Conveyancing AI Software: What a Fee Earner Should Test in a Demo

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Conveyancing AI Software: What a Fee Earner Should Test in a Demo.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Conveyancing AI Software: What a Fee Earner Should Test in a Demo.

Ask five things before buying: does it connect to your case-management system without re-keying, where is client data processed and kept, can you export a full audit trail of what the AI did on each file, is pricing per file or per seat on your real volumes, and can it handle one of your own awkward closed files in the demo.

Vendors demonstrate on tidy sample matters. A small firm's real files have scanned PDFs, a seller who changes solicitor, a lender's instructions that arrive late and a mortgage offer that contradicts the contract. The demo that matters uses one of those, with the fee earner, not the salesperson, at the keyboard.

Follow me on Instagram@sagnikteaches

Run the demo on your file, with you driving

Before the call, choose a completed purchase from the last twelve months and remove client names, addresses and account details. Pick one with at least one problem: a missing management-pack document, a late change to the completion date, an enquiry that took three rounds. Ask the vendor to load it, or load it yourself in their sandbox, and tell them in advance. A vendor who cannot accept your file in a demo will have trouble with it in practice.

Connect on LinkedInSagnik Bhattacharya

Take notes in a simple scorecard so that two fee earners can compare views afterwards. One line per test, a score from 0 to 3, and a column for "what the vendor said" against "what I saw". The gap between the two columns is the information you are buying.

Subscribe on YouTube@sagnikteaches
AreaTest in the demoPass looks like
IntegrationOpen a matter in your case-management system and trigger the AI feature from thereOutput lands on the file with no copy and paste
Data locationAsk where documents are processed, stored and backed up, and which sub-processors touch themNamed locations and a written sub-processor list
Audit trailAsk to see who ran what, when, on which document, and which version of the modelAn exportable log tied to the matter number
AccuracyRun it on a messy scan and a document with an inconsistencyIt flags the inconsistency or says it cannot read the page
PricingRequest a quote on your monthly file count and headcountA total, with what triggers extra charges
ExitAsk how you get your data and logs out if you leaveA stated export format and timescale
The Complete Excel Course with AI Integration thumbnail

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 now

Integration with your case-management system

The single most common regret after buying is a tool that works well in its own window and nowhere else. Each copy and paste is time, and each transfer is a place where a figure gets mis-keyed or a document lands on the wrong matter. Ask three questions and watch them being answered rather than hearing them described.

  1. Is the connection native, through an API, or a manual upload? Ask the vendor to name your case-management product and show a customer using it. "Works with anything via a spreadsheet export" means manual upload.
  2. What flows back? A summary text is the minimum. Better: the extracted dates, parties and figures arrive in the right fields, and the generated letter is saved in the matter's documents with the right name.
  3. What happens when the connection fails? Switch off the network in the demo, or ask what the log shows if a sync fails. A product that fails silently is worse than one that fails loudly.

Time the whole path for one task. Say the task is turning the seller's replies to enquiries into a client report. In the demo it takes four minutes with the AI feature, against your usual twenty-five. If the integration forces eight minutes of exporting and uploading, the real saving is smaller than the headline. Write both numbers on the scorecard.

A worked sum on the integration gap

Illustrative figures. A four-fee-earner firm completes around 70 purchases and sales a month. The tool saves 12 minutes per file on enquiry reports when integrated, but only 5 minutes when files must be exported and re-uploaded by hand. At 70 files, that is 14 hours saved a month against 5.8. If a fee earner's time is worth $45 an hour to the firm, the integrated route is worth about $630 a month and the manual route about $262. A product costing $400 a month wins in one case and loses in the other, and the integration question decided it, not the AI.

Where the data goes, and who can see it

Conveyancing files contain identity documents, bank details, source-of-funds evidence and sometimes health or family information. Ask the vendor for answers in writing, because a verbal answer in a demo does not bind anyone. The questions that matter are narrower than a general security questionnaire:

  • In which region are documents processed by the AI, and in which are they stored? These can differ, so ask about each.
  • Which other companies handle the data? Ask for the sub-processor list, including whichever company supplies the underlying AI model.
  • Are our documents used to train or improve any model, for you or for the model supplier? Ask for the contract clause, not the marketing line.
  • How long is a document kept after the file closes, and can we set the period ourselves?
  • Can a client's data be deleted on request, and how do you show it has gone, including from backups and logs?

Compare the answers with what your own client documents promise. If your privacy notice says documents stay in a particular region and the vendor processes them elsewhere, you have a gap to close before signing, not after. Where in doubt, ask your data-protection adviser or insurer, and check the current guidance from your regulator on using AI with client data. Do not assume a product built for lawyers has already cleared this for your firm.

