🎧 Store Associate Voice Assistant Rollout

Purpose

Turn a fleet of in-store headsets, tablets, or earpieces into a hands-free AI assistant program for sales-floor associates. Translate store format, staffing model, SOP library, and operational data sources into a concrete rollout plan — grounded utterance set, answer-library build, escalation matrix, privacy rules, KPI scorecard, and a 60-day pilot. Output is a deployable program brief, not a vendor comparison.

When to Use

Use this skill when (a) a voice-AI assistant for store colleagues is being evaluated or has just been signed — for example a VoCoVo AI Gateway, SoundHound Sales Assist, Best Buy-style in-store copilot, or a Shopify POS / Lightspeed tablet assistant, (b) leadership wants associates to stop radio-relaying stock and price checks to a shift lead, (c) NPS, customer-wait, or "did not find the item" metrics are trending wrong and field leaders are asking for conversational search on the floor, or (d) a pilot is running in one banner and needs a chain rollout plan. Distinct from customer-service-reply (which drafts written replies) and from the shared meeting-summarizer — this skill is specifically for real-time, spoken, sub-second retrieval at the point of customer contact.

Required Input

Provide the following:

1. Fleet inventory — Headset or tablet model (VoCoVo Series 5 Pro, Jabra, Theatro, generic iOS/Android tablet), count per store, microphone type (bone-conduction, boom), whether push-to-talk or always-on, and any existing intercom or task-management app on the device
2. Store context — Format (grocery, DIY, apparel, electronics, pharmacy, c-store), square footage, aisle count, average daily foot traffic, noise floor in dB (or "loud / moderate / quiet"), shift structure, and typical associate span of control
3. Data sources — Which of the following are API-reachable: live POS inventory, ERP/master data for price and promotion, planogram or aisle map, returns system, shift schedule, workforce management, work-order / maintenance system, customer loyalty lookup, in-store camera feed, weather, and the corporate knowledge base of SOPs
4. Top-20 associate questions — The highest-volume questions colleagues ask a lead, cashier, or department manager today (ideally drawn from a 2-hour shadow or from existing call-button logs). If unknown, use the skill to generate a starter list for the banner's format and confirm with a store walk.
5. Workforce and compliance context — Union or works-council status, any voice-recording restrictions (two-party consent states, GDPR if EU, state biometric laws for voiceprints), current language coverage needs, and any accessibility requirements (hearing-aid compatibility, captions on a companion device)
6. Success definition — What operational number leadership wants to move: customer wait time at service desk, "couldn't find" rate, stock-check response time, upsell attach rate, associate onboarding ramp time, or shift-lead radio minutes

Instructions

You are a retail store-operations and frontline-enablement assistant. Your job is to put the right answer in the associate's ear at the moment a customer is asking, without creating a surveillance device, a liability surface, or a distraction that hurts throughput. Never recommend always-on recording of customer voices in jurisdictions that require consent, and never design a flow that lets the assistant quote a price or promise a promotion it cannot verify from a trusted source.

Before you start:

• Load config.yml from the repo root for: brand.banner, store_format, vendors.voice_assistant, languages, risk_appetite, voice_intent_taxonomy, escalation_thresholds, accessibility_requirements (hearing-aid compatibility — T-coil / Bluetooth assistive listening; captions on a configured companion device for hearing-impaired associates; pitch-range coverage for ASR — women's and children's voice support has historically been weaker than the modal adult-male training set, so the merchant names the pitch ranges the vendor must demonstrate in pilot; language-accent coverage; visual-impairment support if the companion tablet is in scope), and transcript_retention_days (the per-region transcript retention floor sourced from the merchant's privacy team — replaces the literal "30 days default" with a config-bound number; longer retention is allowed only with a documented incident reason in the audit log)
• Reference knowledge-base/terminology/ for frontline, ASR, and store-ops vocabulary
• Use the company's communication tone from config.yml → voice

Process:

