AI lead scoring for Canadian small business: what the score is allowed to see

Short answer: Lead scoring ranks your follow-up order using signals you already collect. For a Canadian business the constraint is which signals you may use: your own observed interactions are generally straightforward, third-party and inferred data are not, and Law 25 adds transparency obligations around automated decisions. Most small businesses also score far too early, with too few closed outcomes for any pattern to exist.

Lead scoring is usually explained as a technology question. For a small Canadian business it is mostly two other questions: whether you have enough history for a score to mean anything, and which inputs you are actually allowed to feed it.

Scoring too early is the common failure

A score is a pattern-finding exercise. Patterns require enough closed outcomes to exist.

The threshold is not fifty enquiries, it is roughly fifty to a hundred won and lost deals, because the model learns from what happened rather than what arrived. Below that it will still produce confident-looking scores, built on coincidence.

This is why many small businesses try scoring, see it rank leads badly, and conclude the approach does not work. What actually happened is that they asked for a pattern in a sample too small to contain one.

If you are below that threshold, the useful work is capturing outcomes properly so that scoring becomes possible later. Recording why each deal was won or lost is worth more now than any model.

What the score is allowed to see

This is the specifically Canadian part and it is worth getting right before building anything.

Generally straightforward. Information the person knowingly gave you, and behaviour you observed in your own interactions: what they enquired about, pages they viewed on your site, response speed, stated budget or timeline. This is your own relationship data, collected with their knowledge.

Harder. Data acquired from third parties, and characteristics inferred about the individual rather than observed. Both raise consent questions under PIPEDA, and appended data frequently arrives without any consent chain you can evidence.

Requires care. Under Quebec’s Law 25, decisions made using automated processing of personal information carry transparency obligations. There is a meaningful difference between a score that orders your callback list and a score that decides whether someone is offered a service or a price at all. The first is low-risk; the second sits closer to what the law is concerned with. Our guide to AI lead generation in Canada sets out the specific duty involved, a right for the person to be told and to ask for a human review, and the penalties behind it, up to C$25 million or 4 percent of worldwide turnover, which is the detail worth reading before building the second kind of score rather than the first.

None of this prohibits lead scoring. It shapes which inputs are worth using, and the answer conveniently overlaps with the inputs that actually predict well.

What predicts conversion in practice

The strongest predictors in most small businesses are behavioural and unglamorous:

Enquiry specificity. Someone describing a particular problem converts at a far higher rate than someone asking for general information.

Response speed to your first reply. How fast they come back is one of the most reliable signals available, and it costs nothing to capture.

A stated timeline. People who mention when they need something are usually further along than people who do not.

Source intent. A service-specific search converts better than a broad one or a social click, and this is recorded automatically.

Firmographic detail predicts less than owners expect. Demographic inference predicts little and carries the most compliance risk, which makes it the worst trade available.

Start with rules, not a model

Four or five behavioural signals weighted sensibly capture most of the available value, and they have three advantages over a model at small scale.

You can explain them, which matters if anyone ever asks how a decision was made. You can correct them when they are wrong. And they do not require the volume of closed outcomes a model needs.

Most CRMs used by Canadian small businesses already include scoring capability that nobody has configured. Auditing what you own before buying anything is the usual cheapest step, and it is usually skipped.

Move to a learned model when the rules become unmanageable and you have the outcome history to train on.

A working example: four signals and their weights

Take the four behavioural signals already covered in this guide and turn them into a simple points-based score.

Award two points for a specific enquiry rather than a general one: someone naming a make and model, a square footage, or a particular service beats “just looking into options.” Award two points for replying to your first message within an hour, since this is one of the cheapest and most reliable signals available and needs no judgement call to record. Award two points for a stated timeline, whether that is “next month” or “as soon as someone can start.” Award one point for arriving through a service-specific search rather than a broad one, a paid social click, or a directory listing, since source intent already correlates with the other three.

That produces a score from zero to seven. A lead at six or seven gets a phone call today, not an email queued for later. A lead at two or fewer goes into the normal automated sequence rather than being dropped.

None of this requires machine learning, a data scientist, or a tool beyond whatever spreadsheet or CRM field you already have. Its accuracy comes from checking the weights every few months against which scores actually closed, and adjusting the two-point and one-point signals if the pattern in your own numbers says otherwise.

Where scoring already lives in your CRM

Before buying anything, it is worth checking what the CRM you already pay for can already do.

HubSpot’s lead scoring tool builds both a fit score, from details like business type and size, and an engagement score, from interaction history, and can combine the two. It offers AI-assisted recommendations that analyse past conversions to suggest more precise scoring criteria, and every score is visible on the contact record alongside the interactions that produced it. The feature sits in premium editions of Marketing Hub rather than the free tier, so check which plan you are actually on before assuming it is unavailable.

