AI can help flag mismatches in a law firm's trust account reconciliation, but it cannot do the reconciliation. A person must agree the bank balance to the statement, confirm outstanding items, review every difference, and sign off. AI never approves, posts, transfers or moves client money.
That division works because a reconciliation is mostly comparison, and comparison is the part a tool does well and a tired person does badly at 6pm on the last day of the month. The accountability is the part it cannot carry. So the monthly routine splits cleanly: at one marked step a tool may look at the records, and at every other step it keeps its hands off.
What the three records are
A three-way reconciliation compares three things that should agree at the end of the month: the bank account's balance (adjusted for items not yet cleared), the firm's trust journal or cash book, and the total of all the individual client ledgers. Regulators and law societies describe the check in these terms; the details of frequency, records to keep and who signs differ by jurisdiction, so read your own regulator's current rule before changing any routine.
- Bank statement balance, plus deposits made but not yet cleared, minus cheques or payments issued but not yet cleared, gives the adjusted bank balance.
- The trust journal is the firm's own running record of every receipt and payment through the account.
- The client ledgers are one running account per client or matter. Their sum, often called the trial balance, is the figure the firm owes clients.
The bank balance and the journal can differ because of timing; the journal and the client ledgers should differ never. If all three agree after adjustment, the account balances. If they do not, the difference has to be found and explained before anything else happens.
A worked example with a deliberate error
An illustration with invented figures, for a small firm's single trust account at month end.
| Step | Item | Amount |
|---|---|---|
| A | Closing balance on the bank statement | $148,250 |
| B | Add deposits recorded but not yet cleared | $12,000 |
| C | Subtract payments issued but not yet cleared | $9,400 |
| D | Adjusted bank balance (A + B - C) | $150,850 |
| E | Closing balance in the trust journal | $150,850 |
| F | Sum of 31 client ledger balances | $150,350 |
D and E agree, but F is $500 short. Bank and journal agree with each other, so the problem sits in the client ledgers. Possible causes: a receipt posted to the journal as $5,000 but to a client ledger as $4,500; a payment entered to a client ledger twice; a ledger entry made in the wrong month. A flag-only tool earns its keep here. Give it the month's journal lines and the month's client-ledger lines (with client names swapped for codes) and ask which amounts appear in one and not the other, or appear with different values. It returns the $5,000 versus $4,500 pair in seconds. A person then opens the original deposit slip, confirms $5,000 is right, and posts the correction to the ledger with a dated note.
Notice what the tool did not do: it did not decide which figure was right. It cannot know the deposit slip said $5,000; it can only say two records disagree. Deciding which record is wrong requires the source document, and that is a human step.
A second thing the same check should catch, even when totals agree: one of the 31 ledgers has a negative balance of $1,200. The totals can still balance because other clients' positive balances offset it, which means money belonging to one client has effectively been used for another. A negative ledger is an automatic stop-and-escalate, whatever the total says. Ask the tool to list every ledger below zero as part of every run; it is a simple filter and a serious flag.
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 nowWhere AI may help, and where it may not
| Step | AI may | AI may not | Person must |
|---|---|---|---|
| Get the bank figure | Nothing; read the closing balance off the statement | Be the source of the balance | Agree the figure to the original statement |
| List outstanding items | Suggest which lines are not on the statement | Declare an item cleared or cancelled | Tick each item against the statement and later clearance |
| Compare journal with ledgers | List mismatched amounts, dates, duplicates, negatives | Choose which side is correct | Trace each difference to a source document |
| Draft a correction | Describe the proposed entry in words | Post it or send an instruction to the bank | Approve, post and date the correction |
| Sign-off and record | Format the report for filing | Sign, certify or declare the account reconciled | Review, sign and keep the record for the required period |
| Any transfer or payment | Nothing | Initiate, approve or schedule it | Authorise per the firm's two-person rule |
The last row is the line to write into the firm's policy in plain words: no AI tool has a login to the trust bank account, the practice software's posting functions, or an approval queue. If a vendor offers "autonomous" posting or payment, that is a feature to leave off for client money, and a question for your regulator and insurer before it is ever switched on. For broader limits on what a firm should hand to a tool, see AI for Small Law Firms: What to Automate First.
The monthly routine, step by step
- Freeze the month. Post all receipts and payments for the month before starting; late entries mean redoing the comparison.
- Take the bank closing balance from the statement itself. Do not copy it from an email summary or a bank-feed figure.
- List outstanding deposits and payments. Tick each against the following month's statement when it arrives, and carry unticked ones forward with a note.
- Calculate the adjusted bank balance. Do the sum yourself, in the firm's spreadsheet or accounting system.
- Run the flag-only comparison. Journal lines against client ledger lines, plus a list of negative ledgers, plus duplicates. This is the step where AI helps.
- Investigate every flag. Open the source document. Record what was found and who found it.
- Re-run until the three figures agree. Never plug a difference into a suspense line to make the report balance.
- Review and sign. A different person from the preparer, where the firm has the numbers for it; otherwise the lawyer responsible signs after reading the flags.
