How to brief your first AI agent

If you have decided to bring an AI agent into your business, the temptation is to start with the technology. Which platform. Which model. Which integrations. Which prompt.

That is the wrong starting point.

We have built and deployed AI agents for our own business and for customers, and the single biggest predictor of success is not the technology. It is the brief. The agents that work are the ones that started with a clear, honest answer to "what do we actually want this thing to do?" The ones that fail are the ones that started with "we want AI" and worked backwards from there.

This is the playbook we use ourselves. It is the same brief we ask our customers to fill in, and the same questions we walk through in discovery. If you are about to recruit your first AI agent, this is the work to do before you commit.

A founder sitting at a kitchen table with a notebook and laptop, working through a hiring brief checklist.

Why "AI agent recruitment" is the right framing

The phrase you want in your head while you read this is recruitment, not procurement.

Procurement implies buying a tool. You define the spec, find the vendor, sign the contract, and integrate. The tool either does the job or it does not. You replace it if it underperforms.

Recruitment implies onboarding a teammate. You define the role, find the right shape of teammate, write the brief, agree the scope, set the success criteria, and onboard them properly. You also expect to coach them in the first few months as they learn your business.

AI agents work much better when treated as the second. They have a job description, a scope, a voice to learn, and edge cases to be coached on. The customers who think of it as procurement get frustrated when their agent does not magically know things. The customers who think of it as recruitment get an agent that works.

Question 1: What is the actual problem?

The most common mistake is starting with the solution.

"We want an AI agent for our customer service" is a solution. It assumes the answer is an AI agent, and the question becomes which one.

The right starting point is the problem. What is happening in your business right now that is not working? Some examples of well-formed problem statements:

"Our shared inbox has 90 minutes of triage every morning before the team can do real work."
"We get 30 inbound enquiries a week and the founder personally replies to each one. Most are not qualified leads."
"Our pipeline has 80 deals in 'warm' stage that nobody has touched in two months."
"We send the same six explanations to customers every week, and they often arrive at midnight when no one is awake."
"We are spending six hours a month manually re-keying data from PDF invoices into our accounting system."

Each of these is solvable by an AI agent, but the agent that solves it depends on the problem. The shared inbox problem maps to an AI Customer Success Agent or a Ticket Triage Agent. The lead-qualification problem maps to an AI Lead Qualifier. The stale-pipeline problem maps to an AI Follow-up Agent. The same-explanations-at-midnight problem maps to an AI FAQ Agent.

Start with the problem. Let the role follow.

Question 2: What does success actually look like?

If you cannot describe what success looks like, you cannot tell whether the agent is working.

Success criteria should be specific, observable, and honest. Some examples:

"Inbox is at zero unread by 9am every weekday."
"First response time on any inbound enquiry is under 5 minutes, around the clock."
"70% of routine support questions are resolved without a human."
"Outreach goes to 50 prospects a week, with at least 5% reply rate."
"Pipeline contains zero deals untouched for more than 14 days."

What success criteria should not be: vague aspirations like "save time" or "be more productive." Those are real outcomes, but they are not measurable. You will not know whether you got them.

Be specific. Be observable. Be honest. Aim for outcomes the agent can actually move, not outcomes that require a hundred other things to also be true.

Question 3: What does the agent absolutely not do?

Just as important as the scope is the anti-scope.

Every agent we deploy has a "what this agent does NOT do" section. It is the boundary that protects the agent from being asked to do things outside its competence, and protects the customer from making promises the agent cannot keep.

For your brief, list at least five things the agent should not do. Examples:

Does not commit to pricing or negotiate discounts
Does not handle billing disputes, refunds, or contractual amendments
Does not pretend to be human when asked
Does not blast generic templates without personalisation
Does not contact prospects on a suppression list

These are constraints that customers sometimes find counterintuitive. "Why are you telling me what your agent cannot do?" Because every agent that works in production has clear limits. The ones without clear limits are the ones that hallucinate, drift, or accidentally promise things they should not.

Question 4: Whose voice does it speak in?

Voice matters more than people expect.