What a good written answer looks like

Here is a compact illustration of a reply that would pass, and one that would not. A passing reply: "Documents are processed and stored in the region your account is set to. The AI model is supplied by [named company] under a data processing agreement that prohibits training on customer content. Retention defaults to the file closure date plus the period you set. A deletion request removes content and derived text within 30 days, with a confirmation record." A failing reply: "We take security extremely seriously and use industry-standard encryption." The second says nothing you can check. Ask the follow-up until the reply names a place, a company, a period and a clause.

The audit trail you will want when something goes wrong

Treat the audit trail as the feature that decides whether you can defend the firm after an error. Suppose a lender complains that a charge was missed in a title summary. You need to know what the AI was shown, what it returned, whether a person edited it and who signed it off. Ask the vendor to produce this live, for the closed file you loaded.

Log fieldWhy it matters
Matter number and userTies each AI action to a file and a named person
Date and timeShows what was known when
Document names and versions usedShows which version of a contract the summary came from
Output as generatedPreserves the original before any human edit
Human edits and approverShows the fee earner checked it
Model or feature versionExplains why the same file gives different results later

Ask whether the log can be exported in a form you can store with the file, and for how long the vendor keeps it. A log that disappears after 30 days does not help a complaint raised eighteen months later. Also check that the tool makes a human approval step easy: a button that stamps "reviewed by" with the fee earner's name is worth more than a long report nobody opens.

Accuracy tests that take ten minutes

Run three small tests and keep the outputs. None needs technical knowledge.

Test 1: the inconsistency. Take a contract and mortgage offer that disagree on a detail, for example a completion date or a named party. Ask the tool to compare them. Pass: it flags the mismatch. Fail: it produces a smooth summary that treats them as consistent.

Test 2: the bad scan. Load a skewed, low-resolution page, such as a faxed management pack. Pass: it says some text was unreadable and shows you where. Fail: it fills the gap with plausible figures. A tool that guesses is the most dangerous kind.

Test 3: the question with no answer. Ask about something the documents do not contain, such as "What is the ground rent review date?" on a freehold file. Pass: "not found in the documents provided". Fail: an invented date.

Here is an illustrative result sheet from a three-test run on one product, and what a fee earner would conclude:

TestResultVerdict
InconsistencyFlagged the completion date mismatch, missed the party name spelling differenceUseful first read, not a check
Bad scanMarked two pages "low confidence", still summarised themAcceptable if the flag is visible to the reader
No answerSaid "not stated" for ground rent reviewPass

The conclusion for this product would be: it can speed up the first read of documents, so long as the fee earner still reads the original on anything marked low confidence. That is a sensible purchase if the price fits. It is not a replacement for the title check, which the tutorial on reports on title covers in detail.

Per file or per seat: price it on your numbers

Pricing models differ, and the cheaper one depends on how your firm works. Ask each vendor for a written quote built from your own figures, and ask what happens in a busy month. Do not accept a headline price alone, because several vendors add charges for extra document pages, additional users, integrations or support tiers.

ModelSuitsWatch for
Per fileLow or uneven volume; firms trying a tool on a few mattersCost climbs in busy months; charges per document or page
Per seatSteady volume with a few heavy usersPaying for seats that nobody opens; minimum seat counts
Bundle with case managementFirms already buying a new systemLock-in; hard to price the AI part on its own
Usage creditsOccasional use by many staffCredits that expire; unclear what counts as one use

A worked pricing comparison

Illustrative figures, not vendor prices. A firm with three fee earners and two case handlers opens 60 files in a normal month and 95 in the busiest. Vendor A charges $9 a file. Vendor B charges $65 per seat a month with a minimum of three seats.

  • Vendor A: 60 files cost $540 in a normal month, 95 files cost $855 in the busiest. With nine normal months and three busy ones, the year comes to about $7,400.
  • Vendor B with three seats (the fee earners only): $195 a month, $2,340 a year, but the case handlers cannot use it, so they keep doing the first-read work by hand.
  • Vendor B with five seats: $325 a month, $3,900 a year.

Per-seat is cheaper on paper in this case, but only if the five seats are actually used. Check by asking what each person would do with the tool on a typical day. If two seats would be opened twice a week, the per-seat price effectively buys less than it appears.

Also ask what the saving needs to be. At $45 an hour, a $325 monthly bill is covered by about seven hours of saved time. That is a little over seven minutes a file across the firm's 60 files. A tool that saves ten minutes a file clears that comfortably; one that saves two does not. For wider pricing patterns in legal software, the tutorial on per-seat and pay-as-you-go legal AI pricing sets out the usual trade-offs.