- File the record. Statement, outstanding-items list, three totals, flag log, signatures and date. Keep for as long as your regulator requires.
Only step 5 involves a tool, and it is optional. A firm with 15 ledgers and a clean practice system may not need it at all; a firm with 150 active ledgers and a backlog of small receipts may find it saves an hour of scanning.
A flag-only prompt, with a sample of what comes back
Prepare two lists from your accounting system's exports: journal lines (date, reference, amount, ledger code) and client ledger lines (date, reference, amount, ledger code). Replace client and matter names with the codes before anything leaves your system, and keep the code key somewhere the tool cannot reach.
You are comparing two exports from a law firm's trust accounting.
List A = trust journal lines. List B = client ledger lines.
Task: report ONLY differences. Do not correct, explain causes or
decide which list is right. Report:
1. Lines in A not in B, and lines in B not in A (match on reference + amount).
2. Lines with the same reference but different amounts or dates.
3. Duplicate references within either list.
4. Any ledger code whose running balance goes below zero, with the date.
Output four short tables. If a table is empty, write "none found".
Finish with: "These are flags only. A person must check each against its source document."
LIST A:
[paste]
LIST B:
[paste]An illustrative result: table 2 shows "REF 2231: A = 5,000.00, B = 4,500.00, same date"; table 4 shows "Ledger C-017 below zero from 14th, lowest -1,200.00". Everything else is empty. What you fix in it: confirm the tool actually read all rows (ask it to state how many lines it saw in each list and compare to your export counts), because a pasted list that was truncated quietly produces "none found" for rows it never saw. Row-count agreement is the cheapest control on this step.
If your practice software already includes its own reconciliation report and mismatch flags, use those first; a built-in report works on live data without copying anything out, which removes the paste risk entirely. The tutorial How to Automate Bank Reconciliation With AI and Check the Matches covers operating accounts, where the stakes are lower and more automation is reasonable.
Mismatch types, and who must investigate each
| Flag | Likely meaning | First check | Escalate when |
|---|---|---|---|
| Same reference, different amount | Keying error on one side | Deposit slip or payment instruction | Differences repeat for one staff member or one client |
| In journal, missing from ledgers | Receipt not allocated to a client | Whose money was it, per the covering letter | Money sits unallocated beyond the firm's stated period |
| In ledgers, missing from journal | Ledger entry with no bank movement | Whether a transfer was entered on one side only | Always, until explained |
| Duplicate reference | Double posting or reused cheque number | Bank statement for one or two movements | Two movements really left the account |
| Negative client ledger | Money used for one client from another's funds | Which payment took it below zero | Always; consult the responsible lawyer the same day |
| Stale outstanding item | Cheque never presented, or deposit never made | Contact the payee or confirm the deposit | Older than the firm's review period |
The final column is the reason to keep a named person on the checklist. Some flags are bookkeeping, and some are potential regulatory events. A tool cannot tell the two apart, and telling them apart is the judgement you are paid for. Some jurisdictions require certain shortfalls or irregularities to be reported; your regulator's rules say which, so do not let an AI summary decide it.
Data handling for the one step where a tool sees ledgers
- Minimise. Export only the month's lines, only the columns the comparison needs, with names replaced by codes.
- Choose the plan on purpose. Business tiers from the main vendors do not train on business content by default, while consumer plans may need a setting turned off. Confirm what your plan does with chats, how long it keeps them, and whether admins can delete them.
- Keep the key apart. The code-to-client key stays in the practice system or a locked file, never in the chat.
- Delete the chat once the flags are logged, and record in the file that this was done.
- Check the engagement terms. Clients' consent and your insurer's conditions may have something to say about third-party processing. See Can Solicitors Use ChatGPT Without Breaching Confidentiality? for the confidentiality side.
A cleaner alternative is to run the comparison in a spreadsheet with formulas, with no AI involved at all. A lookup on reference and amount finds the same mismatches. AI adds value when the references are messy: "Smith deposit", "SMITH DEP 11/5", "dep - Smith". Fuzzy matching of free-text descriptions is the genuine advantage, and it is also the case where the tool is most likely to make a wrong guess, so those flags come labelled "possible match" and wait for a person.
Three situations where the routine gets tested
The bank charge that is not client money
The statement shows a $15 monthly fee. Bank charges are generally the firm's cost and not a client's. Some professional rules let a lawyer keep a small amount of the firm's own money in the trust account solely to cover such charges; others handle them differently, so follow your regulator's guidance. The tool will flag it as "in bank, not in journal", which is true and unhelpful. The person decides how it is booked and records the reason, so the same flag has an answer next month.
The deposit that belongs to two matters
One $8,000 receipt covers a client's two matters. The journal shows one line, the ledgers two lines of $5,000 and $3,000. A tool matching on amount flags a mismatch even though the split is correct. The fix is a note in the reference field ("split A/B") and a prompt instruction that split receipts are listed once and checked by a person.
The sole practitioner doing it alone
A one-lawyer firm has no second person to review. The substitute is time and a different medium: prepare the reconciliation on the last working day, and review it printed or on a different screen the next morning before signing. Using the flag-only tool as a second pair of eyes helps here, as long as the reviewer still checks the source documents the flags point to.
