PrecisionDocs

Designing for land development: why trust is the product

In an industry that stakes professional licenses on research accuracy, the AI that hallucinates is the AI that gets fired. A note on the design decisions behind citations, audit trails, and public-source-by-default.

· 7 min read · PrecisionDocs Team

A licensed engineer stamps a report. If the report is wrong, the stamp is wrong, and the license is exposed. The liability is personal. It follows the engineer from firm to firm. It does not transfer to the AI that helped produce the report.

That is the starting condition for any AI tool built for civil engineering or land development. The tool is not the professional. The professional is the professional. The tool has to make the professional's job defensible, not easier.

This is why we built the product around trust instead of speed.

Hallucination is malpractice, not a quirk

When a general-purpose model is asked about a zoning ordinance in a small Midwestern city, it will often produce a confident answer that is not grounded in that city's actual code. The answer reads like a quote. It is not a quote. The model is completing text, and text about zoning tends to follow predictable patterns.

In a consumer context this is annoying. In a professional context, it is a liability event. A planner who sends a client a memo quoting a setback that does not exist has a problem that does not go away with an apology.

The design response is not "make the model better." The design response is: refuse to answer without a source. If the agent cannot locate the ordinance section, the agent says it cannot locate the ordinance section. The gap is visible. The professional sees it and knows to pick up the phone.

Citations are load-bearing

Every factual claim the agent makes links to the source that produced it. Ordinance citations link to the specific chapter and section of the municipal code. FEMA claims link to the specific FIRM panel and date. Soil claims link to the specific USDA Web Soil Survey run that generated them.

This is not a marketing feature. It is the feature. If a claim is not cited, the developer has no way to audit it. If the developer cannot audit it, the developer cannot stake a deal on it.

The latency cost was real. Adding a retrieval step and a verification step to every tool call made the product slower than a straight generative loop. We accepted the cost. The audience we serve needs the citation more than they need the speed.

Public source by default

Wherever possible, the agent pulls from the public sources a land development professional already trusts. FEMA Map Service Center for flood data. USDA Web Soil Survey for soils. USGS National Map for topography. State DOT portals for road classifications. County GIS for parcel geometry. Municode and AmLegal for municipal ordinances. The agent does not invent a proprietary "flood score." It returns Zone AE with a BFE of 543.2 and a panel number. Anyone on the team can verify the answer with twenty seconds and a browser.

The philosophical reason: a product that trains professionals to trust its internal numbers is a product that degrades professional judgment over time. A product that trains professionals to trust the public record makes the record easier to read, without displacing the record. The record outlasts the product.

The audit trail is a first-class object

Every research session produces a chronological record of the questions asked, the tools called, the sources consulted, and the citations returned. The record is exportable. It is the thing a project manager can hand to a reviewer, a client, or a planning board if the basis for a decision is ever questioned.

We treat the audit trail the way accounting firms treat their working papers. It is not optional. It is not a nice-to-have. It is the evidence.

What we refused to build

We refused to build a one-click "auto-fill this variance application" feature. Not because it was technically hard. Because the output would move into a legal filing with the professional's name on it, and the professional would not have read every word of the application before signing. That is a class of failure we are not willing to normalize.

We refused to build "confidence scores" visible to end users. Internal telemetry, yes. User-visible "this is 84% correct" gauges, no. A professional who trusts a number because the interface told them to trust it is a professional who has outsourced judgment to a UI element.

What the industry has the right to demand

A tool that knows what it does not know. A tool that cites every factual claim. A tool that never replaces a stamp. A tool that survives an audit.

We are not there on every feature yet. But that is the design contract. Everything we ship is measured against it.

More insights