07 / AI strategy / Transformation
Relay Pulse
A team AI strategy for an activewear company that did not fail at its work but at noticing: an alarm layer that routes anomalies to a named owner with a clock, chosen over a $2.8M personalization engine.
- VantagePoint engineRefuse as scoped$2.8M over 3 years · metrics can rise while margin falls
- Relay PulseBuildowned data · weeks to pilot · internal only
- CX CopilotPilot, phase 2a human sends every message · only after Pulse proves governance
01 / The question
What actually failed at Relay?
Not the work. Orders were up 11%, gross margin was stable at 57%, and fulfillment shipped what it had in December. What failed was the noticing: the signal of a demand shift existed on December 10 in the daily order feed, an actor with expedite authority existed the same day, and it still took eleven days to act. A legal warning took fourteen months to reach no one, and the board's “AI everywhere” instruction produced twelve scattered initiatives.
02 / The diagnosis
Five jobs, three of them broken.
The Viable System Model splits an organisation into five jobs. The team scored each one against the evidence in the case.
| Job | Working? | Where it fails |
|---|---|---|
| S1 · Doing the work | Yes | Only downstream of missing signals: starved of warning, not of capability. |
| S2 · Coordinating | No | Nobody owns cross-function sync; the pilot launched without Operations or Community informed. |
| S3 · Running today | Partly | Corrects monthly, not daily; an intra-month demand shift is invisible until the next review. |
| S4 · Sensing tomorrow | No | No owner and no competitive-intelligence function; missed the December 10 turn and Cadence's economics. |
| S5 · Holding identity | No machinery | Values live in the Founder, not in process; a Legal flag took fourteen months to reach no one. |
The failing system is not a starved S1 but the link between its units: no function failed at its own job, and the space between functions had no owner. Two coordination traps show it. A Marketing pilot lifted conversion by 22% while Fulfillment still planned to the old baseline, and complaints landed on Pacers who did not know a campaign existed.
03 / Three use cases
Three use cases, three verdicts.
Each candidate was tested for whether the metric it moves is real or vanity, what it costs, and how much autonomy it takes.
| Use case | What it does | Verdict |
|---|---|---|
| VantagePoint engine | 12 pre-built personalization modules; auto-generated customer content. $2.8M over 3 years plus data rights. | Refuse as scoped |
| Relay Pulse (the team's proposal) | Detects anomalies across owned data and routes alerts to named owners with clocks. Small build, weeks to pilot. | Build |
| CX Copilot | Drafts context and replies for human CX agents; a human sends every message. | Pilot, phase 2 |
VantagePoint moves clicks and sessions, which can rise while margin falls, and its contract would train competitors on Relay's data. Relay Pulse uses data the company already owns, stays internal, and keeps AI at recommending and drafting while humans commit. It is also the alarm channel the company lacks, with a named owner, the CDOO.
04 / The instrument panel
Three KPIs that name a decision and an owner.
The panel is built on one sentence: Relay gets healthier when signals reach someone who can act before the customer feels the gap. Every KPI names the decision it drives and is paired with a guard-rail against gaming it. The team built it as a live dashboard.
| KPI | Decision it drives | Guard-rail | Value |
|---|---|---|---|
| Signal-to-action latency | Median hours from anomaly to owner action. Drives escalation and staffing; owner CDOO. | Paired with % acted within SLA (81%) | 14h |
| Emerging-issue index | 0–100 composite from transcripts, reviews and forum text. Drives response briefs; owner Head of Community and CX. | Paired with tenure-weighted Pacer churn | 57 |
| LTV:CAC ratio | Includes effective CAC with 24% returns waste. Drives spend rebalancing toward community channels; owner CFO. | Paired with cohort LTV | 2.81 |
05 / Governance and roadmap
Ethics with machinery, and a small first step.
Relay's fifth job failed because identity lived in the Founder's voice instead of in process. The team rebuilds it with the CLAIM rail: a one-step ethics flag any employee can trip, a review gate for any AI that touches people, published charter lines, provenance for generated content, and a visible register of AI use. The recommendation is checked against GDPR, EU AI Act Article 50, US endorsement rules and a light compute footprint, using statistical anomaly models rather than a large model on every signal.
| Phase | What happens |
|---|---|
| Phase 0 · weeks 1–4 | Demand playbook pilot on the Checkpoint replay, plus a tool registry. One product line, one playbook. |
| Phase 1 · weeks 5–12 | Trust playbook, Pacer council integration and training for the people who receive alerts. |
| Phase 2 · Q2 | Full ethics playbook, mining of verbatim feedback, and an outward-looking radar team. |
| Phase 3 · Q3 onward | Scale decision on hard margin evidence, and the gate for a CX Copilot pilot. |
Each phase has a named owner, a training cohort and a kill criterion. Phase 0 is deliberately small: one product line, one playbook, four weeks.
06 / The strategy
The most human company in activewear.
The winning aspiration, sharpened from the Founder's: the most human company in activewear, powered by the best operational AI in apparel, measured by health and trust rather than applause. It plays on owned channels and in the community, the signal-rich ground that the competitor, Cadence, does not have. It wins by responding at machine speed while staying human at the boundary, on capabilities that models cannot rent: owned first-party data, the Pulse sensing layer and consent-clean community data.
Where the team corrects the Founder: the vault needs windows, not just walls, because Pacer churn rose unseen inside the protected community. It keeps the no-AI-touches-customers line for now, and proposes a Phase 3 evidence gate to revisit CX augmentation with full disclosure.