Our AI
AI built for work somebody has to sign
Most AI products ask you to trust the answer. In surveying and civil engineering that is the one thing you cannot do: the answer goes under a seal, and the seal is yours. So we built our AI around three rules that let a licensed professional check it rather than trust it. Here they are, in the plain version, along with what we deliberately do not ask the AI to do.
The three rules
The AI reads. It never computes.
A language model is good at finding and transcribing words on a messy page: the grantor and grantee, the calls in a legal description, the comments in a review letter, the requirements in an ordinance, the due date in a solicitation. It is not good at arithmetic, and we do not let it try. Every bearing, distance, curve, closure, area and coordinate is computed by ordinary arithmetic from the text as it was read, on every setting, every time.
It quotes, or it is discarded.
Everything the AI reports has to point at the passage of your document it came from, and that passage is shown beside the finding. A finding the AI cannot tie to the page is thrown away before you see it. This is a property of the software, not a habit of the team, and we test for it.
It flags. It never fills in.
A reading the AI could not make cleanly — a smudged bearing, a clause that could be read two ways, a sheet with no scale — is marked as uncertain and handed to a person. The AI is not allowed to guess and present the guess as an answer. That silent kind of wrong is the one that costs a licence, and it is the failure we spend most of our care on.
Your AI provider, your key, your terms
CivitasOPS is not tied to one AI vendor. A firm administrator connects the firm's own account with Anthropic, OpenAI or Google, and the firm's documents are read under the firm's own agreement with that provider — never on our account unless your firm is on the Team or Premium plan and chooses the CivitasOPS key, in which case they are read on our vendor account under our agreement with that vendor, and the menu names the vendor. Switch whenever you choose; the choice is recorded against the person who made it, and only the key's last four characters are ever shown again.
This matters for two reasons. The first is contractual: your client's security review asks who sees the text of a deed, and the answer is a vendor your firm chose and signed with, not one we chose for you. The second is practical: the vendors leap-frog one another every few months, and a firm should be able to follow the best one without changing software.
What we will not claim
We will not tell you the AI cannot be wrong. The arithmetic is exact, but it is arithmetic on the text as it was read, and a reading can be wrong: a bearing transcribed from a bad scan, a clause mis-located on a crowded page. That is why every finding is quoted, every uncertain reading is flagged, and the output is research support for a licensed professional's review — not a survey, a legal description, a title opinion or a permit determination. Responsibility for sealed or filed work stays with the licensee, and we built the product so that it can.
We publish what the AI gets wrong. The blog is where we write about the readings that failed, what the flag looked like, and what we changed. If you have a document that broke it, we would like to see it.
Where your documents go
The privacy page names every provider that receives text, what each one receives, where it is hosted and for how long it is kept. Two answers are still not yet: there is no service-level agreement and no third-party audit, and we will say so in writing rather than in a contract.
Watch the AI work on a parcel you know
Bring a deed you have already run. You know what the calls should be, so you will know inside ten minutes where the AI was right, where it flagged, and whether it earns its seat.