approved, and passes anything else to a person with the conversation so far.
It is not the right first agent when you have no FAQ yet, or when most messages are complaints and refunds, which go to a person anyway.
Before it works, you need the FAQ and documents signed off, and a hand-off path: who picks up the conversations it passes on. Change the material and its answers change.
Company knowledge assistant
It is the right first agent when customers are looked after but staff keep asking questions that a document already answers. The company knowledge assistant answers in Slack, Teams or a web chat from the documents you choose, links to the source behind each answer, and says plainly when it does not know, pointing to the person who owns the answer.
It is not the right first agent when the documents are out of date or contradict each other. It will repeat whatever they say.
Before it works, you need the folders and pages it may read, a person who owns each document, and permissions set so that staff get answers only from documents they are allowed to open.
Where not to start
Do not start with the most impressive demo. An agent that plans, decides and acts across several systems is hard to judge and hard to correct.
Do not start with work whose answer is not written down. The agent will either guess or hand everything back, and neither helps.
Do not start where a wrong answer is expensive and nobody checks before it goes out: commitments to customers, anything that moves money, anything that changes an account.
And do not start with an agent whose output lands somewhere new. If your team has to open another screen to see what it did, they will stop looking.
How a person stays in control
Where an agent writes for you, a person approves anything that leaves the building. That covers reply drafts, outreach emails and social posts. Where an agent answers customers directly, on the phone or in chat, nobody approves each answer as it goes. Instead: It answers only from material you approved. Outside that, it takes a message or hands off.
Who owns it
Every agent we build sits inside your own accounts, so you own the agent and the intelligence in it: the rules, the prompts and the data. We build it and set it up for you, and answers land in the system of record you already use.
Where it runs, whether Copilot Studio, n8n or custom code, is a separate decision that comes after this one. For the wider choice of workflow tools, see Zapier vs Make vs n8n vs custom code.
Where to start
Pick one place where work visibly gets stuck. Put it through the five questions. Write down the material the agent will answer from and the person it hands to. If you cannot do both yet, that is the first job.
Talk to sales. Tell us where the work gets stuck and we will tell you which agent, if any, should go first.
