There is a particular kind of message we get, and it is not the one people expect from a business asking about lead generation.
"The problem is to get the data, and convert that into leads, that's why I'm looking for a CRM expert in order to do that."
He was not short of enquiries. He was short of a way to tell which of them mattered. That distinction runs through most of what gets requested under the heading of lead generation: in our archive of 30,096 public briefs, 690 ask for lead generation and 176 of those also ask for a CRM — more than the 137 that mention advertising at all.
The bottleneck is usually sorting, not sourcing. Which is exactly the shape of problem an AI layer is good at, and also exactly the shape of problem where handing over too much of the decision goes badly.
AI lead qualification is three jobs, not one
"AI qualification" is sold as a single feature. It is three, and they carry very different risk.
| Job | What it does | Safe to automate? |
|---|---|---|
| Enrich | Fill in what the enquiry did not say — company, sector, size, language, source | Yes, fully |
| Score | Rate how well it fits what you sell | Only as a visible suggestion |
| Route | Send it to the right person, in the right channel, immediately | Yes, and this is where the value is |
Enriching is safe because it adds information and removes none. Routing is safe because a misrouted lead is still a lead — somebody just has to pass it along. Scoring is the one that deletes things, and scoring is the one everybody wants to automate first.
What the model should be allowed to do
Ask it to read the enquiry and produce structure. Not a verdict — structure.
- What is being asked for, in one line
- Which of your services it maps to, or none of them
- Whether the person named a timeline, a budget or neither
- What language they wrote in, and which they would prefer a reply in
- What is missing that somebody will have to ask
That last item is the underrated one. A qualification layer that surfaces the three things nobody asked turns a two-day back-and-forth into a single reply. In a market where the same buyer has messaged four suppliers, being the one who asks the right question first is most of the advantage.
What it should not be allowed to do
Decide who gets ignored.
The failure mode is quiet and it is expensive. A model trained on your historic wins will learn your historic biases, including the ones you would disown if they were written down. Small company, unfamiliar sector, imperfect English, an enquiry sent at 2am — all of these correlate with lower historic conversion in most datasets, and none of them is a reason to bin somebody.
You will never hear about this error. The customer who was scored low and left in a queue does not complain. They just buy elsewhere, and the dashboard shows a healthy qualification rate.
The rule we work to: the model can sort the queue, but nothing may leave the queue without a person, and every automated decision has to be visible and reversible in one click.
Where WhatsApp changes the calculation
In this market a large share of first contact arrives on WhatsApp, and the expectations attached to that channel are different from email. People who message a business on WhatsApp expect the pace of a conversation, not the pace of correspondence. Our archive shows what it sounds like when that expectation is missed:
"Hello, any update? It's almost 4 hours past."
Four hours is nothing on email and an eternity on WhatsApp. This is where an automation layer earns its cost honestly: not by pretending to be a person, but by acknowledging within seconds, asking the one question that unblocks the reply, and putting a structured summary in front of a human who can answer properly.
Two things matter in how you build that:
- Say it is automated. People are markedly more tolerant of a bot that admits it than one they catch out. It also keeps you the right side of platform rules.
- Never let it dead-end. Every automated conversation needs an obvious route to a person. A bot that loops when it does not understand converts worse than no bot.
Wiring up an automated lead qualification process
The order matters, and most implementations do it backwards by starting with the model.
- One inbox first. WhatsApp, web form, email and phone into a single queue. Without this you are automating five separate problems.
- Then enrichment. Cheap, safe, and immediately useful even with nobody reading the scores.
- Then routing rules, written as plain conditions before any model is involved. Many businesses find they never need the model — sector plus service plus language covers most of it.
- Then scoring, as a suggestion, displayed next to the lead and never acted on alone.
- Then review. Sample the low-scored leads every week and read them yourself. This is the step that gets dropped, and it is the only one that catches the expensive error.
What to measure
Not leads processed, and not hours saved. Those numbers go up whether or not the thing is working.
Measure time to first useful reply — not an acknowledgement, the first message that actually moves the conversation. Then measure what share of enquiries got one at all. A qualification layer that halves your response time and leaves nobody unanswered has paid for itself, regardless of how clever the scoring turned out to be.
And keep a list of the leads the system ranked lowest that went on to buy. If that list is empty after three months, either the system is working or nobody is checking. Both are worth knowing.