AI Lead Qualification Workflows: DIY Scoring, Human Review, and Vector

What Is an AI Lead Qualification Workflow?

An AI lead qualification workflow takes an inbound inquiry, scores how well it fits, writes a first response, routes it to the right next step, and keeps a person in charge of the decisions that matter. It is not one model making one call. It is a short chain of steps, and the useful question is which step should be rules, which should be a model, and which should stay with a human.

This post covers two versions of that chain. The first is the one on our own contact form, which we built in February 2026 and describe here as it runs in October 2026. The second is Vector, our lead qualification product, built for sales teams whose volume and routing have outgrown a simple scorer.

How Does Our Own Contact Form Qualify a Lead?

Our contact form qualifies a lead in one server route. As of October 2026, it checks the submission, scores it with plain rules, writes a short reply with a model, stores the request, emails the person, and notifies our team when the score is high enough.

The visible form asks for five things: a name, an email address, a topic from a short menu, an optional budget range, and a message of 10 to 1,000 characters. Each one feeds the score or the reply, so there is nothing on the form that we do not use.

When someone submits it, the route runs these steps in order:

1. Rate limit. The route accepts three submissions an hour from the same client and blocks requests it cannot identify, so a script cannot flood the inbox.
2. Validate and sanitize. The fields are checked against a schema and cleaned before anything else touches them.
3. Score. A rule-based function turns the fields into a number from 0 to 100.
4. Store. The request is saved. If the save fails, the person still gets their reply and the route still tries to notify our team, because losing the lead would be worse than losing the row.
5. Reply. A model writes two or three sentences that respond to what the person actually wrote.
6. Email. The route sends the person a confirmation email that includes the same reply. Which version they get depends on the score and the topic.
7. Notify. Depending on the score, our team gets an alert, a standard notification, or nothing.

One rule sits underneath all of it: a contact request is not marketing consent. The form never adds anyone to a mailing list. Broadcast email only goes to people who opt in through a separate flow.

How Is the Score Calculated?

The score is a sum of points, capped at 100. As of October 2026, these are the rules in our route, and the same submission always gets the same score:

These rules are deliberately simple. Anyone on our team can read them, predict a score, and change a weight in one line. That is the main argument for rules over a model at this step: when a score surprises you, you can see exactly why it came out that way.

They are also crude. A two-line message from a serious buyer scores lower than a long message from someone who will never buy. Rules can only count what is on the form, which is why the score decides who hears about a lead, and never whether the lead is answered.

What Does the Score Decide?

The score decides two things: who on our side is notified, and which version of the confirmation email the person receives. As of October 2026, every valid submission gets an on-screen reply and a confirmation email, whatever it scores.

A partnership inquiry gets its own confirmation email, whatever it scores. Two exceptions keep good leads from slipping through on our side: a detailed agency partnership brief notifies the team even when its score is lower, and if the request could not be saved, the route tries to notify the team regardless of score. Every email here is an attempt: if a send fails, the route logs the failure and the request still goes through.

What Does the AI Reply Do, and What Does It Not Do?

The model writes the reply, and only the reply. As of October 2026, GPT-4o-mini gets the person's name, topic, budget and message, and writes two or three sentences that acknowledge the specific need, suggest a relevant next step, and say what happens next.

The instructions are short and strict: use the person's first name, refer to what they actually wrote, be honest if the budget looks low for the goal, and keep it warm and professional. The reply is capped at 300 tokens.

What the model does not do matters as much:

• It does not score the lead. The number comes from the rules above.
• It does not decide who is notified. The score does.
• It does not block the submission. If the model is switched off or fails, the person gets a fixed reply instead, and everything else still runs.

That split is the core design choice. A model is good at reading a message and answering it in plain language. It is a poor place to hide a decision you will later need to explain.

“Automation makes the call. Humans keep authority.”

The Vector product page, Pixelmojo (October 2026)

How Much Does It Cost to Run?

The model is the cheap part. Our estimate, as of October 2026, is under $0.0003 per reply, using the GPT-4o-mini rates in our own price table ($0.15 per million input tokens and $0.60 per million output tokens).

That figure is a ceiling, not a measured average. It assumes the longest message the form accepts (1,000 characters), our full instructions to the model, and a reply that uses the entire 300-token cap. A reply of two or three sentences, which is what the instructions ask for, uses only part of that cap. At the ceiling, a thousand replies cost under 30 cents in model fees.

The real cost of a scorer is everything around the model: the time to build the route, choose the signals, write the emails, handle failures, and keep the rules in line with what actually converts. That cost does not shrink with volume, which is why a simple scorer makes sense for a small inbox and stops making sense for a busy one.

Where Should Human Review Sit in an AI Lead Qualification Workflow?

Human review belongs wherever a wrong call is expensive: before a lead is dropped, before a record changes in your CRM, and wherever the rules cannot see enough to decide. This is our view of where the line should sit, not a rule about how AI systems behave.

On our own form the answer is simple, because the score never drops anyone. Every inquiry is answered, and our team is notified about what the score flags. The automation saves the time it takes to sort, not the judgment about who is worth a conversation.

