Ideal Buyer Fit Scoring: How an AI Agent Decides Who Is Worth a Meeting

TL;DR: Ideal buyer fit scoring is deciding who somebody is, not how they happen to behave in the first two minutes of a chat. Get that distinction wrong and an AI agent will either book everyone or write off your best early buyers. In Aijent, two questions you answer at setup teach the agent who your ideal customer is and who it should politely screen out. From then on it may judge durable facts (budget, authority, size, segment, whether someone is buying at all) and is forbidden from counting six normal behaviours against anyone. Where judgement is not enough, because a rule is a rule, SureGate takes over and the same answers get the same outcome every time.

Every inbound sales team already scores fit. It just does it in a meeting room, around minute six, with a silent glance across the table.

The question is whether that judgement happens before the meeting or during it. An AI sales agent on your website can do it before, in the conversation itself, while the visitor is still on the page. If you are researching an AI agent for B2B sales, this is the part worth understanding properly, because it is where the demos say least.

This post covers one narrow job: how an agent decides whether a visitor is worth your team's time. For the wider picture, start with what an AI sales agent is and come back.

What ideal buyer fit scoring actually means

Most people hear "scoring" and picture a number. Points for job title, points for company size, minus five for a personal email address, a threshold and an alert to sales.

That model was built for forms, which only have fields. A conversation has something better: a person telling you, in their own words, who they are and what they need.

So in practice, ideal buyer fit scoring in an AI agent comes down to a label, not a total. In Aijent there are three:

• Good fit. Matches what you said you want. The agent gets on with it.
• Poor fit. A real business, but not the one you asked for. Still helped, still recorded, and you decide what happens next.
• Not a prospect. Not a genuine buyer at all. A bot, a test, or somebody with no commercial intent.

A number invites you to argue about whether a 68 should have been a 72. A label invites the only question that matters: what should happen to this person next? The harder question is what the label is allowed to be based on.

Who somebody is, not how they behave

Here is the trap every fit scorer falls into, human or machine.

A visitor says they are just having a look. They ask a lot of questions. They have not mentioned a budget. They are vague about timing. On a lead-scoring sheet, every one of those is a deduction. On a tired Friday afternoon, every one of those reads like a tyre-kicker.

And every one of them is also exactly what a serious buyer looks like in week one of a six-month decision. Writing them off for behaving like an early buyer is the most expensive mistake a qualifier can make, because you never see it. They do not complain. They just go elsewhere.

So the rule Aijent works to is simple to say and strict in the code: the agent may only judge who somebody is, never how they happen to be behaving early in a chat.

Who somebody is means durable facts, the kind that do not change halfway through a conversation:

• No budget for what you sell
• Not the decision maker
• The wrong size or segment
• An industry you do not serve
• Wanting something you do not offer
• Not a commercial visitor at all

How somebody behaves means transient signals, which change by the next message and say more about the stage of someone's research than whether they will buy.

Two questions set it up

You do not build any of this with a scoring sheet. You answer two questions when you set your Aijent up:

1. Who is your ideal customer? The customers you most want to win. Their size, the person you would actually speak to, the budget, where they are, and the problem they bring you.
2. Who should your Aijent politely screen out? The leads you would rather not sit through. No budget, not the decision maker, wrong size, wanting something you do not offer, or not buying at all.

The second question does the work, and you can skip it. Skip it and every visitor is treated as a prospect.

Your agent reads your website alongside your answers and turns them into the signals it will look for, plus worked examples of a lead to take and a lead to let go. The examples matter more than they sound: without a concrete example of a poor fit, an AI tends to wave everyone through.

What happens to your words on the way in is the interesting part. Setup translates rather than transcribes. Write "tyre kickers with no money" and the behaviour is dropped while the durable part, "no budget", is kept. "Students wasting my time" becomes "non-commercial student". If nothing durable is left once the behaviour is removed, nothing is added. And it never invents a number you did not state: no staff count in your answer, no staff count threshold.

You can change both answers any time in your agent's settings, and the criteria follow them.

The six behaviours that never count

These are six behaviours your Aijent is not allowed to hold against a visitor, whatever you write in your answers:

1. Just browsing.
2. Only researching.
3. Not ready to buy yet.
4. Has not mentioned budget.
5. Asking a lot of questions.
6. Being vague early on.

All normal buying behaviour. None of it counts. Two deserve a closer look, because they catch people out.

"Has not mentioned budget" is not "no budget". The second is a durable fact the agent may judge. The first is silence, and the rule underneath is that the absence of information is never disqualifying. Nobody is downgraded for what they have not said yet.

"Asking a lot of questions" is a fit signal, not a warning. The screening instructions say so directly: asking several genuine evaluation questions, asking to buy, book, see a demo or start a trial, or naming their own CRM, team, process or volume all point toward a good fit. A visitor asking question after question is qualifying themselves the healthy way.

Patience is a rule in its own right. The agent is told never to rush to a verdict to look decisive. A bare opening, a greeting or "just looking" with nothing real volunteered, is held as "not yet known", and that hold is never a downgrade. The one exception is a durable fact rather than a behaviour: if a visitor clearly says they are a student, a researcher or simply curious, and has shown no real business or need, the agent can call that early.

The guardrails on the label

A few rules decide which way the agent leans when it is unsure.

