Short answer: the lead scoring built into the big CRMs learns from your past deals, and it needs a lot of them. Salesforce asks for 1,000 new leads in 200 days, with 120 converted, before it builds your own model. Many 10–200 person companies in India don't have that history in one clean place. You can still score leads with AI: write down what a good lead looks like, have a language model read each enquiry against that page, and check the scores against real outcomes every month.
What the CRMs actually need
Predictive scoring sounds like a switch you flip. Read the setup pages and it's closer to an entry exam.
Salesforce's Sales Cloud Einstein setup guide (Winter '27 edition) says Einstein Lead Scoring needs at least 1,000 leads created in the last 200 days, and at least 120 of those converted to an account and contact. Below that, if you score all leads together, it falls back to a global model built from anonymous data across many Salesforce customers. It isn't trained on how your own buyers behave.
Zoho's help page for Zia Scores sets a lower bar. It's available on the Professional edition and above, needs at least 200 records to start scoring, and wants at least 75 examples each of what you call ideal and non-ideal records. It scores out of 100 and sorts the result into three bands: 0–50, 51–75 and 76–100.
HubSpot's predictive lead scoring is only on the Marketing Hub and Sales Hub Enterprise plans. It gives each contact a percentage chance of becoming a customer in the next 90 days, then splits contacts into four equal tiers. HubSpot also says the model is a black box: you can see what goes in and what comes out, but not how each input moved a contact's score.
Look at what HubSpot feeds into that model: page views, website visits, email opens and clicks, form submissions, logged notes, days since a meeting was last booked, whether the contact has a phone number, and whether their email is on a free domain like Gmail. Those are good signals for a company whose buyers fill in web forms and read nurture emails.
Why this misses many Indian sales teams
Think about where your enquiries actually come from. A lot of them arrive as a WhatsApp message, a missed call, a marketplace enquiry or a walk-in. The detail that tells you whether the lead is serious ("we need 40 units before March", "can you deliver to Hosur?") sits in a chat thread or in the salesperson's head, not in a CRM field.
A model that learns from CRM activity can only score what's in the CRM. HubSpot doesn't say which way the Gmail input pushes a score, but where plenty of real business owners run their company on a Gmail address, the email domain says little about intent.
For a small team, the useful starting point is simpler than picking the CRM with the smartest scoring: can you describe a good lead in plain words?
Score leads from a one-page rubric
This is what works without a data team.
1. Look at your last few months of closed deals. Sit with your best salesperson and go through the deals you won and the ones that died. Note what the good ones had in common when they first got in touch. Usually it's a handful of things: the kind of business, size, location you can serve, a named need, a timeline, and whether they asked about price or a site visit.
2. Write it down as a short rubric. Split it into fit (could this person buy from us at all?) and intent (are they trying to buy now?). Add the things that disqualify a lead, such as a location you don't serve or a budget far below your smallest job.
3. Have an AI model read each enquiry against the rubric. Give it the form entry, the WhatsApp message or the call note, plus your rubric. Ask for three things back: hot, warm or cold; one line on why; and what's missing ("didn't mention quantity"). Three labels beat a 0–100 number here. Your team will trust a label with a reason more than a figure they can't explain.
4. Put the result where the team already works. The score and the reason should show up next to the lead in your CRM or shared sheet, and in the alert your salesperson gets. If they have to open another tool, they won't.
5. Check it every month. Pull the leads that closed or died and compare them with the label each got on day one. If won deals keep showing up as cold, the rubric is wrong, so fix the line that misled it. This review is the part teams often skip, and it's the part that makes the scores worth anything.
If you want help wiring this into the tools you already use, that's the kind of thing our sales pipeline automation work covers.
Where it goes wrong
The model can only score what's written. A message that just says "price?" will look cold, and some of those people are ready to buy. Never let a low score stop a first reply; use it to decide who gets called first.
Rubrics also drift. Your best customers this year may not look like last year's. Keep each version of the rubric with a date, so when the monthly check shows a change you know which edit caused it.
And keep a person in charge of it. Someone in sales should own the rubric, read a sample of the reasons each week, and have the final say on any lead that matters.
When to move to the CRM's own scoring
Once you've run the rubric for a while, you'll have something most small teams lack: months of leads with a label, a reason and an outcome, all in one place. That's the history predictive scoring needs. When your volumes clear the bar your CRM sets, switch on its model and run both side by side for a month before you trust it.
Arthat AI
Working on something similar at your company?
Arthat AI builds these automations for Indian businesses every week. If you want to talk through whether this fits your situation, we'll give you a straight answer — even if the answer is "don't build it."
