Cheap Model for the Platted Resale, Strong Model for the Probate Chain — and Who in Your Agency Draws That Line

Which of the ninety files you opened last month actually needed your best examiner?

Ask that at a Monday meeting and you get an argument, which is the point. Everybody in a title agency knows that a platted single-family resale with a prior policy from 2025, one payoff, individual vesting and no survey is not the same animal as a probate chain with three heirs, a foreclosure eleven years back and a mineral reservation nobody has looked at since 1974. Nobody has written the line down. In 2026 you have to, because that line is now also a cost decision every time an AI agent picks up a file.

What a 90,000-person rollout has to do with a six-person agency

Cisco is rolling a personal AI agent out to roughly 90,000 employees, and one of the design choices they built in from the start is routing: the routine request goes to a fast, cheap model and only the genuinely hard one escalates to the expensive one. That is not a big-company luxury. It is the same instinct you already use when you decide which files your newest searcher can handle and which ones go to the examiner who has been doing this since 1998.

Model routing in a title plant means deciding, before any work starts, whether a file is ordinary enough to be handled by the cheap fast option with a human spot-check, or unusual enough that it goes straight to the expensive one and then to a licensed examiner — and writing that test down so the decision is not made file by file under deadline pressure.

The economics are what make this newly worth your attention. Capable models cost roughly a tenth of what they did in 2025, which means the cheap tier is now genuinely cheap — pennies a file — while the strong tier is still meaningful money at volume. Route badly and you either pay the strong price on every routine resale, or you send a probate file to the cheap tier and get an answer that reads confidently and is wrong.

Where the line sits on a residential order

Here is a workable first draft. Treat a file as routine only if every one of these is true:

• Single-family or condominium, inside a recorded plat, one parcel, no metes-and-bounds description.
• A prior policy or starter in your plant less than about three years old, on the same legal description.
• Vesting in one or two natural persons, taking title by warranty deed, no trust and no entity.
• One institutional mortgage to pay off, with a payoff obtained through the servicer's own portal.
• Purchase price within your agency's ordinary liability limit, so no underwriter approval is required.
• Financed purchase — because a cash sale to an entity or trust brings a reporting obligation with it under the FinCEN residential real estate reporting rule that took effect on 1 March 2026, and that is never routine.

Fail any single line and it escalates. Not "probably escalates" — escalates. The value of the list is that a coordinator on her third week can apply it without judgment, and judgment is exactly what you do not want being exercised at 8:15 on the last Tuesday of the month.

flowchart TD
  A["Order opens in the closing software"] --> B{"Six routine tests all pass?"}
  B -->|Yes| C["Cheap fast model: read search, draft commitment"]
  B -->|No| D["Strong model: chain, curative, exceptions"]
  C --> E{"Cheap model reports low confidence or a surprise?"}
  E -->|Yes| D
  E -->|No| F["Examiner spot-checks 1 in 10"]
  D --> G["Licensed examiner reviews every file"]
  F --> H["Commitment issues"]
  G --> H

The files that escalate, every time, without discussion

Every agency's list differs slightly by state, but this one holds up almost everywhere. Post it where the order entry happens.

• Probate, heirship, or any deed signed by a personal representative, guardian or conservator.
• Foreclosure, deed in lieu, or a tax deed anywhere in the last twenty years of the chain.
• Mechanic's or materialman's liens, or visible recent construction on a property with no notice of commencement resolution.
• Severed mineral or water rights, which in Texas, Oklahoma, Colorado and the Dakotas is routine enough in volume and never routine in analysis.
• Access questions — landlocked parcels, prescriptive easements, a driveway crossing the neighbour's lot.
• Entity or trust vesting on either side, and every non-financed transfer that triggers a report.
• Divorce, bankruptcy, or a party under a power of attorney.
• Anything commercial, multi-parcel, or part of a 1031 exchange with a December 31 identification clock on it.
• New construction on a lot split out of a larger parcel where the plat has not been recorded.

