Platform concept · v3 · reviewed
Connect your data. Know the score. Find the opportunity.
Ask My Restaurant is an AI operating system for independent restaurants. It connects the software the restaurant already runs, turns the numbers into a scoreboard the whole team can read, and ends every answer with a recommendation — never just a report.
Real sample week from the founding restaurant's spreadsheet. Every number carries a trend against last week, the year average, and the rolling four-week average.
The problem
Bookings in Now Book It. Sales in Bepoz. Rosters in Loaded. Accounts in Xero. Guests in inboxes and review sites. The owner reconciles it in a spreadsheet on Mondays — when there's time. AMR brings it together and turns it into actions.
Reporting tools stop at what happened. Every AMR answer ends with a recommendation — or an explicit "no action needed," never filler.
Incumbents pitch "growth on autopilot." AMR's stance is the opposite: computers do the counting so the team does more noticing.
Avero runs $400–1,200 per location per month. AMR targets the single-site independent at a $100–300 envelope — architecture included.
| Player | What it does | Where it stops |
|---|---|---|
| Tenzo | AI analytics & forecasting, 70+ integrations, multi-site | Shows what's happening; doesn't run the operation |
| Nory | AI ops — inventory, staffing, payroll, forecasting | Weakest on guest relationships |
| Avero | Multi-location analytics for restaurant groups | Priced out of the independent segment |
| MarginEdge | Invoice & food-cost automation | Silent on guests and marketing |
| SevenRooms / Resy | Reservation platforms with guest data & "autopilot" marketing | Walled garden — you live on their booking platform |
| AMR | Guest intelligence + coaching + huddle execution over the stack independents already run | The gap all five leave open |
The product
Each layer earns the next. Data that flows becomes a score; a score becomes understanding; understanding becomes a plan; the plan walks onto the floor at 5:55pm.
Now Book It, Bepoz, Loaded, Xero, Me&U, Google reviews, website analytics — plus a per-tenant forwarding address for feedback email. CSV upload is the universal on-ramp; live connectors replace it source-by-source with zero downstream change. No manual reporting.
The digital evolution of the Monday spreadsheet. Revenue, covers, spend per head, beverage match, wage %, guest segments — every number with its trend, every week with its wins celebrated.
Ask AMR anything in plain English — "Did Michelin increase bookings?" "Who buys the beer?" — and get answers grounded in governed metrics. Analyst, accountant, interpreter. It refuses rather than guesses.
The biggest opportunity, the critical number, three owner focuses ranked by net-profit impact, questions for the team, a weekly game plan. Opportunities persist with status — next Monday asks whether the number moved.
Tonight's service, table by table: who's returning, who's celebrating, what to ask, what to notice. Phone-first, cached offline, staff-visible — the layer that walks onto the floor.
The noticing engine
Deterministic rules watch the guest graph and the week's bookings. The model writes the story and the questions. Rules find; the model narrates — and every answer a waiter records becomes a dot note that sharpens the next story.
Kim Wilson
Kim's back after her longest absence since 2024. Eleven visits puts her one short of a loyalty milestone. Tonight is the visit that decides whether the gap was a blip or a drift.
Thresholds are personal — multiples of each guest's own rhythm, never a fixed number of weeks. A weekly regular quiet for a month means something; a quarterly visitor doesn't.
July 2026 bookings were up 18% on July 2025. First-time guests rose 32%; repeat guests rose 4%.
The system
A multi-tenant SaaS from day one — shared-schema Postgres with enforced row-level security, one ingestion contract for CSV and API alike, and an AI that is never allowed to do arithmetic.
