Launched AI in 3 weeks, achieving 30% of the savings target.

A checker opened a pack from another company. It was one PDF, sometimes over 100 pages. They split it by hand into ownership papers, home loans, tax records and wills. Then they read each paper against a checklist and compared the land descriptions. They marked the land's border points in a separate tool, a deed plotter. Last, they picked check codes from a list of hundreds.

The rules change by state. A warning in Florida might mean nothing in Ohio. The right check code depends on the kind of paper, the state and the deal.

It was skilled work, done over and over. The leaders wanted to cut that cost.

AI would take the repeated work. The team could then focus on the hard calls that need a person's skill. The leaders did not want AI to do it all alone overnight.

It began as a first test version. A person in charge, not a tech person, built it. It had only the part you see: no data store, no safety, no link to real orders. Ready for real use means all of those, in a tool thousands of people count on each day.

The designer listed what the tool must do and put it into Figma Make, an AI design tool. Draft screens came in hours. The team for the part you see wrote code with Cursor, Claude and Copilot, three AI coding tools. The team for the part you cannot see built features with Claude Sonnet and Cursor. In Jira, a task tracker, a custom helper built on Rovo, Jira's AI, wrote and fixed up notes on what users need. It used the project papers.

A small team worked as fast as a much bigger one. After the first tool, new features kept coming. The client now uses the same way of working in its own teams.

The people in the talks had never used AI tools. They did not know what AI could do or what they wanted from it. The talks took longer than building the first test version. They gave a list of fuzzy points. The leaders never agreed to them or put them in order.

So the team changed the order. First build the must-have parts. Then go live. Then make it better by watching real use. The leaders gave a list of features. The team helped sort it by what helps the business most. Work moved in small, clear steps.

That rule shaped every choice after it. A step runs with no check only when the numbers say it can. Years before, other companies had tried to jump straight to letting the AI do it alone. They still had not reached the results this project already has.

With legal papers, being right every time matters more than being creative. So for quality checks, the AI gets yes-or-no questions it must follow. It does not guess freely at what might be wrong. It follows a path. If a paper talks about mobile homes, it checks the rules for that kind of paper. If not, it moves on. Every path ends in a clear answer.

Checkers also had to accept the tool. At first the AI marked problems by saying “this is incorrect.” Checkers pushed back, and complaints shot up. The fix was to rewrite every note as a tip: “We see this, but your title report mentions X. Worth a look.” Same facts, a different reaction.

TitleFlow plugs into it. The help desk is being linked now. Another in-house project, SDA, comes next. One engine serves many tools, so nothing has to be built twice.

The engine works with private legal data. The safe way for the AI to read order data sits on the open internet. So the AI never gets the real keys. The system hands out short-lived passes that open only one folder. A pass lets the AI read, not change. If a pass leaks, the harm stays small.

Some sets of tasks take 10 to 15 minutes. They run in the background, so people keep working while the AI runs.

The goal was to cut cost by letting AI do the hand work. Each live skill takes over one hand step on every order. These are splitting packs, drawing borders, running quality checks and suggesting check codes.

The system was not built again. An AI layer was added. Users were moved to it step by step.

The engine is still growing. New sets of AI tasks keep going live from what real users say.