Features you can trace.
Numbers you can interpret.

This edition pairs the original four-part method with a short closed-loop introduction and ten case-study slides. Team A usage is measured; feature mechanisms are checked in source. ClearAds is the seller-event audience. Six animated product simulations illustrate existing mechanisms using fictional Morrow & Finch and Kitewell brands, invented people, products, dates and example values. These are stylised simulations, not recordings of the live applications.

Team A · four complete recorded weeks

10 August–6 September 2026, with week boundaries in Europe/Berlin. Reviewed on 7 September using complete production weekly aggregates, excluding test authentication. Internal database identifiers are withheld.

Week starting	Calls	OK / logged	No data	Other problems
2026-08-10	1,221	1,155	2	64
2026-08-17	4,147	4,061	28	58
2026-08-24	4,798	4,609	84	105
2026-08-31	3,526	3,414	32	80
Total	13,692	13,239	146	307

13,239 ÷ 13,692 = 96.7% successful tool responses (rounded to one decimal). Other problems include errors, refusals, rate limits and exceptions. This rate does not measure the correctness of the assistant's explanation.

Count the right thing.

One conversation may call several tools. The 17 identities are distinct non-null OAuth emails across the period, not a headcount of all employees. The 11,021 shared-key calls are the Slack door; individual Slack speakers cannot be counted from this aggregate. Database query calls (12,186) and Slack calls overlap. The 73 request-logging calls are not 73 unique requests. Telemetry is best effort and can omit events when recording fails.

Download the aggregate snapshot (JSON). No names, emails, prompts, query text, customer records or credentials are included.

Implementation evidence

Source was reviewed privately. Client repository names, links and commit identifiers are withheld. “Implemented” means the behaviour is present in reviewed source; this review did not operate Team B production or revalidate every feature flag.

Team A usage definitions
Weekly production aggregates used by the operations dashboard. Public snapshot contains only totals, no identities or prompts.
Team A learning fix
Documents the 28 August review: 69 of 103 failed queries received hints, 62 recovered within ten minutes. Implements proactive dictionary guidance. Historical engineering evidence, not a new experiment.
Team A grouped requests
Documents six repeated break-even ACoS asks in seven days and implements conservative matching; uncertain matches go to human review.
Team A shipped-wish feedback
On a subsequent eligible interaction, verified OAuth requesters receive a shipped-wish notice with its proof note. This is not a push-notification claim.
Team A decision ledger
Permanent decision records; outcomes must reference an earlier event. Structured idea, knowledge and how-to capture. No causal attribution is inferred.
Team B task intake
Read including the August addenda: one catalogue and one selection rule. Historical rollout counts are deliberately not presented as current adoption.
Team B plan snapshots
Stores the resolved flow plan so retries preserve original steps and dates.
Team B assignment and dates
Roles resolve to brand owners and initial due dates skip weekends.
Team B lifecycle and cascade
Records lifecycle changes and shifts downstream ClickUp flow task dates by the exact delta; wrapper follows the latest date.
Team B tracking
Implements cycle time, on-time rate, time-in-status, flow-step duration and freshness views. No current live totals were queried.
Team B native tasks
Feature-gated in-app creation and completion, with ledger events. Distinct from ClickUp flow creation.

Fictional examples and measured evidence

The fictional £1,240 ad spend, three linked requests, 2,000-test sample, 0.8% → 1.0% CTR, four tasks and two-hour cycle belong only to the animations. They are not client results. The 13,692 Team A calls and associated usage figures above remain the separately sourced real telemetry.

What the presentation does not measure

There is no measured revenue uplift, ROAS uplift, time saving or Team B adoption claim. Historical Team A engineering-review counts are labelled separately from the fresh aggregate query. An outcome linked to a decision preserves context; it does not establish causation.

A philosophy for sellers

The closing examples explain a shared pattern: own the data, turn decisions into owned work, record what happens and improve the playbook from that evidence. Team A and Team B remain distinct implementations.

Open the animated deck →