How to automate lead qualification

Inbound arrives through a form, an inbox, a WhatsApp number and occasionally a phone call somebody wrote on a notepad. Someone reads all of it. The serious enquiry that came in at nine o'clock waits behind four suppliers pitching SEO services and a student asking about work experience.

That queue is the problem, and it is rarely solved by trying harder. It is solved by deciding, for each enquiry, what should happen next, which is what qualification actually means.

Qualification is a decision, not a score

Most scoring projects produce a number nobody acts on. Start from the other end: there are only three outcomes worth building for.

Hand it to a person now. Ask two questions and wait for the answer. Or decline politely, so the sender is not left waiting and your team never sees it again.

Design the system to produce one of those three, and the score becomes an implementation detail. A system that produces a number between 0 and 100 without a rule attached to it produces meetings.

The pipeline, step by step

1. One place for everything. Forms, the shared inbox, WhatsApp, and a note field for calls, all landing in the same table with the source recorded. Attribution has to be captured at intake, because reconstructing it later is guesswork.
2. Enrich from company data, not personal data. Take the email domain and look up what the business does and roughly how big it is. Use the company as the unit wherever possible: it is more stable, more useful, and it keeps you away from building profiles of individuals. Where personal data is involved, enrichment is processing like any other, which the GDPR basics covers.
3. Classify the intent from the text. This is the part where a model genuinely beats rules. Request for a quote, support question from an existing customer, job application, vendor pitch, partnership request, spam. Ask for a fixed set of categories and a confidence, and send anything ambiguous to a person rather than guessing.
4. Score fit against evidence, not opinion. Pull your last two years of closed deals and look at what they had in common: sector, size, the problem they arrived with, which channel they came through. Build the definition from that. A scoring model built from what the sales team believes about good customers encodes the belief, not the customers.
5. Route with rules you can read. High fit and clear intent go straight to a person with the enquiry summarised and the enrichment attached. The middle gets an immediate, useful reply that asks the two questions you always end up asking anyway. Out of scope gets a short, honest no.
6. Write back the reason, not just the number. The CRM record should say why something was routed as it was, in one line a salesperson can read. This is what makes the system correctable, and what lets you find out later that it was wrong.
7. Check it against outcomes monthly. Take the leads the system scored low and find out how many became customers anyway. That number is the only honest measure of whether the scoring works, and it is the one nobody looks at.

The scoring mistake that hides its own damage

A scoring system trained on which leads your team converted learns your team's habits as much as your customers' characteristics. If salespeople have historically ignored enquiries from free email addresses, from companies with foreign names, or from sectors they find unfamiliar, the model learns to score those low.

Then the loop closes. Low-scored leads never get called, so they never convert, so the next round of training confirms that they were low quality. The system is not wrong by its own measure, and its own measure is the problem. Worse, in a hiring or credit context this pattern is not only commercially damaging, it is the kind of indirect discrimination regulators take an interest in.

Two defences, both cheap. Always sample: route a random small percentage of low-scored leads to a person regardless of score, and compare. And look at what the score is actually keying on, by taking twenty low-scored examples and reading them, which takes an hour and is more informative than any dashboard.

The other failure mode

Speed. A qualification system that adds two minutes of processing before anything reaches a person has slowed your response, and response time is the one part of inbound handling where the advantage is undisputed. Design so that the fast path stays fast: acknowledge immediately, classify in the background, and never let a queue sit behind a step that can fail quietly.

The version of this we see most often is a system that batches. It runs every fifteen minutes, which is fine until the enquiry that arrives at 16:58 on Friday is processed at 17:13 and nobody is watching the queue it landed in.

When lead scoring is the wrong tool

If you receive a dozen enquiries a month, read them. The time it takes to build, tune and monitor a scoring system will never be repaid at that volume, and the general test for this applies unchanged.

If your problem is that most inbound is junk, the fix is upstream. Junk arriving in volume usually means the advertising, the targeting or the form is attracting the wrong people, and a better filter treats the symptom while you keep paying for the traffic.

And if the actual complaint is that replies are slow, you may not need scoring at all. An immediate acknowledgement, a link that lets someone book a time, and a rule that routes by intent will fix more than a model will. That is a week of work rather than a project, and it is often the honest recommendation.

What is worth automating here tends to be the boring middle: capturing everything in one place, summarising it, and getting it in front of a person quickly. That is the shape of most of our workflow automation work, and the rest of the workflow guides treat neighbouring processes the same way.

Frequently asked questions

Should a model decide which leads to reject outright?

Only for categories where a mistake is cheap and reversible, such as obvious vendor pitches. For anything that could be a customer, let the model route and let a person decide, because a wrongly rejected enquiry produces no signal at all: nobody complains, they just buy elsewhere.

What data do we need before we can score anything?

Two years of closed deals, or as much history as you have, with the outcome recorded. Without outcomes you are encoding opinions. If the history does not exist, start by recording it for a quarter while routing by simple intent rules.

Is enriching leads with public data allowed under the GDPR?

Enriching company information is generally straightforward. Building profiles of individuals is processing personal data and needs a basis, a record and transparency like anything else. Prefer company-level enrichment, and avoid quietly assembling profiles of named people.

How do we know the scoring is not filtering out good customers?

Send a random sample of low-scored leads to a salesperson anyway and track what happens to them. If that sample converts at a meaningful rate, your scoring is wrong, and no amount of internal confidence in the model outranks that result.