A demo script you can paste into your invitation

Send this to the vendor a week ahead, so there are no surprises on the day and you can see who is prepared.

Subject: Demo request - conveyancing AI features

Thank you for arranging the demo on [date]. To make it useful for a firm of
[number] fee earners completing about [number] files a month, please plan to:

1. Work in our case-management system, [name], not a separate sample screen.
2. Load a closed purchase file we will provide (identifying details removed),
   including one scanned document of poor quality.
3. Show the AI feature turning the seller's replies into a client report,
   with the fee earner (not your team) at the keyboard.
4. Show the audit log for that action, and how to export it.
5. Confirm in writing before the call: processing and storage regions,
   sub-processors, whether our content is used for training, retention period.
6. Send a quote based on [number] files a month and [number] users, listing
   every extra charge, and the notice period to cancel.

We will test three things ourselves: a document inconsistency, a poor scan
and a question the documents cannot answer.

The tone is plain on purpose. A vendor who responds with a prepared checklist and a named technical contact is showing you how support will feel later. One who says that point five "needs to go to legal" and never replies is showing you something too.

Questions about the vendor, not the product

A few questions fall outside the demo but belong in the decision. How many conveyancing firms of your size use it, and can you speak to one? Ask for a reference, then ask the reference what went wrong in the first month. What does the contract say about changes to the product, price rises, and the model underneath being swapped? Has the vendor had a security incident, and how was it handled? What is the support route when a file is due to complete tomorrow: a person or a ticket queue?

Also ask what the vendor will not do. Honest vendors say that the tool does not replace the title check, does not verify identity, and does not decide whether a lender's requirements are met. Treat a product that claims to do all of that with caution. Identity checks and money-laundering checks stay with people and approved providers; the onboarding tutorial shows where AI helps around them without taking over.

Deciding after the demo

Within 48 hours, sit down with the scorecards and apply three rules. First, any "fail" on data location or audit trail ends the evaluation until it is fixed in writing, because these cannot be repaired after you have uploaded client files. Second, if integration needs manual steps, subtract the lost minutes from the savings and rerun the sum. Third, start with a small, reversible commitment: one fee earner, one task, a month-to-month contract if the vendor will agree, and a written list of what you will measure.

A sensible pilot measures time per task before and after on at least ten files, the number of corrections the fee earner makes to AI output, and any complaint or query from a client or lender that traces back to the tool. At the end of a month the numbers tell you whether to widen the rollout. If you want to pick the first task with someone else, the tutorial on cutting admin on each transaction lists the places conveyancers usually start.

Two mistakes to avoid when signing

The first is buying for the demo's best feature. A dazzling document-comparison tool is no use to a firm whose real bottleneck is chasing clients for documents. List your three slowest steps on a normal file before any demo, and weight the scorecard towards them.

The second is skipping the internal rule that goes with the tool. Write one page: which tasks may use the AI, which may not, who reviews output, and how a mistake is reported. Fee earners need to know that the summary is a draft, that source documents still get read on anything that affects a client's money, and that nobody pastes client details into a different, unapproved AI tool because the approved one is slow. A tool bought without that page tends to be used differently by every person in the office, which makes the audit trail you paid for harder to interpret.

Buying questions fee earners ask after the demo

Should we pay per file or per seat?

Compare both against your own completions. Count files opened in a normal month and fee earners who would use the tool. Per-file pricing suits low or lumpy volume, per-seat suits steady volume spread over few people. Ask for both quotes on your real numbers and for the cost in your busiest month.

Is a free trial enough to judge the software?

Only if you run it on a closed file, copied with client details removed, and time every step. A vendor's sample matter will always go smoothly. Your odd files, with missing documents and late amendments, show how it behaves under pressure.

Who is responsible if the AI misses something on a title?

The firm and the fee earner, not the software. Supplier terms usually exclude liability for output, so check what yours say and what your insurer expects. Treat every AI summary as a first read that a qualified person checks against the source document.

Do we need to tell clients we use AI?

Check your regulator's current guidance and your insurer's requirements, then say so plainly in your terms of engagement and privacy notice, covering what is processed and where. Clients rarely object to a clear explanation; they object to finding out later.

Further reads

Sources: vendor demo practice and general contract and data-protection checks; no vendor pricing is quoted because plans vary by firm.

Want a second pair of eyes before you sign?

On a 1:1 call we'll go through your shortlist against your own case-management set-up, map the demo questions that matter for your files, and pick what is worth piloting first.

Book a 1:1 call with me