Skip to content

Evaluation Rubrics

Evaluation rubrics define the criteria used to score calls in COPC Evaluation. Each rubric contains a list of weighted criteria that calls are evaluated against.

You do not have to rewrite your quality form before you start. The usual path is:

  1. Import the form your supervisors already use — Excel, Word, PDF, CSV or pasted text — or start from a template
  2. Go through the review: it points out wording that would make the AI score inconsistently and suggests a fix for each
  3. Calibrate: score a few calls your supervisors already scored, and see where the AI agrees with them before you go live

Click Settings in the left navigation, then open the Evaluation Rubrics (COPC) tab.


  1. Click New Rubric (or import your rubric, or start from a template)
  2. Fill in the rubric details:
    • Name — A descriptive name (e.g. “Customer Service Evaluation Q2”)
    • Description — Optional note for your own team about when to use this rubric. It is not shown to the evaluator and does not affect scoring
  3. Add evaluation criteria (see below)
  4. Click Save

Each criterion in a rubric has the following fields:

FieldRequiredDescription
IDYesUnique identifier (e.g. greeting, data_verification)
NameYesDisplay name (e.g. “Greeting Protocol”)
DescriptionNoWhat the evaluator should look for
Max PointsYesMaximum score for this criterion (must be > 0)
SectionNoGroup label (e.g. “Opening”, “Body”, “Closing”)
Critical ErrorNoIf checked, scoring 0 on this criterion fails the entire call
Goal of the callNoMarks the criterion that records the call’s outcome. Its result fills Goal reached in the reports — see The goal of the call
Excused if the customer hangs up before the farewellNoFor closing-script steps. Not docked on calls the customer left before the agent’s goodbye — see Closing steps when the customer hangs up

Critical errors: When a criterion marked as critical receives a score of 0, the entire call automatically scores 0% regardless of other criterion scores. Use this for non-negotiable requirements like identity verification or compliance disclosures.


Most criteria score what the agent did. The goal of the call says what the call achieved: a sale closed, a payment promised, an issue resolved. Tick Goal of the call on the one criterion that records that outcome, and every call scored on the rubric gets a Goal reached answer — no extra analysis needed.

A criterion marked as the goal of the call

The goal criterion…Goal reached
scored full pointsYes
scored less than full pointsNo
did not apply to the callunknown

Goal reached appears in the agent performance table, in the exemplary calls of the Call Center Report — where calls that reached their goal are listed first — and in the PDF export.

Writing a good goal criterion:

  • Describe the outcome, not the effort. “The customer accepts the offer” is a goal; “the agent attempts to close” is a behavior and belongs in an ordinary criterion.
  • Make it all-or-nothing. Anything below full points counts as not reached, so partial credit only blurs the answer.
  • Keep its weight low. The goal criterion counts toward the call’s score like any other. A customer who says no lowers the agent’s score by the criterion’s points, so a small weight — 5 points out of 100, say — keeps the outcome from drowning out the quality of the work.

Rules:

  • One per rubric. Ticking it on a second criterion moves it there.
  • Never critical. A goal criterion cannot also be a critical error, because a refused sale would fail the whole call. Ticking one clears the other.
  • It applies straight away. Ticking, moving or clearing the goal updates Goal reached on every call already scored with this rubric, without re-analyzing anything. The scores themselves do not change.

The evaluator already skips the steps a call never reached when the transcript visibly breaks off — see Calls that end early. The commonest real hang-up does not look like that. The customer says “okay, thanks” and hangs up while the agent is still closing. Nothing in the transcript says who ended the call, and a step like the satisfaction survey or “was everything clear?” can always be pinned on some earlier turn where the agent could have done it — so it gets docked.

If your dialer records who hung up first, two steps fix this:

  1. Send who hung up first with each call, as ended_by — see Who hung up first
  2. Tick Excused if the customer hangs up before the farewell on each closing-script step

On every call where the customer hung up first and the agent never said goodbye, each ticked criterion that lost points is given its full points and marked N/A, with a Customer hung up badge beside it.

What happenedTicked criteria
The customer hung up first, before the agent’s goodbyeFull points, marked N/A
The customer hung up first, after the agent’s goodbyeScored normally
The agent hung up firstScored normally
Who hung up first is not known for the callScored normally

A goodbye is the agent taking leave or wishing the customer well, or transferring them to the survey. “Okay”, “agreed” or “fine” in reply to the customer’s thanks is not a goodbye, and neither is a sentence left unfinished.

