A title agent or escrow officer can use AI to turn a commitment into a table of exceptions and requirements, with each item classified, linked to the document that might satisfy it, and cross-checked against the vesting deed and payoff letters. AI never clears a requirement. The examiner reviews the source and the table, and signs off.
That last sentence is the design principle. A commitment is a promise by an underwriter, conditioned on requirements and qualified by exceptions, and every line of it carries consequences for the buyer, lender and insurer. So the job of any tool here is to make the examiner faster at reading, not to make decisions. Every check that follows is there to catch the two ways a tool can hurt you: dropping an item, and changing what an item says.
What the commitment gives you to work with
Layouts vary by underwriter, but most commitments follow a recognisable pattern: a first part describing the transaction (effective date, the interest to be insured, the proposed insured, the vesting and the land description), a part listing the requirements that must be met before a policy issues, and a part listing the exceptions that will appear on the policy unless removed or insured over. On the widely used standard commitment form these parts are Schedule A, Schedule B, Part I (Requirements) and Schedule B, Part II (Exceptions), and the form's conditions say the commitment terminates if the Part I requirements are not met within a stated period after the Commitment Date, and that the company may amend the commitment at any time. Endorsements, standard exceptions and conditions sit alongside. Check how your underwriter labels the sections, and use those labels in your prompts, because a model that does not know the layout will merge requirements and exceptions.
The information is dense but regular: numbered items, each a paragraph or two, referring to recorded instruments by book and page or document number. That regularity is why a table works well. It is also why a mistake is dangerous: one dropped line in a list of eighteen is invisible unless you count.
The company rules to settle first
- Which tool. Use one approved for non-public personal information under your company's information-security policy. If you are not sure, ask the person who owns that policy before you start. Check the retention and training terms of the plan, not the marketing page.
- What goes in. A commitment often holds the names of buyer, seller and lender, loan numbers and addresses. Redact party names and loan numbers where you can; exception wording and legal descriptions are usually fine, but check your own rules.
- Who signs off. The examiner, under your underwriter's guidelines. Write it in the procedure so there is never doubt that AI output is a working draft.
- Where the record lives. The table is saved in the file with the version of the commitment it was built from, and the date.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowStep 1: extract every item into a numbered table
Do this first and alone. Do not ask for analysis in the same prompt; extraction and judgement should be separate runs so you can verify each. The prompt below asks for a faithful table with the original wording.
The text below is the requirements and exceptions section of a title
commitment (names and loan numbers redacted). Produce two tables.
Table 1, Requirements: item number, requirement text copied
verbatim, instrument or party referred to, any date or amount
mentioned. Table 2, Exceptions: item number, exception text copied
verbatim, instrument reference (book/page or document number),
recording date if stated. Do not summarise. Do not skip any item.
At the end give the count of rows in each table.
The count at the end is the first control. Compare it with the numbers in the commitment. If the commitment lists requirements 1 to 14 and the table says 13, stop and find the missing one. Ask the model to list the numbers it captured; a gap is obvious. Also compare three random rows to the source word for word. If a verbatim column has been tidied, your later steps are standing on sand.
Step 2: classify each item so the examiner reads the hard ones first
Now add columns, as a second run on the verified table. The classification is a suggestion for ordering work, not a legal conclusion. A sample scheme:
| Class | Typical items | Who acts |
|---|---|---|
| Payoff or release | Existing deed of trust, lien, court-ordered charge | Escrow officer gets payoff letter; examiner confirms release path |
| Authority and capacity | Entity documents, trust certification, power of attorney, estate | Examiner reviews documents |
| Survey and land | Survey matters, boundary or encroachment items, easements | Examiner decides on endorsements or exception language |
| Taxes and assessments | Unpaid taxes, special assessments | Escrow officer confirms status and payment at closing |
| Recorded restrictions | Covenants, easements, rights of way | Examiner reviews instruments; buyer informed |
| Standard exceptions | Rights of parties in possession, matters a survey would show | Examiner decides on removal if offered |
| Unclear | Anything the model is unsure about | Examiner reads first |
Tell the model to use "Unclear" rather than force a fit. A good run has a few. A run with none is suspicious, because real commitments contain items that do not fit tidy categories. Re-run the classification on a commitment you have already closed and compare with what your team actually did.