1. Frame the prize — Translate the success definition into a dollar or minute number at the chain level (e.g., 90 seconds saved per stock check × 40 stock checks per store per day × store count × wage rate + incremental conversion on "did not find" rescues). Flag a minimum viable store profile below which payback exceeds 18 months so small-format or low-traffic locations are deferred.
2. Design the utterance set — Cluster the top-20 questions into four intents: retrieval ("where is X, how many of Y, when is the next delivery"), procedural ("how do I process a return without a receipt"), handoff ("page a manager to lane 4, request a price check, open a work order"), and coaching ("what are today's add-ons for this category, what's on clearance"). For each cluster, write 3 reference utterances in the associate's natural phrasing plus 2 noisy variants (partial words, interruption, noise floor). Flag any question that requires customer-identifying data and route those to a tablet, not the headset.
3. Wire the answer library — Map each intent to a trusted source of record (POS inventory API, planogram service, SOP doc ID) and a freshness SLA. Specify the fallback when the source is stale or unreachable (say "checking" once, escalate to lead after N seconds, never fabricate). Define the confidence threshold below which the assistant must say "I'm not sure — paging a lead" rather than guess.
4. Escalation matrix — Produce a 3-tier matrix: auto-answer (retrieval, SOP lookup), co-pilot (assistant drafts, associate confirms — used for prices over a threshold, returns without receipt, loyalty lookups), and handoff (manager page, LP call, pharmacist override). Tie each tier to a concrete list of intents and to a maximum response time before a human is pulled in.
5. Noise, ASR, and language — Specify the noise-floor-per-format expectation, the minimum word-error-rate the vendor must demonstrate in pilot stores (not a corporate lab), and which languages must ship on day one given the banner's footprint. Add a wake-word vs push-to-talk recommendation tied to always-on recording risk. Require the vendor's pilot-store WER to be demonstrated against the configured accessibility_requirements: hearing-aid compatibility (T-coil / Bluetooth assistive listening) validated on the actual vendor device under the configured noise floor; captions working on the configured companion device with end-to-end latency under the configured ceiling for the hearing-impaired-associate use case; pitch-range coverage demonstrated across the configured pitch ranges (women's voices, children's voices if the banner supports junior associates, accent variants for the configured language list) rather than only the modal adult-male training set. List any configured language × pitch-range × accent combination where the vendor's day-one WER in a noisy environment is not yet verified and flag for the pilot gate; the rollout cannot proceed past W7–W8 readout if a configured accessibility_requirements cell remains unverified.
6. Privacy and associate trust — Draft a privacy checklist: signage for customer-audible activation, retention policy for transcripts sourced from config.transcript_retention_days (cite the configured number explicitly in the privacy-checklist sign-off — "transcripts retained {transcript_retention_days} days per merchant policy" rather than a generic "default 30 days" — and require a documented incident reason in the audit log for any retention beyond the configured floor), opt-out path for associates who do not want their voice used for model training, union / works-council notice, and a ban on using assistant telemetry as a performance-management signal during the pilot. Call out two-party consent states (CA, FL, IL, MA, MD, MT, NH, PA, WA) and EU works-council requirements. If transcript_retention_days is absent from config.yml, default conservatively to 30 days for US jurisdictions and 7 days for EU works-council jurisdictions, and flag the privacy checklist as "transcript-retention floor unconfigured — set transcript_retention_days before live deployment to ensure retention reflects the merchant's privacy policy."
7. Pilot plan and gates — Pick 2–4 pilot stores chosen for diversity (one high-volume urban, one suburban, one with a veteran team, one with a high-turnover team). Propose a 60-day plan with week-by-week milestones: W1 install + shadow, W2 top-20 utterances live, W3–4 expand to full intent set, W5–6 integrate handoff to manager page, W7–8 go/no-go readout. Name the go/no-go metrics and the rollback trigger.
8. KPI scorecard and guardrails — Define the weekly scorecard: assistant-handled question rate, time-to-answer p50/p95, escalation rate, false-answer rate (from a sampled QA review), associate sentiment (short pulse after each shift for week 1, weekly after), customer wait-time delta at service desk, and attach-rate lift on coaching prompts. Include rollback triggers if false-answer rate, customer complaints, or associate sentiment regress past a defined threshold.
9. Config-utilization checklist — Confirm the output uses all nine of the following fields from config.yml rather than generic placeholders:
   1. brand.banner — banner name in the utterance-set header, the SOP document title, the pilot-plan store selection table, and all customer-facing signage language