It screens out, it does not screen in. Screening exists to flag the few people who clearly should not be pursued. It starts from the assumption that every visitor is a real prospect and only moves off that on positive evidence.

A real business is never "not a prospect". If there is a real business or a real buyer at all, the worst label they can get is poor fit. When the agent is unsure between the two, it chooses poor fit.

Anyone who asks for a meeting is never written off. A visitor who asks to buy, book, see a demo or start a trial cannot be labelled "not a prospect". This is checked twice, once in the instructions and again in code afterwards, so a visitor who asked to book is at worst a poor fit.

Labels only move up. A visitor can move from not a prospect to poor fit to good fit as they say more. Good fit is final.

Every conversation is tagged. Your dashboard shows the split, with a short reason naming the signal that decided it. If a reason reads wrong, you change your answers.

Fit runs both ways

There is a second fit judgement in every sales conversation, and it gets less attention than it should: whether your product actually suits what the visitor described.

Aijent works to a fit rule on this side too. Whether the product can do a particular job is treated as a claim like any other. Before the agent says the product suits, fits or can do what the visitor described, your website content or the verified facts you gave it must describe that job. If the visitor's words could mean two things, and the answer would differ between them, the agent asks one short question first. It never confirms the fit and checks afterwards.

That matters for fit scoring because a meeting booked on a misunderstanding is a bad meeting, however well the visitor scored. Fit is not only whether they are your buyer. It is whether you are their answer.

Where judgement stops and a rule takes over

Everything so far is judgement. Carefully bounded, but judgement, and for most businesses that is the whole job.

Some decisions are not a judgement call. If a project under a certain size is never worth a meeting, or you only work in certain states, that is a firm rule rather than an opinion, and it should give the same answer every time.

That is what SureGate is for. It sits on top of Visitor Screening (screening is step one, and SureGate does not run without it), and you write the rule yourself. Criteria come in three kinds:

• A number. How many staff?
• An amount of money. What is the budget?
• A category. Which state are you in?

Each criterion gets its own thresholds: where Go begins and where Review begins. For numbers you choose whether higher or lower is better. For categories you list which answers go and which get reviewed, and choose what happens to everything else. The criteria, the thresholds and the wording are entirely yours.

For each criterion you also choose how the agent learns it. Either it asks, using the exact question you wrote, woven into the conversation one at a time in its own voice, or it only uses the answer if a visitor volunteers it. The visitor never sees a rule. They have a normal conversation, and your criteria are answered inside it.

Then the rule decides. A fact counts only if the visitor stated it in their own words, and plain code compares it against your thresholds. Same answers, same outcome. If several criteria are known, the most restrictive one wins. Two protections carry straight over from screening:

• An unknown answer is silent. No value, no block, no downgrade.
• Refusing to answer is never a no. A declined question resolves to Review, never No-go.

Three outcomes, and the warm decline

Every SureGate conversation lands in one of three places.

Go. They meet your criteria. Your Aijent proceeds exactly as it would have anyway: it books the meeting or shares your link.

Review. Borderline, so a human decides. Review follows the single "when unsure" setting you already chose for screening: route the visitor to a person, or let them proceed with the conversation tagged for review. This is the safety net.

No-go. Not a fit, and told so kindly.

A badly handled no is brand damage, so the decline is built so the agent cannot fumble it. From the moment a visitor is out, the agent keeps answering their questions from your content, in the same name and voice, but it no longer has any way to book a meeting or share your conversion link. Those tools are removed, and your link is kept out of what it can see.

The agent is never asked to perform a rejection, and never shown the decline wording. When the visitor reaches for the next step, the decline is posted once, verbatim, in words you wrote yourself. It never names the line that fired, so nobody hears "your budget was too low". If they insist, the no is not repeated: they are offered a callback form, once, and your team decides. The full transcript is saved, so you can audit every no your agent has ever given.

We wrote more about the thinking in the door policy guide and in We Built an AI That Says No.

A worked example

Illustrative, not a real customer. A commercial fit-out business answers the setup questions: the ideal customer is a business fitting out an office or shop, with someone who can approve the spend, in the eastern states. Screen out homeowners wanting small residential jobs, job seekers, suppliers pitching, and students.

Three visitors arrive the same afternoon.

The first is "just looking at options" for a new office, asks six questions about timelines and materials, and never mentions money. None of that counts against them. They are held as a real prospect, and when they ask to book, the agent books.

The second is a homeowner asking about a bathroom. A real buyer, but not the one the business asked for: poor fit, still helped, still recorded, and the "when unsure" setting decides whether a person follows up.

The third is a student writing an assignment: not a prospect, helped anyway, and politely declined a meeting.

Now the business moves to Growth and adds one SureGate rule: projects under its minimum size are a warm no. A fourth visitor, a genuine office manager, says the job is well under that figure. That is no longer a judgement. The rule decides, the same way it would for anyone giving that answer, and the decline is the one the owner wrote.

Where to start

Visitor Screening is included on every plan, at no extra cost. You set it up by answering two questions during onboarding, and you can change your answers whenever your idea of a good customer changes.

SureGate is available on Growth and above. One flat monthly price, conversations included, no setup fee and no contract.

Starter is $99 a month and comes with 14 days free, no credit card. Everything each plan includes is on the pricing page.

Fit scoring was never about collecting more points on more people. It is knowing who somebody is, being patient with how they behave, and drawing a firm line only where you mean one.