Scoring

What makes a lead worth ringing first, in points you decide.

1Overview #

Every open lead and deal carries a score out of 100 and a band word - Hot, Warm or Cold. This screen decides what that score is made of.

A score is never a black box. It is a starting 20 points, plus every rule that fired, minus what going quiet costs - and each of those is kept as a sentence, so a record can always show exactly why it reads what it does.

A rule is three things: when it applies (a condition, in the same language as assignment rules and approval workflows), how many points, and the words the explanation will show. The third is not decoration: it is what an executive reads.

Later, once there is enough finished history, a model can be fitted on the company's own won and lost leads. It is fitted here, judged here, and only made live here - by a person, never by itself.

2Who it is for #

An administrator or a sales manager deciding what makes a lead promising here.
Anybody reviewing whether the scores are telling the truth - the Records column is where that shows.

3What it is not for #

Reading one record's score - that is the Score panel on the lead or deal itself.
Setting up the AI - that is Settings > AI Assistance.
Changing who may see which records - that is CRM > Teams and sharing.

4Step by step #

Choose leads or deals. - They have their own rules and their own bands.
Read the rules that are there and change what does not fit. - The seeded set is a starting point, meant to be argued with.
Watch the Records column. - A rule that has never once fired is either wrong, or about something that does not happen here.
Set where Hot, Warm and Cold begin, and save the bands. - The lowest must start at 0, so every score lands in one.
Set how scoring behaves, and save. - When the nightly pass runs, and how long silence is allowed.
Much later, press Fit a model now. - Only once there are enough finished leads. Read what it did before making it live.

5The fields #

A rule

When this is true, these points, for this reason.

Name - What you would call it in a meeting. Required Nothing but the list. Two rules of the same kind may not share a name.
About - Who they are, or what they have done. Required How the rules are grouped when read. It changes no arithmetic.
Points - What the rule is worth, either way. Required The score, directly. -100 to 100, and not 0 - a rule worth nothing does nothing.
The reason it gives - The words shown on the record when this rule fires. Required What an executive reads. Write it as a reason: "they have replied to us", not "replies rule". 200 characters.
In use - Whether it counts. Optional Switched off it stops counting at once, not merely stops being listed.
When this is true - The condition, built from the record's own facts. Optional Blank means it fires on every record of that kind. The facts offered are the ones that kind actually has.

A band

Where Hot, Warm and Cold begin. They are saved together, because they only mean anything as a set.

Band - What this band is called. Required The WORD shown beside every score in it. 20 characters.
Starts at - The lowest score in this band. Required Which band a score falls in. 0 to 100. The lowest band must start at 0.
Look - How the badge is coloured. Optional Colour only - the band's word is always shown as well, so colour is never the only signal.

How scoring behaves

Settings 277-283. They apply to the whole company.

Work scores out again at - The hour of the nightly pass over every open record. Optional When time-only changes appear - chiefly a record going quiet. 0 to 23. Setting 277.
Days of silence before it counts against a record - How long nothing may happen before points come off. Optional Every open record's score, and the suggestions. 1 to 365. Setting 279.
Work it out at once when something happens - Whether a reply, a call, an activity or a stage change rescores within moments. Optional How fresh a score is between nightly passes. Setting 278.
Chance of winning is worked out from - The stage alone, or the stage adjusted by age and activity. Optional The win probability on deals only. Setting 283.
Keep the score history for - How long the one-row-a-day history is kept. Optional The shape over time on each record, and what a future model can learn from. 30 to 5000 days. Setting 282.

The model

Fitted on this company's own finished leads. Nothing leaves the building.

What the model may do - Off, Beside, or Live. Optional Whether the model's number is shown, and whether it decides the score. Setting 280.
Finished leads needed before one is fitted - How much history a model needs. Optional Whether fitting is attempted at all. 50 at the very least; 500 is the sensible figure. Setting 281.
How it did - What it scored on the fifth of the leads it was never shown. Read-only Nothing - it is there to be judged. Below about 0.65 it is barely better than guessing.
State - Fitted, Live or Retired. Read-only Only a Live model is used, and only when the mode above is Live too.