Why the goodbye matters. On most normally closed calls the customer is also the one who hangs up — after the agent’s goodbye. An agent who reached the goodbye had the chance to do every closing step before it, so on those calls the steps are scored as usual. Without that check, ended_by alone would excuse the closing steps of almost every call.

What to tick. Steps that belong to the close of the call: the summary of what was agreed, final reminders and offers, payment or self-service channels, “is there anything else?”, the farewell and the survey. Leave unticked the opening, steps from the middle of the call, and conduct judged across the whole call such as tone or interruptions.

Rules:

  • Never critical. A hang-up does not excuse forbidden conduct. Offensive language still fails the call, whoever hung up first. Ticking Critical Error clears the box, and a rubric that has both on one criterion is rejected.
  • Only lost points change. A ticked step the agent completed before the customer left keeps its score and stays evaluated, so the scorecard’s Applied to count still shows it was done.
  • The justification is kept as written. The evaluator reports where the agent said goodbye; the platform then applies the rule. A justification may therefore still describe the step as missed. The N/A and Customer hung up badges are what tell you it was excused.
  • It applies to calls scored afterwards. Unlike the goal of the call, ticking or clearing it does not change calls already scored. Analyze them again to apply it.

Suggest hang-up exemptions in the rubric editor reads your criterion descriptions and ticks the ones they show to be closing-script steps. A line under each criterion — Suggested: excuse. or Suggested: do not excuse. — gives the reason.

  • It only ticks boxes. It never clears one, never ticks a critical criterion, and saves nothing. Review the result, then click Save.
  • It runs on its own after an import that leaves no criterion excused on hang-up.
  • It only goes by what the text says. A criterion described as happening “anywhere in the call” is not suggested, even when in practice it is a closing question. Tick that one yourself.

The Description is what the evaluator actually reads. The points below separate a criterion that scores the way you intend from one that does not. The review checks most of them for you.

1. Say what earns full marks, in proportion to the weight. This is where rubric effort pays. One rubric we worked through described its heaviest criterion — 35 of 110 points — in a single 196-character sentence about “responding coherently to objections”. Nothing in it said where the line between partial and full credit sat, so the score wandered. Rewritten to 826 characters, with the behaviors spelled out, it scored consistently. A criterion carrying a third of your points deserves more than one line.

2. Put applicability in its own sentence. The single most valuable thing you can do, and the one place ambiguity actively hurts:

Ambiguous — Offer the retention discount, or explain the cancellation consequences (applies to calls where the customer does not accept the offer)

Does the parenthetical govern the whole criterion, or only the clause beside it? Both readings are grammatical, and the evaluator will not settle on one. On a 30-call sample, the same criterion written this way was treated as not applicable 10 times in one run and 14 times in the next.

Clear — Explain the cancellation consequences. This criterion applies only to calls where the customer does not accept the offer. When the customer accepts, criterion 4 is evaluated instead.

An ambiguous criterion does not produce a wrong score. It produces an unstable one — which is much harder to defend to an agent disputing their result.

3. Name any exemption specific to your process. Hang-ups, dropped lines and situations that never arose are handled automatically — closing steps of a call the customer leaves before the goodbye need the hang-up flag — see How Scoring Works. If your process exempts a criterion for some other reason — a transfer to another queue, a system outage, a script the agent is not permitted to read — say so, in its own sentence.

4. When order is what you judge, say so. If a criterion is really “X must happen before Y” — verify identity before giving account details, state the cost before taking the order — listing the steps in order does not make the order part of the criterion. Say that the order counts, and what the score is when it is wrong:

Verify the caller's identity with at least two details. The order is part of the criterion: verification must come before the first account detail the agent gives. 20 points: two or more details verified first. 10 points: one detail verified first. 0 points: no verification, or verification after account details were given.

5. For a prohibition, say what is not the violation. Prohibitions are only docked on evidence, but the evaluator still has to tell the forbidden action from a permitted one that looks like it. A rule against asking for the customer’s password sits right next to the required identity check — both are the agent asking for a number — so name what is allowed: “Asking for the ID number, full name, date of birth or last four digits of the card is allowed and must not be penalized.” Likewise, refusing a discount the customer asks for is not a promise outside policy, and an announced hold is not abandoning the call however long it runs.

