Conditional Go.

EvidenceNZ had 9,867 cafe and restaurant geographic units as at February 2025 — 18,243 including takeaway. Sector sales are large; cost pressure is severe.

AssessmentAMR addresses a measurable operating problem rather than manufacturing an AI use case. The strongest need is not more data — it's faster recognition of financially meaningful exceptions and clear actions.

AssessmentA small cafe with limited reservations and a basic POS may produce little usable guest data and not support the subscription or onboarding cost. Premium owner-led restaurants have the data, margin opportunity and management cadence AMR requires.

AssessmentGoverned metrics → weekly report → critical number → owner focuses → statused opportunities → pre-service stories → follow-up outcomes. A closed loop — observe, interpret, act, record, evaluate — is more defensible than a chatbot or dashboard because it becomes part of the operating cadence.

AssessmentOwners often praise reports but stop opening them once novelty declines. Fit requires evidence that recommendations are accepted, assigned, discussed and revisited.

AssessmentA small dataset won't be statistically useful, and restaurant comparisons are sensitive to concept, price point, service model and region. Poorly normalised benchmarks can mislead customers. NZ privacy law limits retention and overseas disclosure; Australian rules require de-identification when data is no longer needed.

AssessmentAn absolute category claim ("no incumbent combines guest intelligence, coaching and huddle execution") is difficult to prove and easy for an investor or competitor to rebut. It weakens credibility even where the underlying distinction is real.

EvidenceNow Book It activates API keys through its support and partner process; public documentation does not establish that a new third party can obtain unrestricted access on demand.

AssessmentA connector roadmap based on presumed API access can fail for commercial, contractual or technical reasons. CSV exports reduce the dependency but add onboarding and schema-drift costs.

EvidenceNatural-language analytics is becoming a standard capability; Tenzo already exposes AI access to restaurant data and released an MCP server beta in 2026.

AssessmentAsk AMR will be useful but is unlikely to remain differentiating. The harder-to-copy asset is a governed restaurant ontology, source mappings, service workflows, and labelled evidence showing which actions produced outcomes.

AssessmentComputing metrics, trends, segments and lifecycle rules in code — with the model only explaining and recommending — sharply limits arithmetic hallucination, supports auditability, reduces model cost and makes regression testing practical. Consistent with NIST guidance on generative-AI governance.

AssessmentStarting with an application-level registry and deferring a semantic-layer product achieves the important property — one tested definition per metric — without adding a platform dependency before product-market fit.

AssessmentRLS is a strong control, not a complete tenancy strategy: superusers and BYPASSRLS roles still bypass it, and a single privileged application connection, migration role or background worker can expose data.

AssessmentTokenisation reduces exposure but does not make all raw data non-personal — free-text notes, dates, booking context and linkage patterns may remain identifying. Crypto-shredding requires complete key coverage and doesn't address downstream copies, logs, AI outputs or exports. NZ Principle 9 and Australian APP 11 impose retention and destruction duties.

AssessmentPer-guest key lifecycle, rotation, backup restoration, merge/split behaviour and deletion verification add substantial security-critical code. AI generation does not reduce the need for expert design and review.

AssessmentEmail/phone arrays and JSON source-ID maps are convenient initially but make uniqueness, history, confidence, consent status and identifier-level correction harder. Identity resolution is central IP and deserves an explicit model.

AssessmentErroneous merges are inevitable. Undoing them safely requires knowing which source facts and child records came from which original identities — re-pointing all children during a merge can lose that information.

AssessmentString-matching numbers in model output is brittle: wrong labels, wrong periods, transposed values and unsupported arithmetic can pass; legitimate transformations get blocked until exceptions accumulate.

AssessmentAMR ingests reviews, feedback emails, booking notes and dot notes, then exposes them to a model. A malicious or accidental instruction inside feedback can influence output if source text is treated as trusted. Risk increases when the same agent can query metrics, guest records or future write-back actions. (OWASP LLM Top 10.)

AssessmentOffline caching is operationally sensible, but guest names, celebrations, preferences and notes may remain on personal phones after a shift, after staff leave, or after access is revoked.

AssessmentGood controls are necessary, but a small AI-assisted team can spend more time maintaining platform machinery than validating the product. SOC 2 certification before customers require it is unlikely to be the best use of capital.

AssessmentAI creates implementation capacity faster than review capacity. If the engineer is available only after features are generated, defects and architectural drift accumulate faster than they can be corrected.

AssessmentA long design document drifts as code and product decisions change; agents may follow outdated sections or reconcile contradictions inconsistently.

AssessmentHuman review alone will not scale with agent-generated change volume. The system must reject low-quality work automatically before the engineer sees it.

AssessmentLines of code and "% AI-generated" are easy to report but don't measure useful delivery — AI can increase code volume while also increasing rework, defects and review burden.

EvidenceNZ: 9,867 cafe/restaurant units (Feb 2025). Australia: 55,000+ cafe and restaurant businesses in 2024; ~89,000 including takeaway (June 2025).

AssessmentOnly a subset has the systems, revenue, guest identification and management maturity AMR needs.

AssessmentAMR's main early costs are onboarding, mapping, connector support, data reconciliation and customer success — not LLM tokens. A low monthly price creates poor unit economics before self-service exists. MarginEdge's public US pricing includes a US$350/month tier; direct competitors commonly quote.

AssessmentRestaurants are noisy environments — demand, events, weather, menu changes and staffing all move outcomes. Overstated attribution will destroy trust; correlation-only reporting may struggle to prove value.

AssessmentCSV-first with concierge onboarding is the correct way to learn mappings and workflows, but it creates a services burden that must be measured rather than hidden.

AssessmentOwners rely on accountants, bookkeepers, hospitality consultants, POS vendors and reservation providers for technology decisions. Founder-led sales suits pilots but may be expensive at a single-site price point.