STR Operating Data Platform — Executive Summary

Executive verdict

Build it, but launch as a productized data service with a software core, not as an all-in-one platform.

No mature STR product was found that combines source-level provenance, conflict resolution, canonical property facts, human approvals, controlled publishing, and deterministic receipts. PMS, guest-experience, AI messaging, and operations tools each solve only part of the problem.

The strategic opening Become the neutral trust and portability layer around the PMS. Clean and govern information once, present it beautifully, then let the client use it through our applications, exports, APIs, webhooks, AI agents, or any future tool they choose.

0 Exact mature STR competitors identified

4 Roam Free properties for the internal MVP

95% Target fact-class coverage before pilot completion

1–3 Design partners after internal validation

What the MVP actually does

It creates one approved, traceable property fact base from messy and conflicting sources.

Observe Ingest claims from OwnerRez, listings, guidebooks, HostBuddy knowledge, documents, photos, and spreadsheets.

Reconcile Normalize facts, preserve every source, detect conflicts, and flag missing or stale information.

Approve Route uncertain or high-risk facts to a human with evidence and clear decision rights.

Publish safely Generate a versioned AI guest-messaging knowledge export with a complete receipt.

The first technical proof should be one outbound workflow. The broader customer package can later synchronize the same approved facts to multiple destinations.

The platform constitution

Beautiful by default. Open by design. Portable by contract.

The product should feel simple and polished to a nontechnical operator while exposing a flexible foundation for technical clients, partners, and AI agents.

Beautiful client experience Clear dashboards, evidence, approvals, search, history, exports, and client-ready information surfaces.

Client-owned operating data Canonical facts, source provenance, versions, relationships, permissions, and audit receipts remain independent from any vendor.

Open extension fabric Stable APIs, outbound webhooks, events, scoped agent actions, and documented contracts let clients build on top of us.

Replaceable system adapters Connectors translate between legacy SaaS tools and the canonical model so switching providers becomes repointing, validation, and controlled cutover.

The Slack-like lesson Do not copy Slack's product. Copy its extensibility: a strong core experience surrounded by permissions, APIs, webhooks, events, integrations, and a surface where third-party agents can safely act.

Five non-negotiable client rights

View all of their operating data in a useful human interface.
Export complete data and history in documented, machine-readable formats.
Transform and synchronize data without needing our professional services.
Connect their own AI tools with explicit permissions and revocable access.
Leave any downstream SaaS vendor without rebuilding the operating data foundation.

Open does not mean uncontrolled. Every integration and agent action needs tenant isolation, narrow permission scopes, approval policies, rate limits, idempotency, and a durable receipt.

Safe agents and private AI

Enable ambitious AI use without forcing clients to surrender control of their data.

Our agents and client-built agents should use the same safety gateway. No agent receives direct database access or special trust because of who built it.

Identify and scope Every agent has its own identity, tenant, purpose, allowed data, allowed tools, spending limits, and expiration.

Classify and route Policy determines whether a task may use an approved cloud model, must run locally, or cannot be performed.

Preview and approve Risky actions produce a proposed change and impact preview before a person or policy grants execution.

Execute and prove Rate limits, idempotency, isolation, monitoring, kill switches, and durable receipts make every action attributable and reversible where possible.

Private Compute mode · MVP decision locked Run document processing, embeddings, retrieval, inference, storage, and logs on hardware or private infrastructure the client controls. A strict local-only policy blocks outbound model traffic, so frontier AI labs never receive the protected data.

The Roam Free MVP will prove one protected local-model workflow. This must be a verifiable deployment property, not a marketing promise. The platform must show where a workload ran, which model handled it, what data boundary applied, and whether any network egress occurred.

Local models trade some frontier capability and operational simplicity for control, privacy, predictable cost, and independence. The routing layer lets each client choose that tradeoff by data class or enforce local-only processing across the entire workspace.

Why the partnership model can work

The research found repeated success in other fragmented verticals. The winning companies entered through one critical workflow, earned a trusted data position, then added adjacent services while customers retained their core work.

