Why we built MAIA’s Yardi connector before we had a customer.

The standard early-stage AI playbook is to ship the model first and wire integrations in later, once a customer is paying and someone is on the hook to make it work. At MAIA we did the opposite. Before any operator had signed anything, our team built a production-grade Yardi connector, then ServiceNow, then MS Graph, then SAP, then MRI.

Five enterprise integrations on a sandbox-default runtime. No design partner waiting on us. No paying customer. Just an architectural bet I was willing to make because I’ve spent fifteen years on the operator side of these systems, and I know what makes a vendor demo land or die.

The bet

In operational AI, the integration is the moat. Models are commoditizing on a six-month cycle. The thing that doesn’t commoditize is the work of getting a decision out of a real system of record, with the right auth, the right tenancy boundaries, the right idempotency, the right write-back semantics, and crucially, the right lineage all the way back to the originating signal.

Operators don’t buy AI on the strength of a demo video. I never did. I bought when I watched a system call into my tenant, pull my ledger, reason on it, and write a decision back that my auditor could trace. Until that loop closes, every conversation is hypothetical. After it closes, every conversation is a procurement conversation.

Why Yardi first

The two products MAIA ships today are Workforce and Buildings. Buildings runs the largest category of operational decisions we handle, work orders, turnover, arrears, energy, compliance, and Yardi sits underneath a meaningful share of the North American multi-family and commercial portfolio market. If we couldn’t move data through Yardi cleanly, we couldn’t credibly say we ran portfolios. So we made it the first integration, built before the second product had a roadmap line item asking for it.

What it took

More work than the line item suggests. OAuth client-credentials and password-grant flows, plus SAML-bearer for the enterprise tenants that won’t do anything else. Tenant-scoped auth with no credential leakage between portfolios. Sandbox-default config so the connector boots and runs against a synthetic tenant when no client ID is set, meaning the demo flow works on a laptop, the real flow works in production, and the surface is identical. Idempotent writes with full lineage threaded from a raw Yardi GL line all the way up through detection, decision, action, and back into the decisions ledger as a single auditable chain.

None of this is glamorous. It is the part of the work that does not show up in a launch tweet. It is also the part that determines whether the next conversation is a pilot conversation or a politely archived email thread.

The first time it worked

The first time we ran the Yardi flow live for a partner, pulling their actual arrears tape, surfacing a top-10 risk list with reasoning, and writing a recommended decision into the ledger, the conversation changed instantly. We had spent twenty minutes preparing to defend the model. We never got the question. The question was about deployment timeline.

This is the part the standard playbook misses. The bottleneck on AI adoption in operations is not model quality. It is institutional trust, and institutional trust is built by doing the integration work upfront, before the customer asks, so they don’t have to.

What we’d do again

Build the integration before the demo. Build the audit trail before the dashboard. Build the sandbox-default before the production credentials. Operators reward systems that are already standing up when they walk in the door, and they discount systems that need to be assembled in front of them. I know because I’ve been the operator on the other side of both kinds of meeting.

We’ll keep posting these as we make them. The changelog is the receipt; the research is the why; the integrations are the moat.

If you’re an operator running a multi-site portfolio and any of this lines up with what you’re seeing, we’d like to hear from you. Request a briefing , thirty minutes, founder-led, no deck.