org · restaurant(tenant, timezone, goals, schedule) — tenant-local, DST-safe scheduling guest(emails[], phones[], visit_count, avg_gap_days, segment) — first_time | friend | regular | vip guest_link(match_confidence, status) — the "matches to review" queue guest_merge_event · guest_redirect(old → canonical) — identity stability contract booking · visit(menu_type, bev_match, tips, server_id) dot_note — what the team noticed team_member · shift · huddle_entry campaign · campaign_result(bookings_48h) feedback_item · theme · feedback_theme — voice of guest, trended metric_value(metric_key, period, value) — the governed metric store opportunity(evidence, est_profit_impact, status, outcome_metric_key) story(type, suggested_questions, status) ai_output(model_id, prompt_version, source_snapshot_ids) — full provenance ingest_run · context_event(weather, holidays, Michelin) audit_log · app_user(role)
GET /v1/scoreboard?period=week&ending=2026-01-18
GET /v1/huddle/{date} POST /v1/huddle/{date}/cashup
GET /v1/opportunities?status=suggested POST /v1/opportunities/{id}/status
POST /v1/ask POST /v1/goals
GET /v1/segments/{lapsed_regulars | lapsed_first_timers | celebrations}/export
POST /v1/imports GET /v1/sources/health
Load-bearing decisions
21% → 98%
Raw LLM text-to-SQL solves ~21% of real enterprise queries. Grounded in a governed metric layer, frontier models reach 98–100% — and refuse when they can't answer. Every KPI is defined once; the dashboard, the Monday report, and Ask AMR quote the same number or no number.
Deferred deliberately: a typed metric registry in code now; a semantic-layer product (dbt SL / Cube) at ~25 tenants.
3–5×
Single Postgres, tenant_id everywhere, row-level security that no code path can bypass — the 2026 B2B default, at 3–5× lower infra cost than siloed models. "Staff can't see wage %" is a database guarantee, not a UI convention.
Discipline: SET LOCAL per transaction, cross-tenant reads proven to fail in CI, and an org layer above the tenant from day one.
$100–300
Commodity connectors are bought; first-party adapters for the ANZ indie stack — Now Book It, Bepoz, Loaded, Me&U — are built, because those integrations plus the guest graph are the moat. Narratives are batch-generated and cached; chat is the only interactive LLM surface. The price envelope is an architectural constraint.
Deferred deliberately: Postgres job queue now; Temporal-class orchestration when connector concurrency actually hurts.
Roadmap
The vision docs' own warning governs: don't build Version 10 before Version 1 has paying users. Validation artifacts and platform foundations are separated so nothing gets thrown away twice.
No platform code. Write the 100 questions, interview 10 owners, run the manual export-and-upload loop. Gate: owners say "I'd pay for that."
Multi-tenant core, PII-tokenized ingestion, historical backfill, metric registry, Scoreboard, Monday report, Ask AMR on the top 20 questions with the full guardrail stack. One pilot tenant, concierge onboarding.
Live Now Book It + Bepoz adapters, identity resolution with the stability contract, stories, Today's Service with offline cache, staff role, action segments. AMR becomes daily-use. 3–5 pilot tenants.
Opportunity templates and the accountability loop, review and feedback themes, Loaded/Xero adapters automating wage % and food cost %, goals, context events. Cohort-wide rollout; first paid tier.
Self-serve onboarding, connector catalog, columnar marts, opt-in anonymized benchmarking — "your beverage match % vs similar restaurants" — and group views for multi-venue operators.
Adversarial review, folded in
An architecture review hunted for the failures that are cheap to fix on paper and expensive in production. Verdict: sound direction, conditionally approved — every condition is now in the design.
"Replayable forever" raw data full of guest PII contradicts privacy-law deletion rights. Fix: PII tokenized at ingest, values in a per-guest key vault — deletion becomes a crypto-shred, the raw zone keeps its integrity. Ships in Phase 1, because retrofitting is a migration project.
Replays and review decisions change matching results over time — and dot notes point at guest IDs. Fix: merges and splits as first-class events, a redirect table consulted by all readers, resolution as a versioned deterministic job. A note on the wrong guest breaks the hospitality promise.
dbt + Cube + Temporal is a platform team's stack; Phase 1 is two engineers and one restaurant. Fix: same interfaces, deferred products — metric registry in code, Postgres job queue, plain versioned SQL — each with a documented trigger for upgrading.
Reading a business inbox puts AMR in Google's restricted-scope regime — an annual five-figure security assessment. Fix: invert the flow. Each tenant forwards feedback email to their own AMR ingest address. Same data, no OAuth.
The vision docs ask nine-year-lookback questions, so onboarding starts with bulk history: batch identity resolution with triage, segments computed as-of each visit, baselines built before the first Monday report. It doubles as the concierge-onboarding runbook.
Requiring advice on every answer guarantees filler. The contract now allows an explicit "no action needed" with rationale — and the eval harness scores recommendation specificity, not just presence.