Who owns the list — and it is not your software vendor

This is the part owners get wrong. The routing test is an underwriting decision, so it belongs to whoever holds underwriting authority in your agency: your underwriting counsel if you have one, otherwise your senior examiner or agency manager, reviewed with your underwriter's agency representative and re-read whenever a bulletin changes. It goes in writing, dated, with a version number, and it is reviewed quarterly.

It is not a setting your closing software vendor picks by default and it is not something an escrow officer adjusts on a busy afternoon because the strong option is taking too long. If the list needs an exception, the exception is granted by the person whose name is on the authority letter, and the reason is noted in the file. That single discipline is what turns routing from a cost trick into something you can show a claims examiner two years later.

A cost sheet for a ninety-file month

Everything below is illustrative, using publicly discussed 2026 pricing levels rather than a quote for your agency. The shape of the answer is what matters, not the cents.

Look at what that table actually says. The model spend is trivial either way — a rounding error against one examiner's morning. The real money is the examiner minutes, and routing earns its keep not by saving fifty dollars on models but by making sure the eight-minute files genuinely get eight minutes and the thirty-four-minute files are identified before somebody promises a commitment by Thursday. If your routing choice is being sold to you primarily as a cost saving on the AI itself, the vendor is measuring the wrong column.

The failure mode that will actually bite you: a routine file that wasn't

The six tests are applied at order entry, on the information in the sales contract and the prior policy. Facts arrive later. The payoff comes back showing a second mortgage nobody disclosed. The search turns up a 2016 quitclaim from a name that does not appear in the contract. The buyer changes vesting to an LLC eight days before closing, and now you have a reporting question on a file you scoped as ordinary.

So build the reversal into the flow: any file can be promoted upward at any point by any person who touches it, and promotion needs no approval — only demotion does. Then track how often it happens. A promotion rate creeping past roughly one file in ten means your entry tests are too loose and it is time to tighten a line, usually the prior-policy age.

The second failure mode is quieter. A cheap-tier draft that reads smoothly makes an examiner less likely to look hard. Counter it the way auditors do: plant a known error in one file a month and see whether the review catches it. If it does not, your spot-check rate is theatre and you need to raise it.

Frequently asked questions

Do I have to choose which model to use, or does the software do it?

You choose the rule; the software applies it. What you should refuse is a product that routes silently by its own judgment with no record of which tier handled which file. When a claim comes in on a file from eighteen months ago, "the system decided" is not an answer you want to give your underwriter.

Our refinance volume swings wildly with rates. Does routing still hold up?

It holds up better under a wave, not worse — refinances skew heavily routine, so a defined routine lane is exactly what lets a small team absorb a doubling of volume without hiring. What breaks under a wave is discipline: the pressure to widen the routine definition so more files qualify. Freeze the tests before the wave, not during it.

Is there a risk in running the cheap model on a file that later becomes a claim?

The risk is not which model read the search; it is whether a licensed person reviewed and signed the commitment. Your policy is issued on your examiner's judgment either way. Keep the record showing the tier, the draft, the review and the reviewer's name, and the model choice becomes a workflow detail rather than a liability question.

How do we build the first version of the routing list without stopping work?

Take last quarter's closed files, sort them by how many days elapsed from order to clear-to-close, and read the twenty slowest. The reasons those files were slow are, almost always, your escalation list already written by reality. Draft it in an afternoon and let your underwriting counsel argue with it.

A closing thought on the other queue. The same routine-versus-hard split applies to the phone, where roughly the same seventy per cent of calls are "has it recorded yet," "when is my signing," "who do I make the check out to." CallSphere builds AI voice and chat agents that handle those calls and web chat around the clock and book signing appointments, passing anything about a payoff figure, a Schedule B question or a disbursement to a named person on your team. It answers the routine so your escrow officers can stay on the thirty per cent that needs them.