Step 3: a worked example on a small residential purchase
An illustration with invented details. A purchase of a single-family property: the commitment has 12 requirements and 9 exceptions. After extraction, the table counts match. The classification puts three requirements under payoff or release (an existing mortgage, a second lien and a court-order lien against a name similar to the seller's), two under authority (the seller is a trust), two under taxes, and five under miscellaneous administrative items. Of the exceptions, five are standard, two are recorded easements, one is a covenant and one is "unclear".
The unclear one turns out to be an old oil and gas reservation, which needs the examiner's attention. The AI made the examiner's reading order better: the unclear item, the court-order lien and the trust requirements first. Nothing was cleared. The examiner's sign-off still reads the commitment itself.
Time, as a quick sum: manual transcription of 21 items into a working sheet is about 25 minutes. Redaction, extraction, count check and classification take about 12, and the examiner's read of the table against the source adds 10. Net saving is small per file but regular, and the real benefit is consistency: every file gets the same table.
Step 4: cross-check the requirements to the vesting deed
The vesting deed is the document that shows how the current owner holds title. Many problems sit in the gap between what the commitment says about vesting and what the deed says. Run a separate comparison, with the deed text (redacted) and the commitment's vesting description:
Compare the vesting in the commitment with the deed text below.
List: (1) names exactly as written in each, (2) the form of
ownership, (3) the legal description in each, (4) the recording
information. Mark each field as MATCH, DIFFERS or NOT SHOWN.
For DIFFERS, quote both versions. Do not decide which is correct.
A sample (illustrative) comes back with three MATCH rows and one DIFFERS: the commitment gives the owner's first name as an initial while the deed spells out the full first and middle names, and the legal description in the deed contains a lot number that differs from the commitment's by one digit. The first is probably a name variation; the second might be a typo or a different lot. The AI has done what it can: found the difference. The examiner decides what to do, which may be a corrected commitment, an affidavit or a further search.
Always read legal descriptions yourself. Metes-and-bounds text is exactly where a model's tidy-up (changing "N 45 degrees 30' E" into something it prefers) can do real harm. Insist on verbatim copying and compare character by character where it matters.
Step 5: match payoff letters to the requirements they satisfy
Payoffs are where money moves, so the check is strict. For each lien or mortgage requirement, build a row with the lender's name, the loan reference as shown in the commitment, the payoff amount, the good-through date and the per-diem. Then compare each against the payoff letter you received.
- Lender and reference. Does the payoff letter identify the same loan as the recorded instrument in the commitment?
- Amount and date. Does the good-through date cover the planned closing, and what is the per-diem if closing slips?
- Wiring or payment instructions. This is a human-only step. Verify any instructions by calling a known number for the lender, never one printed on the letter itself. Never ask AI to confirm bank details are legitimate.
- Release mechanism. Who records the release, and when, and is it on your tracking list?
Illustrative failure: a payoff letter shows a good-through date three days before the scheduled closing, so the amount needed is higher by three days of per-diem, say 3 times 38 equals 114. A table that holds both dates puts this next to the closing date and the AI can flag it. The escrow officer then requests an updated letter. A check that costs seconds and prevents a short payoff on closing day is worth building into your routine.
Step 6: draft the questions for the underwriter or examiner
Once the tables are verified, the model is useful for drafting. Give it the "Unclear" and "DIFFERS" items and ask for a short list of specific questions to put to the underwriter, such as "Please confirm whether exception 7 can be removed on receipt of a survey, and whether any affidavit is acceptable". Also useful: a plain-English explanation of each exception for the buyer's agent, with the caveat that it is not advice. Read these before they go out; models often over-promise ("this exception will be removed at closing"), and that is exactly the wording a title company should never put in writing before the examiner decides.