An agent that does its job well but sounds wrong will undermine the brand it is supposed to support. An agent that sounds like the founder, but with the patience and consistency the founder cannot maintain, will quietly become one of the most valuable things in the business.

Before you deploy any agent, give them voice samples. Specifically:

5 to 10 examples of emails or messages the founder has actually sent
The brand voice rules (warmth, formality, humour level, prohibited phrases)
Words and phrases the brand actively avoids
The signature touches that customers have come to recognise

We have seen agents that worked operationally but failed on voice get pulled within weeks. We have also seen agents whose voice was so well calibrated that customers did not realise they were talking to an AI until it told them so.

Question 5: What are the integrations?

Standard integrations are usually fine. Custom integrations are where projects slip.

Make a list of every tool the agent will need to read from or write to. CRM, email, help desk, calendar, project tool, accounting, billing, analytics, the platform your store runs on. For each tool, note whether it has a public API and whether the agent will need read-only or read-write access.

Most reputable platforms (HubSpot, Salesforce, Pipedrive, Gmail, Outlook, Chatwoot, Intercom, Stripe, etc.) are well-supported. Niche platforms, internal tools, or legacy systems are where time and money compound. If the agent needs to integrate with a custom in-house dashboard, that is a custom integration with a separate scope.

The honest answer to "can this agent integrate with our system" is: it depends on what the system exposes. Public API: usually yes. Webhooks: probably yes. Database access: maybe, with caveats. Screen-scraping a legacy interface: technically possible, but think hard about whether you really want that.

Question 6: What is the escalation pattern?

Every agent that works in production has an escalation pattern. The question for your brief is: what is yours?

Specifically:

What categories of conversation always escalate to a human? (Legal, financial, contractual, sensitive customer issues, public complaints)
Who do they escalate to? (Specific person, specific Slack channel, specific email address)
What does the escalation handoff include? (Full context, agent's reading of the situation, suggested next action)
What is the response time the customer should expect after escalation?

Get this right and the agent feels like a colleague who knows when to ask for help. Get it wrong and the agent will either bottleneck on every edge case or push through situations it should not handle.

Question 7: How will you know if it is working in 30 days?

Set the 30-day review now, before the agent goes live.

Define what you will measure, where the data will come from, and what threshold counts as "this is working." Examples:

"30-day review covers: time spent on inbox triage by the team, first-response time on customer enquiries, percentage of conversations resolved without human intervention. We will pull the data from Chatwoot and the team's calendars."
"Working = inbox triage time below 30 minutes/day, first-response time below 5 minutes, resolution rate above 60%."

Anything below the threshold gets adjusted. Anything above is locked in. The 30-day review is not optional. It is how the agent learns the parts of your business it could not learn from the brief.

What to avoid

Three traps we have seen kill more AI agent projects than anything else.

The "let it figure it out" trap. Customers who want the agent to discover its own scope, voice, and limits in production. This never works. The agent needs a brief. Without one, it will drift, and you will spend the first month firefighting outputs that should have been defined upfront.

The "everything at once" trap. Customers who want a single agent to do sales, customer service, content, and operations simultaneously. We do not deploy this shape. Each function is a different role with different success criteria. One agent doing everything is one agent doing nothing well.

The "set and forget" trap. Customers who think AI agents are a one-time deployment. They are not. The 30-day review is real. The 90-day review is real. The agent that worked perfectly six months ago might need tuning as your business evolves. Treat it as an ongoing relationship, not a delivery.

What to do next

Take the seven questions above and write your answers down. Even a rough first draft is enough. Then either:

Use them as the basis of a hiring brief if you are working with us. Our brief covers the same ground in a structured form.
Use them as the spec for any other AI tool you are evaluating. The questions are platform-agnostic. Anyone selling you AI should be able to answer them clearly.

If our model fits what you are trying to do, here is where to go next.

Explore the catalogue to see agents by department and tier. Each role page shows what the agent does, what it does not do, what integrations it supports, and what it sounds like in its own voice.

Briefing your first AI agent: the questions to answer before you start