6. Say which criterion takes a shared fault. When one mistake could be charged to two criteria — a missing price is both an incomplete offer and arguably a misstatement of the price — say which one takes it. Otherwise the same omission is docked twice, and if the second criterion is critical it fails the whole call.

7. Do not put scoring rules in the rubric’s Description field. Only each criterion’s name, description, section, points and critical flag are shown to the evaluator. The goal flag is not, so marking a goal never changes how a call is scored. Neither is the hang-up flag: the platform applies it after the evaluator has scored the call. The rubric-level Description is for your own reference — a note about when to use this rubric, visible to your team and nobody else. Rules written there have no effect on scoring.


  1. Click Add Criterion in the rubric editor
  2. Fill in the fields for each criterion
  3. Repeat for each criterion you want to add
  4. Optionally, tick Goal of the call on the criterion that records the outcome
  5. Optionally, tick Excused if the customer hangs up before the farewell on the closing-script steps, or click Suggest hang-up exemptions
  6. The editor shows a running total of criteria count, total points, and critical criteria

Click Import rubric in the rubric editor, then choose a file or paste the text:

  • Excel (.xlsx), Word (.docx), PDF, plain text, or text pasted into the box — the form your supervisors use, as it is. Each item that awards or deducts points becomes a criterion. Its wording, scoring levels and notes are copied word for word; nothing is rewritten at this stage. Headings become sections, and items the form marks as critical or fatal become critical errors.
  • CSV in the column format below — read exactly as written, with no AI involved.

A scanned PDF contains only images and cannot be read; export the form from the program it was written in instead. Files are limited to 10 MB and about ten pages of text.

Values the form does not give. An item with no weight gets 10 points, or 1 point if it is a critical error (failing it scores the call at 0% anyway). An item whose mark only might mean critical is left as not critical. Either way the criterion shows Value not in the document: check those before saving.

Importing replaces the criteria in the editor. Nothing is saved until you click Save.

One row per criterion. A header row is optional: the importer recognises one and skips it.

id,name,description,max_points,is_critical,section,is_goal,excused_on_customer_hangup

Example CSV:

greeting,Greeting Protocol,Agent greets customer and identifies themselves,10,false,Opening
verification,Data Verification,Agent verifies caller identity,15,true,Security
active_listening,Active Listening,Agent demonstrates active listening skills,10,false,Body
resolution,Issue Resolution,Agent resolves or properly escalates the issue,20,false,Body
closing,Closing Protocol,Agent summarizes and closes appropriately,10,false,Closing,false,true
resolved,Issue Resolved,The customer leaves with the issue resolved or a solution agreed,5,false,Closing,true

Notes:

  • is_critical, is_goal and excused_on_customer_hangup accept true, 1, yes, si, sí or x for yes, and false, 0, no or blank for no
  • is_goal and excused_on_customer_hangup are optional and may be left off every row
  • A file that marks more than one goal, a goal that is also critical, or two criteria with the same ID is rejected with an error
  • When no criterion is excused on hang-up after the import, the editor runs Suggest hang-up exemptions
  • Files saved by Excel in Windows encoding are read correctly

Rows the importer cannot use are listed above the criteria, with the row number and the reason: too few columns, a missing ID or name, points that are not a positive whole number, or a flag such as todo or TBD that is neither yes nor no, which is read as no.


After an import, the editor reviews every criterion on its own. Click Review wording to run it again at any time. Criteria with something to fix show a Review box with what was found and, usually, a suggested description.

FindingWhat it means
Unclear what earns the pointsA heavy criterion does not say what earns full, partial or zero points, so its score can land anywhere in its range
Unclear when it appliesThe criterion only covers some calls, but does not say so in a sentence of its own
Order not statedWhat is judged is that one thing happens before another, but the text does not say the order counts
Unclear what is allowedA prohibition sits next to a permitted action that looks like it, with nothing saying which is which
Same fault in two criteriaOne mistake could be docked by this criterion and another, and neither says which takes it
No rule for a customer hang-upA closing-script step has no rule for a customer who says goodbye and hangs up first

The suggested description keeps every standard, threshold and point value of yours. Where your rubric leaves a choice open — which of two readings you mean, which criterion takes a shared fault, what earns partial points — the suggestion makes it, following the rest of your rubric, your supervisors’ scores if you have calibrated, and common quality practice. Each choice is listed under the suggestion with its main alternative. There is nothing to fill in: you only check.

