Field notes
Deterministic prospect scoring beats AI guesswork for recruitment BD
6 July 2026
On this page
A recruitment founder I know spent three months running her BD list through an AI tool that promised to surface "high-intent" prospects. The output looked credible. Named contacts, seniority levels, a confidence score next to each row. She worked the list. Conversion was worse than her gut instinct from the previous year.
The problem was not the data. The problem was that nobody could explain what the confidence score meant. It was not calibrated against her close rates, her sector, or her average deal size. It was a vibe, dressed up as a number.
This post is about why deterministic scoring, a rubric you build and can audit, outperforms LLM-generated scores for recruitment BD, and how to build one that actually maps to revenue.
Deterministic scoring means you assign explicit point values to observable signals, sum them, and rank your prospect list. The rules are visible. You can change them. You can trace why a prospect scored 74 and not 52.
AI-generated scoring (as most BD tools implement it today) means a model produces a number or a tier based on patterns it has learned, usually from training data that has nothing to do with your firm's close rates. The score is a black box. When it is wrong, you cannot tell why, and you cannot fix it without going back to the vendor.
The choice matters now because the market for AI-enriched prospect tools is expanding fast. Apollo, Cognism, and a growing number of recruitment-specific platforms are bolting on "AI scoring" as a feature. Some of it is useful enrichment. Some of it is a confidence score with no denominator. Knowing which is which determines whether your BD team works a good list or a decorated bad one.
When a deterministic score is wrong, you know exactly where to look. If a prospect scored 80 and went cold after first contact, you can open the rubric, see which signals inflated the score, and recalibrate. That is a feedback loop. You are building institutional knowledge about your market.
An AI score that is wrong gives you nothing to work with. You can flag it to the vendor. You can ignore it. You cannot learn from it in any structured way.
The only scoring system worth using is one calibrated against your own historical data. What percentage of prospects in a given segment actually converted to a retained or contingency mandate? What average fee did they generate? Those two numbers, probability and revenue, are the foundation of a useful score.
An LLM does not have that data. It has patterns from a training corpus. Those patterns may correlate loosely with intent, but they are not your close rates.
In recruitment BD, seniority of the contact is a hard filter, not a soft signal. A warm intro to an HR coordinator at a 200-person firm is not the same as a cold email to the CFO. A deterministic model can enforce a floor: prospects below a certain seniority level do not score above a threshold regardless of other signals. That rule is explicit and enforceable.
An AI model may weight seniority, but you cannot see the weight, and you cannot enforce a floor without building a separate filter on top of the model. At that point you are doing the deterministic work anyway.
Your market changes. A sector that was hiring aggressively in Q1 may be in a freeze by Q3. A deterministic rubric takes an afternoon to update. You change the weight on "active job postings in the last 30 days" and re-run the score. Done.
Retraining or reconfiguring an AI scoring model requires either vendor involvement or a level of ML infrastructure most recruitment agencies do not have.
Here is a rubric built around three dimensions: segment probability, revenue potential, and signal bonuses. The numbers below are illustrative. You calibrate them against your own data.
| Signal | Points | Notes |
|---|---|---|
| Segment close rate 20%+ | 30 | Based on your last 24 months of mandates |
| Segment close rate 10-19% | 15 | |
| Segment close rate below 10% | 0 | Exclude or deprioritise |
| Average fee over £15k | 20 | Revenue potential floor |
| Average fee £8k-£15k | 10 | |
| Average fee below £8k | 0 | Below your margin threshold |
| Contact is C-suite or Director | 20 | Seniority floor enforced here |
| Contact is Head of / VP | 10 | |
| Contact below Head of | 0 | Does not qualify regardless of other signals |
| Active job postings (3+ in 30 days) | 15 | Hiring signal from LinkedIn or job board scrape |
| Recent funding round (Series A+) | 10 | Growth signal |
| Previous relationship (met at event, referral) | 10 | Warm signal |
| Company headcount growth 10%+ YoY | 5 | Expansion signal |
Maximum score: 110. A prospect scoring 70+ goes into the active outreach queue. 40-69 goes into a nurture sequence. Below 40, park it.
The scoring logic in your CRM or outreach tool should be transparent enough that any BD lead can open it and explain why a prospect sits where it does. If they cannot, the model is not doing its job.
I am not arguing against AI in BD workflows. I am arguing against using AI to replace the scoring rubric.
AI is genuinely useful for enrichment: pulling in signals you would not manually track, like funding data or new job postings. Tools like Cognism or Apollo do this well. Feed those enriched signals into your deterministic rubric as inputs. That is a sensible division of labour.
AI is also useful for drafting outreach once a prospect has qualified. You have a scored, segmented list and need personalised first-contact messages at volume. That is a reasonable prompt workflow. The scoring that determined who gets the message should still be yours.
The brief bottleneck problem in agency workflows is a useful parallel here: AI accelerates the output stage, but if the intake logic is broken, the acceleration makes things worse. Prospect scoring is intake logic for your BD pipeline. Get it right before you automate the outreach.
Building the rubric is step one. Closing the loop is where most agencies fall short.
Every quarter, pull your won mandates and map them back to their entry score. Did the 80+ prospects convert at a higher rate than the 50-69 band? If not, which signals are over-weighted? Did the seniority floor hold, or did you win mandates from contacts who should have been filtered out?
This is not complicated analysis. It is a pivot table and an honest conversation with your BD lead. But it is the work that turns a rubric into a calibrated model, and a calibrated model into a durable advantage.
AI tools do not do this for you. They retrain on their own data, not yours. The feedback loop has to be yours.
If you want to understand why AI pilots in recruitment tend to degrade quietly rather than fail loudly, the reasons most AI pilots fail comes down to exactly this: the feedback loop was never built.
Pick deterministic scoring when you have at least 12 months of closed mandate data to calibrate against, or when your BD team needs to explain and defend their pipeline to a founder or investor.
Pick AI scoring when you have no historical data at all and need a starting point, or when the vendor can show you the features the model is weighting and let you override them. That last condition rules out most of what is currently on the market.
The default position for a recruitment agency doing serious BD is a deterministic rubric, calibrated quarterly, with AI enrichment feeding the signals in. That is a system you own. Everything else is renting someone else's black box and hoping it knows your market better than you do.
If your BD workflow is producing a list nobody fully trusts, that is a process problem before it is a tooling problem. The AI Workflow Audit is where we start: mapping what your current scoring logic actually is, where the gaps are, and what a calibrated rubric would look like for your sector and deal sizes.