Zoho CRM’s scoring rules work on a similar principle from the settings menu: assign positive and negative points to specific fields and touchpoints, including email engagement and survey responses, and let the points accumulate automatically as a record updates. Zoho also offers an option to let its Zia assistant build a model rather than setting the rules manually.

Neither tool decides which inputs are appropriate for a Canadian business to use. A CRM’s scoring engine will happily build a fit score from purchased or inferred demographic data, which is exactly the category of input flagged earlier in this guide as carrying the most compliance risk for the least predictive value. The software handles the mechanics; the judgement about what goes into it stays yours.

The failure that costs money

Individual mis-scored leads are not the problem. Systematic bias is.

A score trained on historical wins learns who you have sold to before. If your past customer base skewed toward a segment for reasons of history rather than fit, the score encodes that and steers effort away from segments you simply never pursued. The business then stops pursuing them, which confirms the pattern.

The mitigation is straightforward and rarely implemented: keep a portion of low-scored leads in the normal follow-up process, and check periodically whether they convert at rates the score did not predict. If they do, the score is describing your history rather than your market.

What to do with the score

The point is ordering effort, not filtering people out.

High-scoring leads get a call rather than an email, and get it faster. Low-scoring leads get the automated sequence rather than being discarded. Nobody should stop hearing from you because a model ranked them low, which is both commercially sensible and avoids the automated-decision territory that Law 25 is concerned with.

Consent still governs the messages that follow regardless of score. Our CASL compliance guide covers that side.

Frequently asked questions

How many leads do you need before scoring is worth doing? Enough closed outcomes for a pattern to exist, which in practice means at least fifty to a hundred won and lost deals rather than fifty enquiries. Below that the model is learning from noise and will confidently rank leads on coincidence. Most small businesses that try scoring too early conclude it does not work, when what happened is that they asked a pattern-finding tool to find a pattern in a sample too small to contain one.

Which signals can a Canadian business legally use in a lead score? Information the person knowingly provided or that you observe from your own interactions with them is generally straightforward under PIPEDA: what they enquired about, which pages they viewed on your site, how quickly they responded, what budget they stated. Where it gets harder is data acquired from third parties or inferred about the individual, which raises consent questions and, under Quebec’s Law 25, transparency obligations about automated decision-making. Take advice on your specific setup rather than assuming a vendor’s compliance claim covers you.

Does Law 25 restrict automated lead scoring? It does not prohibit it, and it does add transparency obligations around decisions made using automated processing of personal information. The practical position for a small business is that scoring which prioritises your own follow-up order is low-risk, while scoring that determines whether someone gets offered a service or a price at all sits closer to the kind of automated decision the law is concerned with. That distinction is worth getting advice on before you build the second kind.

What actually predicts whether a Canadian lead converts? In most small businesses the strongest predictors are unglamorous and behavioural: how specific the enquiry was, how quickly the person replied to your first response, whether they named a timeline, and whether they arrived from a high-intent source such as a service-specific search rather than a broad one. Firmographic detail like company size predicts far less than owners expect, and demographic inference predicts little while carrying the most compliance risk.

Can a lead score be wrong in a way that costs money? Yes, and the expensive failure is systematic rather than individual. A score trained on your historical wins learns who you have sold to before, which quietly encodes any bias in your past customer base and steers effort away from segments you simply never pursued. The mitigation is to keep a portion of low-scored leads in the normal follow-up process and check periodically whether they convert at rates the score did not predict.

Do I need a dedicated tool for this? Usually not at first. Most CRMs used by Canadian small businesses include scoring features that go unused, and a simple rules-based score built from four or five behavioural signals captures most of the available value. Buy a dedicated tool when you have enough volume that the rules become unmanageable and enough closed outcomes for a model to learn from, which is a later stage than most vendors suggest.

What does a simple point-based lead score look like in practice? A workable starting model adds points for a handful of behavioural signals rather than trying to capture everything: perhaps two points for a specific rather than general enquiry, two points for replying to your first message within an hour, two points for naming a timeline, and one point for arriving through a service-specific search rather than a broad one or a paid social click. A lead scoring six or more gets a call today; a lead scoring two or fewer goes into the automated follow-up sequence rather than being dropped. The exact weights matter less than reviewing them every few months against which scores actually closed.

Do HubSpot and Zoho already include lead scoring for a Canadian small business? Both do, in their paid tiers. HubSpot’s lead scoring tool builds fit and engagement scores from business details and interaction history, with AI-assisted recommendations, though the feature sits in premium editions of Marketing Hub rather than the free tier. Zoho CRM’s scoring rules let you assign positive and negative points to fields and touchpoints such as email engagement, with an option to let its Zia assistant build the model instead. Neither replaces the question of which inputs are safe to score under PIPEDA and Law 25, which is a policy decision no software vendor makes for you.

---

Want to know whether your enquiry data is ready for scoring? Get a free marketing audit.