# Cx Outsourcer Scorecard > Use to compare BPO sites, vendors or partner teams fairly, adjusting for the work mix each is given before concluding anything about performance. Trigger for "compare our BPO sites", "which vendor is performing best", "site A scores lower than site B", outsourcer QBR packs, partner MI reporting, or setting contractual quality targets with a vendor. - Skill: `rulebase-co/cx-outsourcer-scorecard` (Agent Skill) - Install (CLI): `npx skillmds add rulebase-co/cx-outsourcer-scorecard` - Raw SKILL.md: https://api.skillmd.com/api/skills/rulebase-co/cx-outsourcer-scorecard/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: rulebase-co (https://skillmd.com/u/rulebase-co) - Updated: 2026-08-19 - Page: https://skillmd.com/skills/rulebase-co/cx-outsourcer-scorecard --- # Comparing outsourcers and sites A vendor comparison has money and contracts attached, so it gets argued about in a way an internal team comparison does not. The vendor will find any methodological weakness, and they will usually be right about it — because the naive comparison is genuinely unfair. **The default finding of an unadjusted vendor comparison is "the site given the harder work performs worse".** That is not a finding. ## Adjust for mix before comparing anything Sites are rarely given identical work. Differences that move quality and handle time independently of how well the site operates: - **Contact-driver mix.** A site handling disputes and closures will score below one handling password resets, whatever its capability. - **Channel mix**, which has the widest baseline spread of anything. - **Language and market.** Includes whether they are working in agents' first language. - **Tenure profile.** A site mid-ramp after a hiring wave is not comparable to a stable one, and this is often the entire explanation for a gap. - **Hours covered.** Overnight and weekend shifts carry a different mix, thinner support, and less access to escalation. - **Systems access.** Vendors are frequently given narrower tool access than internal teams, then measured on resolution. Adjust by **comparing within like-for-like strata** — same contact driver, same channel, same language — and then weighting back to a common mix. Report both the unadjusted and adjusted figures with the mix difference shown, because a vendor whose raw score is worse and whose adjusted score is better has a legitimate grievance and will make it. Where a stratum exists at only one site, it cannot contribute to the comparison. Say so, and say how much volume that removed. ## Make sure the instrument is fair **Check grader agreement per site before comparing scores.** If the rubric is applied less reliably to one site's work — a different language, a channel with worse transcription, graders unfamiliar with that market — then the sites are being measured with different instruments and the comparison is not valid at all. This check comes before the comparison, not after it is disputed. Also verify: - **Coverage parity.** Are sites evaluated at the same rate? A site sampled at twice the rate has a more precise score, and if the sampling rate rose after concerns were raised, the comparison looks targeted. - **Who grades.** If each vendor grades itself, the scores are not comparable at all. Calibrate across graders on a shared gold set and report the agreement. - **Same scorecard version** across sites and across the whole window. ## Report distributions, not just averages A site average hides the shape, and the shape is where the operational finding is: - Two sites with the same mean and different variance are different problems. High variance means inconsistent execution — usually training or supervision — and the fix is different from a uniformly low mean. - **Report the bottom decile.** A site with a good mean and a bad tail has a concentrated problem that is cheaper to fix than a general one. - **Report per-agent intervals.** A site with 200 agents and 4 evaluations each has no usable per-agent data, and a QBR that ranks individuals on it will not survive contact with the vendor's own analysis. ## Contractual measures need extra care If numbers feed a service credit or a penalty, treat them as a contract, not a report: - **Define every term in the contract**, not in a dashboard: the date field, the status filter, the exclusions, the small-cell rule, the restatement policy. Ambiguity here becomes a dispute later, and the vendor will read it more carefully than you wrote it. - **State the measurement error.** A monthly score with a ±5-point interval should not trigger a penalty at a 2-point miss. Build a tolerance band in deliberately, or you are penalising sampling noise. - **Agree the sample design in advance**, including who draws it and the seed. - **Define what happens when volume is too low** to measure a period. - **Agree a dispute process** with a defined adjudication route, and expect it to be used. **Do not let a metric that was designed for internal coaching become a contractual gate without redesigning it.** Coaching metrics tolerate noise; contracts do not. ## What to compare beyond quality scores A quality score alone is a thin picture of a vendor and easy to argue with. Add: - **Repeat contact rate** on comparable work — harder to game than a QA score. - **Escalation rate**, split by cause. High authority-escalation usually reflects the access you granted, not their capability. - **Attrition and tenure**, which predict next quarter's scores better than this quarter's scores do. - **Coverage and adherence to agreed volumes.** - **Complaint rate** on comparable work. ## Traps - **Reading a ramping site as a failing site.** Check the tenure profile first, every time. - **Comparing across a scorecard change**, or across a period when one site changed scope. - **Ranking sites on a single month.** Monthly variation on realistic sample sizes is large; use a rolling window and show the interval. - **Ignoring what you control.** Access, documentation, training and the routing that hands them the work are yours. If a vendor's escalation rate is high because they cannot see the system, that is your finding. - **Presenting a league table.** It produces defensiveness and gaming, and it obscures that the fix is usually specific and per-driver. ## Present results to the user 1. **Mix comparison first** — how the work differs across sites, before any performance number. 2. **Instrument fairness** — grader agreement per site, coverage parity, grader identity, scorecard version. 3. **Unadjusted and mix-adjusted results side by side**, with distributions and intervals, not just means. 4. **Volume excluded** because a stratum existed at only one site. 5. **The wider measure set** — repeat contact, escalation by cause, tenure, complaints. 6. **What is attributable to the vendor and what is attributable to you** — access, documentation, routing, work allocation. Separate these explicitly. 7. **If numbers are contractual**: the exact definitions, the tolerance band, and the dispute route.