What you need ready before an AI agent is worth installing

Agents fail for boring reasons, and almost never technical ones. Nearly every disappointing project traces back to one of four things being missing on the customer's side before anyone started building.

Here they are, so you can check yourself before you pay anyone.

1. Enough questions to be worth it

There is a floor below which this is not worth doing. If you get five customer messages a week, answer them yourself. You will do it better and it will cost you less.

The point where it starts paying is when the same questions come back often enough that answering them has become a job someone does instead of doing something else. That is usually somewhere north of twenty or thirty messages a week, and it depends more on repetition than on volume. Thirty identical questions is a great case. Thirty different ones is not.

2. Answers that exist somewhere

The agent needs source material. Not perfect documentation, but the answers have to exist outside of somebody's head.

Any of these work: your website, a product sheet, an old email where you explained the policy properly, a spreadsheet of prices, the folder of replies your team already copies and pastes. What does not work is "everyone here just knows".

The good news is that this work is worth doing regardless. Every business that goes through it finds contradictions it did not know it had.

3. A person who owns it

Somebody on your side has to be responsible for the agent. Not full time, and not technical. Someone who reads what it said this week, notices the answer that was off, and says so.

An hour a week for the first month, much less after that. Projects without this person drift: nobody notices the agent has been giving a slightly wrong answer for six weeks, and confidence quietly collapses.

4. A decision about what it must not touch

Before launch, someone with authority has to draw the line. Which questions is the agent allowed to answer, and which ones go straight to a person, always.

The usual list of things it should not decide alone:

• Refunds, discounts and exceptions to your own rules.
• Delivery promises with a date attached.
• Anything about someone's health, money or legal situation.
• Complaints that have already escalated once.
• Prices for anything non-standard.

This is a business decision, not a technical one, and it cannot be delegated to whoever builds the thing.

What you do not need

Some things people assume are prerequisites and are not:

• Technical staff. If the agent is built properly you need to be able to read and write, not to code.
• A tidy CRM. Useful, not required. An agent can be genuinely helpful with no access to any internal system.
• A big website. The channel matters more than the site. Plenty of businesses get most of their questions through Messenger and Instagram.
• Perfect documentation. Scattered and correct beats organised and out of date.

A test you can run this week

Open your inbox and your social messages. Take the last hundred conversations and sort them into three piles:

• Same answer every time. This is what the agent takes.
• Needs a person to decide. This is what it hands over.
• Exists because something is broken. This is a business problem, and an agent will only make it faster to explain.

If the first pile is the biggest, you are ready and it will pay for itself. If the third one is, fix that first. It will be cheaper and the questions will stop instead of getting served.