As volume grows, that stops being enough. When a team receives hundreds of inquiries, nobody reads them all, so the qualification step starts to decide who gets a seller's attention. At that point the workflow needs three things a simple scorer does not have:

• A reason beside every decision, so a seller can see why a lead was routed and disagree with it.
• A way to override, so a person can pull a lead back without editing code.
• A comparison with outcomes, so the rules are judged against what actually closed, not against how they felt when you wrote them.

Lead response speed still matters too. A 2011 Harvard Business Review study of online sales leads found most companies slow to follow up, and its authors put it simply: "most companies are not responding nearly fast enough" (Oldroyd, McElheran and Elkington, 2011). An instant, specific reply is one thing a small team can automate safely.

When Is a DIY Scorer Enough, and When Do You Need Vector?

A DIY scorer is enough when your volume is modest, your routing fits in a few rules, and a person reads every inquiry anyway. Vector is for teams past that point: as of October 2026, it is a fixed-scope build for B2B teams with roughly 50 to 200 or more inbound leads a month, an existing CRM, a defined ideal customer profile, and a seller ready to receive qualified leads.

Vector's own guidance is clear about the other side: fewer than 50 leads a month, no CRM, no salesperson to receive qualified leads, or routing logic you cannot yet spell out usually means Vector is not a fit yet. For that team, a scorer like ours is the right place to start.

Vector does not replace your CRM. It works alongside it: the qualification rationale and the next move can be written into the tools your team already uses, with the exact fields, ownership rules and handoff workflows defined during the build. Pricing is on the Vector page.

Lloyd and Scout walk through how Vector qualifies an inquiry, from the conversation to the handoff:

Read the transcript

This one says urgent. Front of the queue? Maybe. What do they need? It didn't put that on the tag. Inquiries arrive in different ways. A loud request can hide missing details. A quiet one may contain a clear need. So let's open this one. Let's look at the conversation. That's where Vector starts. What problem are they trying to solve? Is there an active project? What is the timing, and who is involved in the decision? This message gives us the problem and the date. Useful evidence. Keep it with the inquiry. And the budget? They haven't told us. That's a question to follow up, not a number to invent. I'll leave it open. Qualification means understanding what we know, what we don't, and what deserves attention next. Start with the customers you can actually help: their problems, requirements, and the situations where your offer fits. So another business might make a different call? Yes. The criteria need to reflect your sales process. Vector is configured around your ideal customer profile, qualification signals, and routing rules. Its assessment should give the receiving team useful context. Can I see why it went that way? The reason belongs beside the decision, with the evidence behind it. Does every inquiry go straight to sales? The next step depends on what we learn. A sales conversation, useful follow-up over time, or a closer look from someone on the team. Each gets a next step. This one doesn't fit the usual pattern. Bring it to a person. Your team can inspect the rationale and change the route. With the original conversation? Yes. Give them enough context to make a considered decision. For the receiving seller, the useful part is the handoff: the need, relevant details, open questions, and why this conversation was routed to them. The inquiry arrives with its story. And the seller can start from there. The build connects qualification to the tools your team uses. We agree the CRM fields, receiving owner, and follow-up workflow. So the next person knows what to do. That's the point. The receiving team helps define the handoff. Vector is a custom build. We establish your baseline, configure the rules, connect the workflow, and test it with your team. And keep checking after launch? Yes. Compare the decisions with real outcomes and adjust where the evidence calls for it. You own the application; Pixelmojo maintains the intelligence platform. What about our urgent one? Check the need. Then choose the next step. The tag can stay. Now it has company.

How Do You Build Your Own Lead Scorer?

If a DIY scorer fits your stage, build it in the same order we did. These steps follow our own route; the effort depends on your stack and how much of it already exists.

1. Decide what the score is for. Ours decides who gets notified. Write that down before you write any rules, because it tells you which mistakes you can afford.
2. Keep the form short. Ask only for what the score or the reply uses. Ours asks for five things.
3. Write the score as plain rules. Points per signal, a cap, and no hidden logic. Put the weights where anyone on the team can read them.
4. Let a model write the reply, not the score. Give it the submission, a few strict instructions, and a token cap. Keep a fixed fallback for when it fails.
5. Route by thresholds, and add exceptions on purpose. Our two exceptions, partnership briefs and failed saves, each protect a kind of lead the score alone would miss.
6. Never lose the lead. If storage fails, still reply and still notify. Rate limit the route, validate and sanitize every field, and log failures.
7. Do not treat a contact as consent. Answer the inquiry; keep marketing email behind its own opt-in.
8. Check the rules against outcomes. Every few weeks, compare the scores with the leads that actually became conversations. Change one weight at a time.

Our ownership guide for AI agency work covers what to ask for if someone else builds it for you, and the multi-channel sales agent post shows how we built the chat side of the same system.

What Changed Since February 2026?

The first version of this post, published in February 2026, compared our form with HubSpot's subscription prices and described Vector as a self-serve product that was coming soon. Both are out of date, so we rewrote the post on 1 October 2026.

As of October 2026, Vector is a 60-day custom build that works alongside CRMs including HubSpot, not a replacement for them, and this post describes our contact form exactly as the code runs today. We also removed a before-and-after results table and a third-party statistic that we could not trace to a source.

Start With the Right Size of System

Start with the smallest workflow that answers every lead quickly and keeps a person in charge. If your volume and routing have outgrown a handful of rules, that is the point to look at Vector.