calle-ai/awesome-phone-call-agents/call-apology-effectiveness-auditor

call-apology-effectiveness-auditor

**"I'm sorry you feel that way" is not an apology. It is a complaint

about the person who had one.**

Apology effectiveness is component-decomposable: research on real conflict apologies shows that a working apology is not one utterance but a stack of parts - acknowledgment of responsibility, concrete repair, brief explanation, expressed regret, forbearance, and (optionally) a request for forgiveness - and that acknowledgment carries the most weight. When a callee raises a grievance on a call, an agent that says only "I'm sorry about that" and moves on to scheduling has produced a rote apology: audible, and empty. This skill detects the grievance and grades the recovery response against that component stack.

When To Use

• after support, collections, or clinic calls where the callee may have

raised a grievance (delay, broken promise, repeated contact)

• in QA pipelines, to find calls where the recovery was rote, deflecting,

or missing entirely

• before recovery callbacks, to craft an apology-protocol goal that names

the miss and offers concrete repair

When Not To Use

• as a measure of sincerity or relationship health; the components are

surface-lexical markers - an agent can say "that's on us" with no feeling behind it

• to force apologies on every failure; sometimes the right posture is no

apology at all, because the agent cannot own fault it does not have

• during a call; strictly post-call analysis plus pre-call goal crafting
• on non-English transcripts; the lexicons are English-only

Workflow

Audit a finished call

python3 scripts/apology_effectiveness_auditor.py analyze \
  --transcript path/to/call-result.json

Reads the real get_call_run result shape ({status, result: {transcript}}) or the flat fixture shape used by sibling skill fixtures. Emits a card:

• grievance_turn_index / grievances_total / grievance_evidence:

where the first grievance was voiced (callee turns only), how many grievance-bearing callee turns exist, and the masked evidence sentence

• apology_detected / apology_evidence: whether any agent turn from the

grievance onward contains an apology, and the masked sentence

• components_present: the six-component markers found in the graded

window (the apology sentence plus the following agent turn, since explanations often arrive one turn late)

• typology: rote / empathic / explanatory (empathic > explanatory

> rote)

• verdict: EFFECTIVE_APOLOGY (acknowledgment plus repair or

explanation) / PARTIAL_APOLOGY / NON_APOLOGY (deflection override) / GRIEVANCE_UNADDRESSED / NO_GRIEVANCE, plus unclear paths (empty transcript, fewer than two turns)

Only the FIRST grievance's response window is graded; later grievances are counted in grievances_total but their responses are not separately graded.

Craft the apology-protocol goal

python3 scripts/apology_effectiveness_auditor.py craft --scenario apology-protocol

Emits the plan_call inputs JSON whose goal implements the protocol: acknowledge the specific miss, offer concrete repair, explain briefly only after acknowledging, and never use "sorry you feel that way" constructions.

Scientific Foundation

Citation notes recorded during verification: all three web-verified 2026-10-01. CORRECTION recorded: the frequently cited "2014 component structure of interpersonal apologies" title does not exist - the empirical six-component study is the 2016 NCMR paper by Lewicki, Polin & Lount (not "Polite").

Differences from sibling skills

• call-disfluency-stress-profiler measures stress markers as they

happen; this skill grades the recovery RESPONSE after a grievance is voiced.

• call-repair-sequence-auditor scores micro-level conversational repair

(huh/repeat/rephrase); this skill scores macro-level relational recovery.

• call-agent-commitment-tracker tracks future obligations; a repair

offer is graded for FORM here and tracked as an OBLIGATION there.