Lead Scoring and MQL Rules That Sales Will Trust

The rules are yours: the criteria, the test, the intake, the MQL rule. Free tools give a first draft of each step, alongside, and you check each one. The lead scoring model skill refuses to tier any lead with a missing or unlisted value. It is in review and not yet installable, so step 2 is followed by a connector-only route.

Fit comes before intent because intent on a poor-fit account is noise. The back-test comes before routing because sales uses a model it has seen tested on its own deals.

Step 1. Draft the fit criteria, and grade them yourself

About this example. TrueFoundry is a real company. TrueFoundry has not reviewed this guide. Public facts are linked alongside with the date I checked them. Every number, deal, lead and customer type marked illustrative is an input I chose to show how the tools work; it is not TrueFoundry's data, and the tool outputs are exercise results, not TrueFoundry's real strategy.

What is public: TrueFoundry raised a $19 million Series A in February 2025 and describes itself as "The Enterprise AI Gateway for LLMs, MCPs & Agents". Its pricing page shows three doors: a free Developer plan, a Pro plan with a "Start free trial" button, and an Enterprise plan through "Contact sales", which adds VPC and on-prem options. Every door produces leads, and the score has to read all three the same way.

Input, all illustrative: five criteria for "AI gateway and agent infrastructure for engineering teams", with the values in the table alongside listed best first, plus a success pattern: deals where a platform engineering lead runs the evaluation and VPC deployment is required close at about twice the ACV of public-cloud-only deals.

What came back: a maximum of 90, tiers at 80%, 60% and 40% of it, weights "preset by importance level", and every point labelled an example.

How to read it: I kept the criteria, the maximum and the tier bands, and rewrote two rows. Two criteria I meant as a scale came back as lists of wanted values: "VPC or on-prem required", "Hybrid" and "Public cloud only" all scored 25, and "Pilot only" scored the same 15 as "Yes, in production". A scale needs points that fall, so I set deployment at 25, 13 and 0, and production at 15, 8 and 0. I also set aside its pattern check, which suggested top points for "Public cloud only", the weaker case in my pattern.

Two limits. A public cloud buyer can reach 65 at most, tier B, so this score favours the Enterprise door; if public cloud deals matter, give them their own model. And my top employee band stops at 10,000: add a band above if you sell to large banks.

Step 2. Back-test the model on closed deals

Input: the same five criteria, graded best first, and ten illustrative closed deals, six won and four lost, entered as leads, with values alongside as the deals closed. The tallies are mine.

Two warnings. Closed deals carry fields reps filled during the sale: deployment, production and platform team are 50 of these 90 points, and a new lead rarely shows them. So score each deal only on what was known the day it became a lead. If your CRM keeps only the latest value, use the first form fill and dated enrichment, and leave the rest blank. And closed deals never meet the junk the score exists to stop, so add a sample of leads sales rejected and hold them to the loss limit in your stop rule.

Run 1, example default weights: five of six wins in A or B, three of four losses in C or D. Two deals looked wrong: Lost 4, tier B with no LLM apps in production, and Won 5, tier C with apps live.

Run 2 swapped two weights: employees to 15, 8 and 0, production to 25, 13 and 0. Neither deal changed tier, and Won 3 fell from 56 in B to 51 in C.

So Lost 4 gets a gate instead: my suggested rule, not the tool's, is that a lead whose production value is "No" is not an MQL at any score. "Pilot only" passes the gate, because Won 3 and Won 4 were pilots.

Judged on the rule sales will use, fit tier A or B with the gate on, run 1 passes five wins, run 2 four, and neither passes a loss. I keep run 1 and the gate, and log the rejected swap with its reason and date.

A stop rule has to be able to reject a change. Mine, illustrative: at least 80% of wins pass and at most 25% of losses and rejected leads pass under the MQL rule, on tuning deals and on held-back deals. Run 1 reaches 83% and 0%; run 2 reaches 67% and fails. I wrote that after seeing both runs, which is thin: write yours before the first run. After three changes that miss, stop and tell sales the fields do not separate wins from losses.

My ten deals break the first warning, and here is the cost. Blank production and platform team, as a new lead would, and the skill holds all ten. The best case at first touch is three fields: deployment from a form, employees and industry from enrichment. On those alone, maximum 65 with tier B from 39, the skill passes Won 1, 2, 4 and 6, and also Lost 2 and Lost 4, which fails my stop rule. The gate cannot help, because nobody yet has the field it reads. So the 0 of 4 shows only that production, once known, separates these ten deals. It does not show a score stopping a loss on arrival. Hence two stages, in step 3.

