Validation and delivery plan

Proposed plan, not a committed schedule. Targets below are experiment criteria chosen for discussion, not observed performance or industry benchmarks.

Fund the next stage only when the preceding stage demonstrates merchant value and controlled execution. A connector demo alone does not justify a broad platform roadmap.

Stage	Deliverable	Proposed exit gate
0. Observe the job	10 owner/operator interviews and 3 agency interviews, including real recent files and current tooling	At least 5 merchants demonstrate recurring pain; 3 agree to a scoped pilot and a commercial discussion
1. Prove the connection	One version-pinned test shop; correct product/stock/order reads; permissions and revocation	Results match authoritative records; wrong-tenant and revoked access fail; supported data semantics recorded
2. Prove one controlled write	Price proposal, approval and durable audit from the outset; apply, read-back and compensation	Duplicate/retry, concurrent-edit, uncertain-outcome and stop scenarios pass on a controlled store
3. Prove a repeated useful job	Supplier-file mapping and bounded updates with review, exceptions and receipts	Three pilot merchants complete repeated weekly jobs for four weeks; compare against their actual baseline
4. Decide whether to expand	Paid continuation, support/compute cost and workflow retention analysis	Continue only if owners save effort, trust outcomes and pay without unsustainable manual support

For staffing discussion only, stages 0–3 might occupy roughly 8–12 weeks for two engineers with part-time product/design and merchant access. This is an estimate, not a delivery promise; revise it after the connector spike. Interview recruitment, platform reuse gaps or extension conflicts can dominate the schedule.

Proposed pilot scorecard

Primary metric: merchant minutes per correctly completed recurring job, including review, exceptions and repair. Collect a baseline from comparable previous jobs and record task size. Do not count a chat reply as task completion.

Suggested continuation thresholds: at least 50% less active owner time, at least 95% of eligible rows verified without manual re-entry, and at least 2 of 3 pilot merchants willing to continue on a paid offer. These deliberately small-cohort thresholds are decision aids, not statistically reliable market estimates. Report the underlying counts and exclusions.

Any unauthorized write or cross-store access pauses write rollout and requires investigation. Track incorrect approved writes, conflicts, uncertain outcomes, compensation attempts and time to resolve. “Zero incidents in a pilot” cannot establish production safety.

Also measure installation/support hours per shop, repeated mapping corrections, inference and infrastructure cost per completed job, weekly repeat usage, and reasons for rejecting a proposal. Reduced clicks alone are not sufficient if the owner must audit every value manually.

Interview guide

Ask the merchant to demonstrate the last update rather than describe an ideal agent. Where did the file come from? Which columns needed correction? What happened to unmatched rows? Who checked the change? What broke last time? Which extension or person does the work today? Which system is authoritative? What access can they grant? What would make them reject our proposal?

Ask about price only after comparing concrete work and the current cost. Offer a clearly bounded paid pilot; distinguish a compliment, a letter of intent and an actual payment. This document authorizes no outreach or payment collection.

Stop, narrow or pivot

If existing imports already solve the job, test a different observed bottleneck such as source-grounded catalog content preparation. If only agencies show demand, evaluate an agency workflow rather than claiming owner adoption. If safe conditional writes or affordable support cannot be achieved, stop the write pilot. If read-only review alone is valued, test it as a different paid scope rather than pretending the execution thesis was proven.

Only after retention should we add recurring authorized jobs, quantity operations or another platform. Automatic refunds, ad-budget allocation and autonomous replenishment require separate product and risk decisions.