6Rules and integrity #

A score is never shown without its reasons - Every point is kept as a sentence against the record. If the reasons and the number ever disagree, the reasons are right and the arithmetic is wrong.
Every record starts at 20 - Not nought, and not fifty. A record nobody knows anything about is not worthless, and it is certainly not average.
Going quiet is priced once, by the engine - No rule needs to say it. The engine takes points off once a record passes the days-of-silence setting, a little more each week, to a limit of 20 - and a rule saying the same thing would charge it twice.
Changing anything here requeues every open record - The message says how many. The worker works through them within a couple of minutes; nothing is scored by this page itself.
A rule is checked against the facts of its own record kind - A lead rule may not name a deal's fields. It would never fire, and a score quietly wrong for six months is the worst way to find that out.
A model is fitted, never live - It arrives as Fitted. Somebody reads what it did and makes it live. Even then, setting 280 must say Live for it to decide anything.
The model is judged on leads it never saw - A fifth are held back - the same fifth every time, chosen by the lead's own number - and the accuracy shown is what it scored on those.
A rule score and a model score are different kinds of number - Points added up, against an estimated chance of winning. A 60 from the model is a strong lead where a 60 from the rules is middling, so the bands need looking at when the mode becomes Live.

7Shortcut keys #

F5 - Save the settings, or the rule being edited.
Esc - Close the rule editor, or leave the page.

8If something does not work #

"That record is not yours to see."

Why: The record belongs to somebody else's circle - its owner, their team lead, a sales manager, or somebody it is shared with. Reading it, its calls, its score or its telephone numbers is refused to anybody else, whichever screen asks.
What to do: Ask its owner to share it with you, or ask a sales manager to hand it over.

"There is nothing to save."

Why: No rule, band or weight was changed.
What to do: Change something first, or close.

"There is no such record to score."

Why: Recompute was asked for a record that does not exist, or of a kind scoring does not cover.
What to do: Refresh the record and try again.

"Scoring is not licensed for this company - lead scores, win probability, suggested next actions and the fitted model need the CRM Scoring add-on."

Why: The CRMSCORE add-on is not on this company's licence.
What to do: Ask your vendor, then sign out and in.

"Give the rule a name - what you would call it in a meeting."

Why: The name is blank.
What to do: Name it. Rules are named so they can be talked about.

"The name is too long - 100 characters at most."

Why: More than 100 characters.
What to do: Shorten it; the reason field is where the sentence belongs.

"Say whether this rule is about leads or deals."

Why: The record kind is missing - only possible if something was sent by hand.
What to do: Reopen the rule from the list.

"Say whether this is about who they are, or about what they have done."

Why: The About field is blank.
What to do: Choose one. It only groups the rules for reading.

"A rule worth no points does nothing. Give it points, either way."

Why: Points are zero.
What to do: Give it points, positive or negative.

"Points must be between -100 and 100."

Why: Outside the range a single rule may move a score.
What to do: Use a smaller figure. A rule worth more than the whole scale is not a rule, it is a verdict.

"Write the reason this rule gives ..."

Why: The reason is blank. It is what the explanation shows, so a rule cannot exist without one.
What to do: Write it as a reason, as it will read on the record.

"The reason is too long - 200 characters at most."

Why: More than 200 characters.
What to do: Shorten it - it appears in a list of other reasons.

"There is already a lead rule called ... " / "... deal rule called ..."

Why: Another rule of the same kind has that name.
What to do: Use a different name, or change the existing rule.

"That rule is no longer there."

Why: Somebody else deleted it while this page was open.
What to do: Refresh. Nothing was lost but the edit on screen.

A complaint about the condition - an unknown field, a missing bracket, an operator that does not fit the field

Why: The condition does not read. The message names the part that does not.
What to do: Fix that part. The field list in the builder is the whole of what this record kind offers.