A second example: a trust seller and a lien that did not belong
An illustration, with invented facts. A commitment for a sale by a family trust carries 15 requirements. After extraction and classification, the model groups four under "authority and capacity": a certification of trust, a copy of the trust pages showing the trustee's powers, identification of the current trustee, and a statement that no trustee has been removed. It also flags one requirement as "Unclear": a release of a lien in favour of a lender "or satisfactory evidence that the lien does not affect the property".
The examiner reads the unclear item first and finds the lien is recorded against a person with the same surname as the trustee but a different middle name. The commitment wording is the underwriter's, not an error on your side, and the next step is an affidavit of identity from the seller and a request to the underwriter on whether it will remove the item. The AI's contribution was to put this item at the top of the reading list. It had no say in what happens next.
Prompts for the buyer's and lender's questions
Buyers and lenders call with questions about exceptions. A plain-English explainer can save the officer ten minutes per call, provided it is labelled and reviewed. Prompt:
Explain this title commitment exception in plain English for a
home buyer, in under 80 words. Say what a recorded easement is in
one clause, what it usually means for the owner, and what is not
known from the text. End with: "Your title officer can tell you
whether this will be removed or insured over." Do not predict
removal. Do not give legal advice.
Exception text: [paste verbatim]
A reasonable sample (illustrative): "This is a recorded right allowing a utility company to run lines across part of the property. Typically the owner can still use the land but cannot build over the strip. The text here doesn't say where the strip lies; the recorded document will. Your title officer can tell you whether this will be removed or insured over." What to fix: many models add "this is standard and nothing to worry about", which is a judgement the examiner should make. Strike it. Keep the explainers in a library and let the examiner approve each wording once, so staff reuse approved language and not a fresh model guess.
A requirements tracker that survives a long file
For files that run for weeks, convert the verified requirements table into a tracker with columns for item, class, document needed, who is responsible, date requested, date received, and examiner status (open, received, accepted, rejected, referred). The AI can pre-fill the first four columns from the table. Humans own the last three. A filled-in row (illustrative): "Req 4, authority, certification of trust and trustee powers pages, seller's lawyer, requested 3 Nov, received 6 Nov, examiner status: referred (trustee name differs from deed)". At any time the file can be read as a list of what is outstanding. Pair it with a daily 10-minute sweep of the open rows, and the long-running files stop hiding problems until closing week.
A mistake to learn from: the table that looked complete
A common failure, described generally. An officer ran extraction on a commitment that had been updated after a supplemental report. The table had 11 requirements and the count matched the original commitment. The revised commitment had 13, two added for a newly recorded instrument. The table was correct for a document that no longer governed. Because the sign-off page recorded the commitment version and date, the examiner noticed the mismatch against the revised date. The control that worked was not clever: it was a date on a form. Add the version to every table, and re-run extraction when a revision arrives, rather than editing the old table by hand.
How to pilot this on ten closed files
- Choose ten files already closed, with a mix: a simple purchase, a refinance, a trust seller, a file with a lien problem.
- Run extraction and count-check on each, timing it, and note any count that fails.
- Run classification and ask the examiner whether the "Unclear" items were the ones that took real work on those files.
- Run the vesting and payoff comparisons and see if they would have caught anything your team caught late.
- Decide: if the tool misses an item twice in ten files, do not use it live until the prompt is fixed.
Ten files take a day or two spread over a fortnight and tell you more than any demo. Keep the results as a one-page note: counts matched, errors found, minutes saved, and whether the examiner would use it.
The sign-off page
Add one page to the file for every commitment reviewed this way. A filled-in example (illustrative):
File: [number] Commitment version/date: [date]
Tool: [assistant and plan] Run by: [initials] Date: [date]
Item counts: commitment 12 req / 9 exc; table 12 req / 9 exc: MATCH
Verbatim spot-check (3 rows): passed
Vesting cross-check: 1 DIFFERS (lot number) referred to examiner
Payoff cross-check: 1 good-through date short of closing; updated
letter requested
Items AI classified Unclear: 1 (exc 8, mineral reservation)
Requirements cleared by AI: NONE (AI does not clear)
Examiner review and sign-off: [name, date]
The line "Requirements cleared by AI: NONE" looks odd, but it is a good habit. It forces the reviewer to confirm the point on every file, and it shows an auditor or underwriter the control worked.