Toast · RestaurantsPOS and payments became the daily operating layer, then expanded into ordering, marketing, payroll, capital, and analytics.

athenahealth · Medical practicesSoftware plus human billing operations created a reusable network-learning loop across claims, records, and administration.

ServiceTitan · Skilled tradesOwns the operating surface around contractors while contractors keep performing the licensed physical work.

Procore · ConstructionBuilt a neutral evidence and collaboration layer across fragmented teams and systems.

Privia Health · Physician practicesPreserves practice independence while supplying administrative infrastructure and operating support.

Best composite analogy athenahealth’s human-plus-software learning loop, Procore’s evidence layer, and Privia’s autonomy-preserving partnership model.

How the wide vision should unfold

Every new service must reuse the trusted data or workflow. Selling something else to the same buyer is not enough. The shared base is a wide operating-data model, not a wide version-one feature set.

Foundation · Property truth and publishing Source evidence, canonical facts, approvals, audit history, connectors, and verified outputs.

First adjacency · Websites and listing content Highest data reuse, visible ROI, and moderate liability.

Second · Social media and CRM administration Start with CRM hygiene and content operations, with consent and deliverability controls.

Third · Compliance administration Track permits, insurance, taxes, inspections, renewals, and evidence. Partner for legal judgment.

Last · Bookkeeping operations Begin with close checklists and document completeness. Partner before owning regulated or high-risk work.

Do not build full modules for these adjacencies in version one. Build only the shared primitives and clean extension points.

What determines whether this wins

Copy these patterns

Make every human-resolved exception improve the software.
Keep source evidence, approvals, versions, and receipts attached to facts.
Preserve the property manager’s autonomy and decision rights.
Give clients full exports, open extension points, and a clean exit path.

Avoid these patterns

Bench-style linear service labor without software leverage.
Zenefits-style expansion faster than compliance controls.
Katerra-style movement into the customer’s core operations.
Custom consulting disguised as a repeatable product.

Commercial kill signals

Setup still takes more than six hours per property after the third client.
Connector breakage consumes more than 20% of engineering capacity.
More than 30% of target facts require fresh manual research each month.
Clients want VA labor but will not pay for setup plus ongoing freshness.
Review queues go stale because customers do not approve conflicts.

Non-negotiable client trust controls

MFA for administrators and publishers.
Strict workspace isolation with automated tests.
No guest PII, payment data, access codes, or secrets in canonical facts.
Append-only audit history, backups, restore testing, and incident response.
Contracts, DPA, insurance, access policies, and deletion guarantees before external launch.

Recommended first 90 days

Days 1–30 · Establish truth Canonical fact model, source observations, STR taxonomy, conflict review, audit ledger, and Roam Free imports.

Days 31–60 · Prove genericity Connector contract, LTA tenant test, agent safety gateway, approval rules, data classifications, execution-routing policy, and freshness dashboard.

Days 61–90 · Prove an outcome AI guest-messaging knowledge export, one protected local-model workflow, workload receipts, ROI measurement, security baseline, and first client package.

Internal MVP exit criteria

95% of the targeted property fact classes represented.
Every canonical fact has source evidence and freshness status.
Conflicts have an owner, status, and resolution history.
No high-risk fact publishes without approval.
One complete restore test and one outbound workflow succeed with receipts.
One first-party agent and one simulated client agent pass through the same safety gateway.
One protected AI task runs locally, fails closed on attempted external model access, and produces an execution receipt.

Current recommendation

One durable core, replaceable applications

The canonical operating-data layer should be the system of record for reusable property knowledge. The PMS and other SaaS products remain systems of transaction.

The first proof remains approved canonical property facts to a versioned AI guest-messaging knowledge export. The architecture underneath it must make the next destination a connector, not a rebuild.

The migration promise A move from Streamline to Guesty should become connector mapping, a dry run, reconciliation, approval, and controlled cutover. The client's data model, automations, history, and AI foundation should survive the switch.

Important naming note

“Octopai” should remain a temporary internal working name. An active data-lineage company already uses it, so legal and trademark clearance is required before any public use.