Rule-Based Scoring vs. AI: Why Simple Wins for Early-Stage SaaS
You don't need machine learning to detect user intent. A 7-point scoring model based on 5 behavioral events outperforms complex AI systems for SaaS companies under 10K users. Here's why.
There's a temptation, when building anything related to "user intent detection," to reach for machine learning. Train a model on historical data. Deploy a real-time inference pipeline. Score every session.
For a SaaS company with less than 10,000 users, this is almost always the wrong approach.
The data problem
Machine learning models need training data. Lots of it. To build a reliable churn prediction model, you need at least 1,000 conversion events and 1,000 non-conversion events, ideally with 6+ months of behavioral history. Most early-stage SaaS companies don't have this. They have 200 users, 3 months of data, and 15 paid conversions. Train a model on that and you'll get noise, not signal.
How rule-based scoring works
Rule-based scoring works with any volume. The logic is simple:
- Visited pricing page — +3 points
- Started checkout — +5 points
- Signed up — +2 points
- Started onboarding — +5 points
- Returned after 24+ hours — +2 points
- Threshold: score >= 7 = high intent
When a user crosses the threshold, flag them for recovery. That's it.
Three advantages over ML
1. It's explainable
When a recovery email goes out, you can say exactly why: "This user visited pricing twice and started checkout." With an ML model, you get a probability score with no clear explanation, which makes it hard to write a relevant recovery email.
2. It's tunable in real-time
If you notice that users who visit your pricing page three times convert at a higher rate, you add a bonus score for repeated visits. With ML, you'd need to retrain, validate, and redeploy. With rules, you edit one number.
3. It doesn't hallucinate
ML models can develop blind spots — correlations in training data that don't reflect real behavior. A model might learn that users who visit your blog about API docs are likely to convert, not because the blog indicates intent, but because your Product Hunt launch drove both blog traffic and signups simultaneously. Rules don't have this problem because they're based on direct behavioral indicators.
The scoring model that works
For most SaaS products, start with five events:
- user_signed_up — account creation
- visited_pricing — evaluation signal
- started_checkout — strongest buying signal
- onboarding_started — engagement signal
- payment_success — conversion (stops recovery)
Then add timing windows:
- Checkout drop — trigger 10-15 min after abandoned checkout
- Activation drop — trigger 24h after signup without key action
- Inactive user — trigger 3 days after high-intent user goes silent
Anti-spam safeguards matter
The safeguards are as important as the scoring:
- Never send more than one recovery email per user per scenario within 48 hours
- Cap retries at 3 total
- Respect unsubscribe requests immediately
- Check for conversion events before sending
These rules prevent the system from becoming annoying — which is the fastest way to destroy its effectiveness.
When ML actually makes sense
Once you have 10,000+ users and 6+ months of conversion data, ML can find patterns that rules miss. It can optimize send timing, predict which template works best for a given user, and identify intent signals you hadn't considered.
But that's an optimization on top of a working system, not a replacement for one.
The bottom line
Simple systems that ship beat complex systems that don't. Every recovered checkout you capture with a rule-based scorer in month 1 is revenue that would have been lost while you were training your first ML model.
Start with rules. Run for 3 months. Look at the data. Then decide if you need ML. In most cases, you won't.