National Compliance & Legal Services Firm

From day one, the mandate was not “build a one-off app per client.” Linkt built a multi-tenant platform in which each of the firm's clients gets a fully isolated deployment — its own database, its own AWS account, its own frontend on its own subdomain — while the actual product capability lives in roughly a dozen shared libraries that every deployment pulls from.

The consequence is structural: every time something is built for one client, the reusable part is extracted into a library, and the next client inherits it essentially for free. Shared foundations now include document ingestion with OCR fallback, a unified LLM access layer with retries and tracing, schema-validated structured outputs with an automatic repair loop, PII detection, multi-tenant single sign-on including enterprise SAML, and Terraform modules that turn a new client's entire cloud footprint into a repeatable exercise instead of a project.

The payoff is already measurable: the second deployment of the document intelligence product line shipped as a near-pure structural clone of the first. What was a multi-month build the first time was mostly configuration the second. New deployments now take weeks, not months.

Contract lifecycle management for a major law firm: AI extraction pipelines pull obligations out of contracts — due dates, survival clauses, automatic reminder ladders — with amendment conflict detection that catches ambiguous supersession language and value contradictions across contract families.

Legal document intelligence for two commercial real estate clients: redline-aware ingestion preserves tracked changes and margin comments, attorneys get section-aware, cited answers, and documents under negotiation are reshaped into a current-versus-proposed view so the AI reasons about the diff the way a lawyer would.

Certificate of insurance compliance review: certificates are analyzed against a hierarchical requirements model and presented as evidence in a split PDF-and-scorecard view, with an audit-grade Excel workbook generated on the way out. Verdicts are computed deterministically from extracted values, never fabricated by the model. The AI extracts; the rules decide.

Automated contract-vs-CRM validation: opportunity snapshots stream in via webhook and are continuously reconciled against contract field requirements — a permanent check on the gap between what sales says and what the contract says.

Plus SLA compliance tracking off live ticket flow, an AI phone intake agent for insurance incident calls backed by a full evaluation harness, and a direct-to-consumer wellness application in a different vertical — proof the architecture flexes well beyond legal and compliance.

Since going live, the validation product has completed 962 validation runs against 1,083 CRM opportunity snapshots, performing 123,583 individual contract field checks — every one a field a person would otherwise have verified by hand. At the current run rate, that is 400,000+ automated checks a year.

At a deliberately conservative 30 seconds per manual check, the system displaced 1,000+ hours of review work in its first 14 weeks — roughly 3,500 hours a year, or $350,000–$500,000 in annual review capacity at blended review rates. Real-world review time per field typically runs 1–2 minutes, not 30 seconds; at realistic times, the system is on track to displace over $1M a year in manual review capacity.

Just as important is what happened to the review model. Roughly 20% of runs pass clean. The remainder arrive with the specific mismatched fields called out and the evidence already attached. Humans stopped checking everything and now adjudicate exceptions.

As the fleet grew, knowing “what is going on with client X” meant an hour of digging across five systems. Linkt built an internal operations platform with an AI assistant on top: client health checks went from ~60 minutes across five systems to one question answered in seconds with every figure cited; weekly client reports became a generated artifact a human reviews; per-client cloud cost attribution went from nobody's job to automatic; and incident awareness became a real-time SLA board. Current status: four clients under automated monitoring, zero open incidents.

Everything is infrastructure as code. Isolation runs to the AWS account level per client, single sign-on is federated with permission sets defined in Terraform, credentials are short-lived everywhere, and every protected route writes an audit row. SOC 2 preparation runs the same way: evidence documents generated from cited sources and live, read-only cloud state, with a standing rule that nothing aspirational goes into an artifact.

The economics of this engagement improve with every release. Client one of a product line pays for the product; client two inherits it as configuration. Three shared libraries extracted in a single modularization wave were consumed directly by the very next client build. The firm's marginal cost of its ninth product will be a fraction of its first.

The firm didn't buy software. It acquired a product and engineering organization — one that gets faster with every product it ships.