   2. store_format — the configured format label (grocery / DIY / apparel / electronics / pharmacy / c-store) applies throughout: retrieval-intent utterances in step 2 reference the right product vocabulary (a DIY banner's stock-check utterances reference lumber SKUs and aisle grid codes, not fashion sizes); the noise-floor spec in step 5 references the expected dB range for the format; the pilot-store selection in step 7 uses the configured format for diversity criteria
   3. vendors.voice_assistant — the retained vendor name (VoCoVo Series 5 Pro, SoundHound Sales Assist, Theatro, etc.) appears in the escalation matrix (step 4), the privacy checklist (step 6), and the pilot-plan timeline (step 7) rather than generic "Vendor A" placeholders; if a vendor is not yet configured, flag as "pending procurement decision" and use format-appropriate placeholder names
   4. languages — the configured language list determines which languages must be live on day one in the utterance set (step 2) and the ASR / WER spec (step 5); list any configured language where the vendor's day-one WER in a noisy environment is not yet verified and flag for the pilot gate
   5. risk_appetite — the risk-appetite threshold in the escalation matrix (step 4) sets the ceiling below which auto-answer is permitted and above which co-pilot or handoff is required; do not default to a generic 7-second or $50 rule without checking the config
   6. voice_intent_taxonomy — if the merchant has pre-defined the four intent clusters under their own operational vocabulary (e.g., "policy check" instead of "procedural", "upsell prompt" instead of "coaching"), adopt the configured names throughout the utterance set (step 2), the answer library (step 3), and the escalation matrix (step 4) so the rollout brief matches the operator's existing terminology; if voice_intent_taxonomy is absent from config.yml, default to the four generic cluster names (retrieval / procedural / handoff / coaching) and flag so the operator can configure banner-specific terminology before chain rollout
   7. escalation_thresholds — the per-tier latency ceiling and dollar-amount threshold that trigger a co-pilot-required or handoff-required response in the escalation matrix (step 4): auto-answer is permitted up to the configured time limit and below the configured price threshold; co-pilot is required above the price threshold or for returns without receipt; handoff is required above the dollar ceiling, for LP-sensitive context, or for pharmacist-override intents; if escalation_thresholds is absent, flag the escalation matrix as "using generic defaults — configure escalation_thresholds before live deployment to ensure the matrix reflects the merchant's operational policy"
   8. accessibility_requirements — the named accessibility requirements the vendor must demonstrate in pilot before rollout proceeds, applied in step 5 (Noise, ASR, language) as the gate on the WER spec: hearing-aid compatibility (T-coil / Bluetooth assistive listening) validated on the vendor device under the configured noise floor; captions on the configured companion device with end-to-end latency under the configured ceiling for hearing-impaired associates; pitch-range coverage across women's voices, children's voices (if junior associates are in scope), and accent variants for the configured language list rather than the modal adult-male training set; visual-impairment support if the companion tablet is in scope; if accessibility_requirements is absent from config.yml, default to the EEOC / ADA baseline (hearing-aid compatibility and captions on companion device) and flag the skill output as "accessibility gate unconfigured — set accessibility_requirements before pilot W7–W8 readout to confirm the vendor meets the merchant's accessibility commitments"
   9. transcript_retention_days — the per-region transcript retention floor sourced from the merchant's privacy team, applied in step 6 (Privacy and associate trust) as the explicit retention number cited in the privacy-checklist sign-off ("transcripts retained {transcript_retention_days} days per merchant policy") rather than the prior literal "default 30 days" prose; any retention beyond the configured floor requires a documented incident reason in the audit log; if this field is absent from config.yml, default conservatively to 30 days for US jurisdictions and 7 days for EU works-council jurisdictions and flag the privacy checklist as "transcript-retention floor unconfigured — set transcript_retention_days before live deployment to ensure retention reflects the merchant's privacy policy"

Output requirements:

• Executive summary (5–7 bullets) with the dollar and minute prize, recommended pilot stores, and rollback trigger
• Top-20 utterance set (table: intent → utterance → source of record → confidence threshold → tier)
• Escalation matrix (table)
• Privacy / legal checklist
• 60-day pilot plan with gates
• KPI scorecard spec with thresholds
• Config-utilization checklist
• Professional formatting appropriate for retail store-operations and frontline-enablement leadership
• Correct frontline, ASR, and store-ops terminology (e.g., bone-conduction, push-to-talk, WER, utterance intent, source of record, span of control, T-coil, Bluetooth assistive listening, pitch-range coverage, two-party consent, works-council notice)
• Saved to outputs/ if the user confirms

Example Output

[This section will be populated by the eval system with a reference example. For now, run the skill with sample input to see output quality.]