Mistakes to design out
- Treating "no flags" as "reconciled". It means the tool found nothing in what it saw. Row counts and the bank-statement agreement prove the rest.
- Plugging differences. A "difference" line that quietly absorbs $500 for three months is exactly what an examiner looks for.
- Letting the preparer be the only reviewer. If one person prepares, flags and signs, the control is weaker than it looks.
- Connecting a tool to the bank. Read-only feeds into your practice system are one thing; a tool that can initiate payments is another. Keep the second out of client money.
The firm's other money routines carry their own risks. Bills and chasers are a separate and much safer place for AI: Law Firm Accounts Receivable: AI-Drafted Chasers for Overdue Bills keeps the tool entirely away from client funds. If you want a first 90 days that sequences this kind of work sensibly, see AI Implementation Plan for a Small Law Firm: The First 90 Days.
How to check the routine itself worked
After the first three months, test the process, not just the numbers. Pick one month at random and re-perform it from the filed record: can a colleague reach the same three totals using only the statement, the outstanding-items list and the exports? If they cannot, the record is too thin. Next, plant a test: ask someone to alter one amount in a copy of the exports ($5,000 to $4,500, say) and run the comparison. If the tool does not flag it, or the reviewer does not chase it, you have found out cheaply what a real error would have done.
Count the flags each month as well. A steady two or three flags, all traced to keying slips, is a healthy picture. A sudden jump, or flags that keep pointing at the same ledger or the same member of staff, is a training or process issue that the tool has revealed, not caused. Write the count in the file alongside the signatures; the trend is useful evidence of a working control, and an unexplained blank month is the sort of gap a reviewer notices.
Finally, keep a short list of what the tool got wrong. If it twice read a split receipt as a mismatch, or truncated a long paste, note the fix in the prompt and the date. Prompts for client-money checks should change only deliberately, with the change recorded, in the way any other procedure would.
Writing the rule into the firm's policy
Put three sentences in the firm's written procedures. First: "No AI tool has access to the trust bank account, to posting functions in the accounting system, or to any payment approval." Second: "AI may be used to list differences between exported records; every difference is investigated by a named person against the source document." Third: "The reconciliation is reviewed and signed by [named role], and the record is retained for [period required by our regulator]." Fill the brackets from your regulator's own rules, not from this tutorial, and ask your insurer whether any condition touches client-data processing. Then re-read the rule each time you change tools, because a vendor update can add features, and features can quietly widen what the tool is permitted to do.
If you want a second pair of eyes on where tools could sit around your routines without touching client money, that is the kind of mapping my AI implementation consultation covers; you can get in touch here.
Questions firms ask about AI and client accounts
Can AI do the three-way reconciliation for me?
It can compare the three records you give it and list differences, which is useful. It cannot be the reconciliation, because someone must confirm the bank figure against the statement, verify outstanding items, and sign the result. The accountable person is the lawyer or the named responsible partner, not the tool.
Is it safe to paste client ledger data into a chat assistant?
Only if your regulator's rules, engagement terms and the plan's data terms allow it. A safer pattern is to replace client names with ledger codes before pasting, keep the key outside the tool, and use a business plan that does not train on your content. If in doubt, ask your regulator or insurer.
How often should a trust account be reconciled?
Many regulators require it monthly. At least one widely used model rule treats quarterly as the minimum and monthly as the preferred practice, because an error is hard to find among three months of transactions. Rules and record-keeping periods differ by jurisdiction, so check your own regulator's current rule, not a software vendor's summary.
Should AI be allowed to move money between client ledgers?
No. A transfer or disbursement needs an authorised person and, in many places, a written instruction or second approval. An AI tool can draft a note describing a proposed correction, but the transfer itself is made by a person in the accounting system.
What if the AI says everything matches but the bank figure is wrong?
That is why the bank figure is taken from the statement, not from a pasted extract that the tool summarised. Always agree the closing balance to the original statement. A tool can only match what it is given, so the check on inputs belongs to a person.
Further reads
- AI for Trust Administration: Inventories, Notices and Accountings — A different kind of trust: administering an estate or trust file.
- How Small Law Firms Use AI to Draft Letters and Routine Documents — Routine letters, another safe place for AI in a firm.
- AI Time Capture for Solicitors: Recover Hours You Never Billed — Recover unbilled time that eventually reaches the bill.
- How Small Law Firms Use AI to Answer Client Status Questions — Answering client status questions without opening the ledgers.
- Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — What per-seat and pay-as-you-go legal AI tools cost.
- 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Mistakes that cost a professional firm its clients' confidence.
- ChatGPT for Accounting With Excel: Seven Tasks, Each With Its Check — Seven Excel accounting tasks for ChatGPT, each with a worked example and the check that proves the answer before you rely on it.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: a national bar association's model rules on client trust account records (Rule 1 and its comment on monthly and quarterly reconciliation) and its model rule on safekeeping property (checked November 2026); check your own regulator's current rules.