"There must be at least one band - a score with no band is a number nobody can act on."

Why: Every band was removed.
What to do: Add at least one, starting at 0.

"The lowest band must start at 0, or a score below every band would belong to none."

Why: The lowest band starts above zero.
What to do: Set the lowest one to 0.

"Two bands share a name." / "Two bands start at the same score."

Why: Bands must be distinguishable, by name and by where they begin.
What to do: Change one of the two.

"... is too long a band name - 20 characters at most."

Why: A band name over 20 characters.
What to do: Shorten it - it is shown beside a number in a small badge.

"... must start between 0 and 100."

Why: A band starts outside the scale.
What to do: Use 0 to 100.

"Six bands is as many as anybody can hold in their head. Use fewer."

Why: More than six bands.
What to do: Combine some. Three is the usual number.

"The hour to recompute at must be between 0 and 23."

Why: An hour outside the clock.
What to do: Use 0 to 23, on the server's own clock.

"Days of silence must be between 1 and 365."

Why: Outside the range.
What to do: Use a figure in days.

"Learned scoring must be Off, Beside or Live."

Why: An unknown mode - only possible if something was sent by hand.
What to do: Choose one of the three.

"A model fitted on fewer than fifty finished leads is guesswork ..."

Why: The minimum was set below fifty.
What to do: Fifty is the floor and five hundred is the sensible figure.

"Score history must be kept between 30 and 5000 days."

Why: Outside the range.
What to do: A year and a bit is the sensible floor.

"Choose how win probability is worked out."

Why: The basis is blank.
What to do: Choose the stage alone, or the stage with age and activity.

"AI assistance is switched on but nothing is allowed. Tick at least one, or switch it off."

Why: Setting 284 is on while setting 285 lists nothing.
What to do: Tick at least one task on Settings > AI Assistance, or switch AI assistance off.

"There are N finished leads. A model is fitted at M, so ... more to go."

Why: Not enough history yet. This is an answer, not a failure.
What to do: Carry on with the rule score, which is no bad thing. Try again when the history has grown.

"There are X won and Y lost. A model needs at least twenty of each to learn the difference between them."

Why: Plenty of finished leads, but nearly all one outcome.
What to do: Nothing to do but wait, or check that leads are actually being marked Disqualified rather than left open.

"There is not enough to hold any back for a fair test. Nothing was fitted."

Why: Too few rows to split into a training part and a held-back part.
What to do: Wait for more history.

"Only N of those leads could be read back. Nothing was fitted."

Why: The features could not be gathered for most of the leads - usually a lead date or a status that is not what it seems.
What to do: Report it: the figures behind the leads need looking at.

"It is barely better than guessing. Do not make this one live ..."

Why: The model was fitted but cannot separate won from lost.
What to do: Leave it Fitted. Either the history is too thin, or what decides a sale here is not in the figures the CRM keeps.

"Note that learned scoring is set to Off/Beside, so this model is carried beside the rule score without deciding it."

Why: A model was made live while the mode says otherwise. Not an error - a reminder that two switches are involved.
What to do: Set What the model may do to Live as well, if that is what was meant.

9For an administrator #

What the server does - CrmScoreWorker runs every two minutes: it drains the queue of records something has happened to, does one nightly pass over every open record at the hour in setting 277, and re-fits a model weekly when setting 280 is not Off.
If the hour passes while the server is off - The pass still happens when it comes back - the guard asks whether a pass has run since the hour struck today, not whether a timer fired.

10Settings it uses #

277 - the hour of the nightly pass.
278 - recompute at once when something happens.
279 - days of silence before it counts against a record.
280 - what the learned model may do: Off, Beside or Live.
281 - finished leads needed before a model is fitted.
282 - how long the score history is kept.
283 - how win probability is worked out.

This is the help InvoiceNET® ERP shows beside the Scoring screen. Whether you have the screen depends on the modules of your licence and the rights your administrator has given you.

See this screen for yourself

The demo is the real product, with a company of your own.

Try the demo Getting started