Do not trust those numbers yet. Ten deals tuned by hand fit themselves: one win moves the rate 17 points, I wrote the gate after seeing Lost 4, and these deals carry no intent data, so the intent floor in step 4 is untested. Export as many deals as you can, set the newest third aside before you look, and score them once at the end. This exercise has no held-back deals or rejected leads to show. If yours fail, do not retune on them: run the score as guidance for a quarter and retest on new deals. At 30 deals and an illustrative 25% win rate, the newest third holds about two or three wins, so 80% means every one: a sanity check, not a measurement. My floor, illustrative: under 30 closed deals, skip the score and use three yes or no questions from your top criteria, plus a rep's read. Count value too: a rule that misses your largest win is worse than its rate suggests.

This is the point where I would want a second pair of eyes: the held-back run fails before sales adopts the MQL rule. Bring me your first-touch export, held-back outcomes, sample size, intent window and sales acceptance decision, and we will work it through in 45 minutes.

Step 3. Score this week's leads, and hold back the unknowns

Most holds are an intake problem, so build intake in two stages. At first touch, ask deployment in one dropdown at all three doors, and fill employees and industry by enrichment. That three-field score orders the work and rejects nobody: Won 3 and Won 5 sit in tiers C and D on it. On the first reply, a rep asks production and platform team; trials and signups get the same two questions in onboarding. No lead is an MQL until production is known, because the gate reads it. Keep a mapping table for the words people write, so "On-prem" becomes "VPC or on-prem required" before the script sees it.

Input, before any mapping: the run 1 model and the six illustrative leads alongside, two from each door, as they stand after the first reply. A contact sales lead wrote "On-prem", which is not exactly on the list, so the skill returned "pending your confirmation, no tier". A Pro trial lead had "N/A" for employees, so it came back "incomplete, no tier".

My suggested rule: pending and incomplete leads go to a research task with each open item named, then back through the script. At signup volume, research only leads that can still reach tier B; send the rest to nurture.

A hold stops the MQL label, never the reply. Mine, illustrative: anyone who asks for sales hears from a rep the same day, in any tier. Other holds get five working days, then go to nurture with open items logged. Lead 4 shows why: a regulated firm that asked for sales, held on one word. Mapped, it has 73 points and one blank, tier A whichever way the blank goes. My rule: if every answer lands in the same tier, clear it, unless the blank is production, which the gate needs.

Step 4. Apply the MQL rule and rank the accounts

The MQL rule, illustrative and my own: fit tier A or B on the back-tested five-field model, no gate triggered, and either intent of 50 or more or a contact sales request. A tier A or B buyer who asks for sales should not wait for a second signal. A request in tier C or D still gets its reply, but is not an MQL. Intent points, illustrative and mine, over the last 14 days: 40 for a Pro trial started or a contact sales request, 30 for gateway requests above half the plan's monthly allowance, and 10 for each other sign, up to 100.

Three illustrative accounts passed: Lead 1, Lead 6 and Lead 5, once a rep confirmed its size band as 250 to 999, which scored 78 of 90, tier A. Fit for the ranking tool is the model score out of 100: 68 of 90 is 76. Relationship and timing are the account owner's calls, and a sales request is always now. Lead 4 joins once its mapping is confirmed. Before ranking, merge people at one firm into one account, and drop customers and open deals.

Input to the ranking tool: company TrueFoundry, the category quoted in step 1, and the accounts alongside, in that order. What came back: the lending fintech ranked first with 73. All three landed in the ranking tool's tier B, a separate scale from the fit tiers. Two printed 66. Treat that as a tie: reversing my list reversed them.

How to read it: I keep the order and drop the action lines, which suggest a warm introduction even for a lead who already asked to talk. A blend is safe here only because these accounts already passed. So send each account's fit tier, intent points and signal with it.

Lead 2 is the price of run 1: intent 90 but fit tier C, so it failed the rule. On run 2 it scored 56 and passed. Ten deals cannot settle that, so I send it to a rep as a manual review, marked not an MQL.

There is a fourth door: product usage. A head of platform engineering on the free plan can show heavy use with fields nobody filled. So usage above half the plan's allowance from a work domain opens a rep's review. It scores only 30, so if the account passes on fit, the rep can hand it over as a sales request would.

One more rule, my own: sales accepts or rejects each MQL within an agreed window, with a reason. Record lead-to-opportunity rate before launch as your baseline, start the accept rate on day one, and read both by fit tier monthly. Save each lead's values as they arrive, so your next back-test is true first touch. Accept rate cannot see good leads the rule rejected, so each month a rep reads 20 of them. If manual reviews like Lead 2 keep getting accepted, reopen run 2.

1. Letting a blend of fit and engagement decide who is an MQL.
2. Shipping weights nobody back-tested, or calling ten deals proof.
3. Scoring a blank as zero.
4. Letting the entry route decide fit.
5. Reading tie order as a ranking.

• How to Size Your B2B Market Bottom-Up and Score Target Accounts Without Inflated Analyst Numbers
• Account-Based Plays Without Burning SDR Hours: Account Research, an Account Plan, Category Traps, Lookalikes and Roundtable Follow-Up
• Outbound Sequences That Read Like a Person Wrote Them
• All 24 guides