AB 1466, signed into law in 2021, requires California county recorder offices to identify and redact racially restrictive covenants in their recorded property documents. San Francisco's corpus spans 7.4 million pages across 11,650 record books. Fewer than 0.4% contain covenant language — a needle in a haystack at enormous scale.
The challenge
Detection is only the first step. Every covenant requires:
- Redaction of the restrictive language
- Metadata extraction (instrument number, recording date, grantors, address)
- A cover sheet with statutory citations
- County counsel approval
- Rerecording into the official record
This downstream work demands scarce, high-skill roles, including county counsel. Vendor contracts for AB 1466 compliance have ranged into the mid-seven figures. Many counties have had staff manually review tens to hundreds of thousands of deeds just for detection.
What we built
Mesa for Public Records handles the full AB 1466 compliance pipeline: from millions of raw deed scans to compliant, rerecorded redaction packages. The pipeline has four phases.
- Detection. A machine learning model scans every page, flagging those likely to contain covenant language. In San Francisco, it identified roughly 15,000 covenant documents out of 4.7 million pages.
- Processing. Each flagged document goes through automated OCR, page segmentation into multi-page deeds, extraction of structured fields, and precise redaction boundary generation.
- Review. Staff open each prepared "redaction packet" in a web interface: an AI-drafted cover sheet on one side, the document with redaction overlays on the other. When the AI gets it right — more than seven in ten packets — a reviewer confirms in under 30 seconds. Corrections typically take under two minutes.
- Approval and export. County counsel reviews approved packets in bulk. The system generates compliant exports for rerecording.
Individual field accuracies range from 90% to over 99%.
AI prepares, staff decides. The system amplifies the expertise of the people using it — handling mechanical work so staff can focus their judgment where it matters most.
Results
62 person-years of manual work, completed in 9 person-weeks. If you hired someone for this job the day they graduated college, they'd retire before finishing. San Francisco is doing it with nine part-time staff in two quarters.
Nine staff members — eight preparers and one county counsel reviewer — are processing the entire backlog. Reviewers complete 59 packets per active hour, with top reviewers regularly exceeding 90.
The collaboration
San Francisco's team brought deep expertise in recording conventions, historical document formats, and edge cases that only experienced staff recognize. As they worked, the system analyzed patterns in their corrections and adapted. More than seven in ten packets now require no staff intervention at all.
The system gets better because of the people using it. Each improvement cycle frees staff to spend more time on complex cases, which generates richer signal for the next round.
Beyond AB 1466
We're past the era where "AI for government" means autocomplete and keyword search.
The same architecture that powers Mesa for Public Records on AB 1466 — agents that apply legal rules, make contextual distinctions, and handle full workflows with human oversight — applies across assessor-recorder operations.
Records requests and intelligent redaction
Beyond AB 1466, Mesa for Public Records handles California Public Records Act requests across documents, video, and audio.
- Legal reasoning, not pattern matching: redaction driven by playbooks that encode your office's policies and statutory exemptions
- Context-aware: distinguishes private cell phones from agency lines, protected health information from routine correspondence
- Consistent at scale: enforces the exemptions and policies your office already follows
Legal and regulatory research
Mesa for Research delivers cited answers from statutes, regulations, BOE guidance, and AG opinions, producing finished work product with full citations to primary text.
Change-in-ownership processing
Change-in-ownership determinations are among the highest-volume, most document-intensive workflows in any assessor's office. Every recorded transfer requires reading deeds and trust documents, applying Revenue and Taxation Code rules, and determining whether a reassessment event has occurred. Prop 19 made this significantly more complex. AI agents can handle the mechanical parts — reading documents, extracting fields, applying routine rules — so staff can focus on the transfers that genuinely require their expertise.
Intelligent records collections
County recorder offices sit on decades of documents, maps, and property records that are searchable only by index lookup. Mesa makes these collections:
- Semantically searchable: ask questions across your archives, not just keyword lookup
- Connected: link deed records to GIS and surveyor data, surface patterns across collections
Proactive analytics
Scheduled AI analysis across your workflows: trends in recording volume, anomaly detection, automated quality checks. A dedicated analyst continuously reviewing what flows through your systems and surfacing what matters, without waiting to be asked.