  • Accept suggestion replaces the description; Dismiss keeps yours. Show the original from your document brings back the imported wording after a change.
  • Accept all suggestions, above the criteria when two or more have one, accepts them all at once.
  • Before accepting, read the listed choices and check that the suggestion still says what your supervisors mean: the review reads only the text. If your supervisors score a case differently, accept the suggestion and change that sentence, or dismiss it.

Start from a template on the rubrics page copies a ready-made rubric into your team: inbound banking service, and outbound sales. Both were written for AI scoring following the advice on this page and tested on sample calls, and both are in Spanish. Adapt the criteria to your process like any other rubric.

To base a new rubric on one of your own, click the Duplicate icon on its card.


Before pointing the whole operation at a rubric, check it against the people who score calls today. Open a saved rubric and click the Calibrate tab:

  1. Calls your supervisors scored — choose a call project and up to 30 calls that already have a transcript. Pick calls your supervisors have scored by hand, and include some poor ones, not only good ones.
  2. Their scores — click Download sheet, write each supervisor score under its criterion (leave blank what was not scored), and upload the sheet as CSV or Excel. Calls and criteria the sheet does not match are listed.
  3. Compare — click Score the calls with version N. Each call is scored twice. The page shows how many minutes are left, and you can leave it while the run continues.

The result shows:

  • Agreement with supervisors — the share of scores where the AI gave exactly what the supervisor gave, overall and per criterion. Criteria under 80% are highlighted.
  • Scores that changed between the two passes — a criterion scored differently on the same call each time has wording that allows two readings. Fix these first: an unstable score is harder to defend to an agent than a strict one.
  • Scores that differ from the supervisors — for each one, the call, both scores, and the AI’s justification and quote, so you can see whether the AI misread the criterion or the supervisor’s score is the one to check.

Suggest wording fixes sends the disagreements to the review and puts its suggestions on the criteria. These suggestions follow your supervisors’ scores, since those are your standard. Accept what fits, click Save to create a new version, then run the calibration again. Compare with an earlier run shows each criterion’s agreement under both versions side by side.

Calibration runs do not count against your plan, and they never change the scores in your reports.


A rubric can also be stored on the project itself. This happens when you save a COPC configuration for automatic analysis: the selected rubric becomes the project’s rubric, and every call that arrives afterwards is scored against it without asking again.

The project’s rubric also drives the Call Center Report — the quality scorecard lists criteria in rubric order instead of by point weight, so the report reads like your rubric document.


  • Edit: Click the edit icon on any rubric card to modify its name, description, or criteria
  • Duplicate: Click the copy icon to create a copy you can change without touching the original
  • Delete: Click the delete icon to permanently remove a rubric

Every save that changes the criteria creates a new version of the rubric; changing only its name or description does not. The rubric card shows the current version and, when there are any, how many calls were scored with an earlier one.

Editing a rubric never changes scores that were already produced: each call keeps the version it was scored with until you analyze it again. The one exception is the goal of the call: changing it updates Goal reached on calls already scored, since that is read from their stored scores. The hang-up flag is not an exception — analyze the calls again to apply it.

When a project’s combined COPC results hold calls scored with different versions, the results page says so: averages then mix two wordings, so analyze the older calls again before comparing like with like.


  • Keep criteria focused — Each criterion should evaluate one specific behavior
  • Use meaningful IDs — IDs appear in exports and reports, so make them descriptive, and keep them stable once results exist
  • Balance point weights — Assign more points to criteria that matter most, and describe the heavy ones in the most detail
  • Use critical errors sparingly — Reserve them for truly non-negotiable requirements
  • Mark the goal of the call — One low-weight, all-or-nothing criterion for the outcome lets the reports separate calls that achieved their purpose from those that did not
  • Tick the hang-up flag on closing steps — If you send who hung up first, closing-script steps are no longer docked on calls the customer cut short
  • Group with sections — Use sections to organize criteria logically (e.g. Opening, Body, Closing)
  • State applicability explicitly — A criterion that only covers some calls should say so in its own sentence
  • Calibrate before going live — Compare the AI with your supervisors on calls they already scored, and fix the criteria that disagree or change between passes

Before writing the descriptions, read How Scoring Works. It covers the rules the evaluator applies to every rubric on its own — hang-ups, criteria that do not apply, and how prohibitions are judged — which is most of what clients used to write by hand.