cx-effort-score

Use to measure customer effort from behavioural signals instead of CES surveys, and to audit whether a composite effort score is honest. Trigger for "customer effort score", "behavioural CES", effort

SKILL.md

GitHub

Behavioural customer effort

CES asks customers how hard something felt. The answers arrive late, from a biased sample, and conflate product pain with support pain. Behavioural effort uses what they already did: repeats, switches, transfers, reopens, and time-to-resolution.

The trap on the other side is a precision theatre composite — five decimal places on an index nobody can action. This skill builds a coarse, auditable effort signal, not a fake science score.

Signals that indicate effort

Each signal should be observable in ticket/contact data with stable definitions.

Signal Definition (typical) What it captures
Repeat contacts Same customer + same driver within N days Failure to resolve
Channel switches Different channel, same issue window Self-service or async failed
Transfers Handoffs between teams/queues in one episode Routing and specialisation friction
Reopens Closed ticket reopened within N days Premature closure or recurring defect
Handle episodes Count of agent touches to resolution Internal thrash
Time to resolution First contact to final close Delay; compare within driver
Self-service failure KB view or bot session then contact Documented in funnel
Customer message volume Inbound messages per issue Ping-pong async

Normalise time to resolution by driver. Billing disputes should not be scored against password resets on the same scale.

CES survey traps (when you still have survey CES)

Trap Symptom Behavioural cross-check
Post-interaction timing High CES, low outcome effort Reopens within 7 days
Agent-triggered sends CES improves, repeats rise Compare solicited vs automatic sample
Scale change Step change in trend Driver-level repeats flat
Channel mix Email CES vs chat CES Switch rate into voice
"Easy" closed quickly Good CES on rushed closes Reopen rate

If behavioural effort rises while CES falls, trust the behaviour until survey solicitation, timing, and response bias are audited.

Constructing a composite (without fake precision)

Step 1 — Binary flags per episode. For each resolved issue episode, mark whether each signal exceeded a threshold (e.g. ≥2 transfers, reopen within 7 days, ≥2 channels). Thresholds should be set from distribution elbows on your own data, not industry benchmarks.

Step 2 — Weight by actionability, not optics.

Weight higher Weight lower
Reopens, repeats on same driver Absolute handle time alone
Channel switch after bot/KB Single long wait on voice
Transfers across silos One transfer to specialist

Step 3 — Bucket, do not over-index. Report:

  • Low effort — zero or one minor signal
  • Medium — two signals or one severe (reopen, ≥3 transfers)
  • High — repeat + switch, or reopen + long TTR

A single 0–100 score is optional for dashboards; always publish bucket counts and top drivers of high effort alongside it.

Step 4 — Segment honestly. Effort varies by driver, channel, and customer type. One global number hides product bugs as "support friction."

high_effort_rate = high_effort_episodes / all_episodes   (by driver, channel)

Data requirements

Field Required
customer_id Stable across channels
issue_id or linked episode key Group touches into one effort story
driver / intent Segment thresholds
channel per touch Switch detection
started_at, closed_at TTR
transfer_count or queue history Transfer signal
reopened_at or status history Reopen signal

Without episode linking, every ticket looks like low effort.

Traps

Counting transfers as skill failure. Some drivers require specialist routing; measure avoidable transfers (return to previous queue, wrong team) via taxonomy.

Penalising proactive outreach. Exclude vendor-initiated contacts from repeats.

Bot sessions as zero effort. Include self-service failure in the episode.

False precision. Weighting tuned to 0.01 on 90 days of data will not replicate. Prefer stable buckets and driver breakdowns.

Present results to the user

  1. Effort bucket distribution — low / medium / high with episode counts and definitions used.
  2. Top drivers of high effort — ranked by volume × high-effort rate.
  3. Signal contribution — which flags fire most often within high-effort episodes.
  4. Channel and segment cuts — where switches and repeats concentrate.
  5. CES comparison (if available) — where survey and behaviour disagree, with likely survey bias named.
  6. Recommended interventions — mapped to signals (reopen → closure QA; switch → KB/bot; transfers → routing audit), not "improve effort score."