Where AI misreads commitments
- Dropped items. Caught by the count. Always count.
- Merged exceptions. Two separate recorded instruments combined into one row. Check that each instrument reference appears separately.
- Normalised text. "Lots 4 and 5" turned into "Lots 4-5". Spot-check verbatim columns.
- Over-confident classification. A restrictive covenant labelled "standard". Classification is a reading order, not a finding.
- Wrong version. Table built on a superseded commitment. Record the commitment version and date on the table.
- Stale tables. An updated commitment arrives and nobody reruns the comparison. Treat every revision as a new file for this process.
Using this with the rest of the closing file
The exceptions table feeds other work. The escrow officer's checklist can pull its payoff and tax lines from it, as in the tutorial on inbox triage and closing checklists, and the pre-closing numbers can be reviewed using the method in the closing disclosure review checklist. If you want a wider view of where AI helps in a title company and where it does not, six jobs and their limits is the broader piece, and the lawyer-side tutorial shows how the same table is used by a lawyer who is doing the legal review.
Keep one more habit alongside the tables: a short list of prompts that have failed on your commitments, with the date and the fix. Underwriter layouts differ, and a prompt tuned on one form may merge sections on another. When you start with a new underwriter's form, run it through the ten-file test again before trusting it on a live file, and ask the underwriter whether they publish guidance on how commitments should be read.
One last rule worth putting in the procedure manual: if the tool and the examiner disagree, the examiner wins, and the disagreement is noted so you can improve the prompts. A week of such notes will teach you more about the tool's weak spots than any vendor demonstration.
Questions title agents ask about AI and commitments
Can AI tell me which exceptions will be removed?
No. Whether an exception is removed, narrowed or insured over is a decision for the underwriter or examiner under their rules. AI can list each exception, classify it and propose what document might address it. It must never mark one as cleared.
Should I paste a full commitment into a chatbot?
Only with a tool your company has approved for non-public personal information, and ideally after redacting names, loan numbers and identity data. Property descriptions and exception wording rarely need personal details. Check the tool's retention and training terms first.
How do I check the AI did not drop a requirement?
Count them. Compare the number of numbered requirements and exceptions in the commitment with the rows in the table, item by item. A missing row is the most common failure, and a count takes under a minute.
Does this replace the examiner?
No. It saves time on transcription and cross-checking so the examiner spends their attention on judgement. The examiner reviews the source commitment and the table, and signs off under your company's procedure.
Further reads
- Real Estate AI Chatbot Demo: A Test Checklist Before You Sign — A test checklist before you sign for any AI tool.
- Title Insurance AI: Pick a First Pilot and Price It for an Agency — Pick and price a first AI pilot for a title agency.
- ALTA Best Practices and AI: What a Small Title Agency Should Document — What to document about AI use in a small title agency.
- Best AI Tools for Commercial Real Estate Brokers, Compared by Job — AI tools for property professionals, by job.
- Escrow Instructions With AI: Draft From the Contract, Then Check — Draft escrow instructions from the contract, then check them.
- AI Answering Service for Title Companies: Cost and Status-Call Rules — Status calls and answering-service rules.
- Signing Agent Software and AI: Scheduling, Quotes, Follow-Up — Scheduling and follow-up for signing agents.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the standard 2021 commitment for title insurance form (Schedule A; Schedule B, Part I Requirements; Schedule B, Part II Exceptions; commitment conditions), read 4 November 2026; layouts vary by underwriter; AI provider data-use pages; the author's workflow design. Property, party and figures in examples are invented and illustrative. Follow your underwriter's guidelines and your company's procedure.