AI for Trust Administration: Inventories, Notices and Accountings

A trust administration practice can use AI to build the asset inventory by extracting figures from statements, to draft beneficiary notices from a template, and to run consistency checks on an accounting, provided a person confirms every number and notice against the trust document and the governing rules. AI speeds the assembly; the trustee remains responsible for what is signed.

Those three jobs carry different risks. Extraction is a transcription risk, notices are a content and timing risk, and accountings are a reconciliation risk. A successor trustee, or the small practice supporting one, needs a separate control for each, and those controls are what keep a hurried afternoon from becoming a surcharge claim.

Follow me on Instagram@sagnikteaches

Why trust administration suits AI help, and where it does not

When a trust becomes irrevocable or a successor trustee takes over, the work is front-loaded and clerical: find the assets, get the values, tell the beneficiaries, keep records, report. Much of it is reading dozens of statements and turning them into rows. That is what AI does well.

Connect on LinkedInSagnik Bhattacharya

It does not suit three things: deciding what the trust document means when it is ambiguous, deciding what the law requires in a given place, and exercising discretion (distributions, investment choices, whether to pay a disputed claim). Those stay with the trustee and their legal adviser. The trust document, and the rules where the trust is administered, set the duties and the deadlines, and they vary; the workflow below travels, the rules do not.

Subscribe on YouTube@sagnikteaches

The successor-trustee workflow in six passes

1. Collect and label. Gather the trust document, any amendments, the certificate or evidence of the change of trustee, and every statement, deed and bill. Name files consistently (asset type, institution masked, date).
2. Extract values into an inventory. Turn each statement into rows: asset, institution, last four digits of the account, ownership, value at the relevant date, source document.
3. Issue notices. Send the notices the trust document and the governing rules require, in the required form and time.
4. Track receipts and payments. Record every transaction from the trustee's date of responsibility onwards, with a category and a document reference.
5. Prepare the accounting. Summarise receipts, disbursements and the closing position for the period, and reconcile to bank and custodian records.
6. Review and sign. A named person reviews and approves before anything is sent to beneficiaries or filed.

AI plays a part in passes two, three, four and five. It plays no part in pass one beyond helping to name files, and none in pass six.

Pass 2: Statement extraction into an asset inventory

An inventory is a list of what the trust holds, with values at a defined date. The date and the valuation method depend on the circumstances and on the governing rules; ask the supervising adviser which applies before you begin, because it changes which numbers you pull from each statement.

The extraction step is repetitive: bank balance on a given day, brokerage holdings with units and prices, a property valuation from a letter, a life-policy value. A document-reading tool or a chat assistant on a business plan can pull these into a table. Before using one, mask what it does not need. Account numbers become the last four digits, tax and identity numbers are removed, and beneficiary names become labels.

From the statement text below, extract one row per holding.
Columns: institution (use the label I give), account (last 4 only),
asset description, quantity, unit price, total value, valuation date
shown on the statement, page number where found.
Rules: copy numbers exactly as printed. If a value is missing or
unclear, write NOT FOUND, never estimate. After the table, state
the grand total as printed on the statement and the sum of your rows.

INSTITUTION LABEL: Bank-1
STATEMENT TEXT: [paste masked text]

The last instruction, giving both the printed total and the sum of the rows, is a built-in check. If they differ, the extraction missed or misread something. A realistic sample result (illustrative): a brokerage statement lists nine holdings and prints a total of 412,880.15, and the tool's rows add to 409,330.15. The gap of 3,550 is a holding that wrapped across a page break. You find it, fix it, and the totals match. Without the check, the inventory would have carried the error forward.

Check three things on every extraction

• Date. Is the value at the date the inventory needs, not the statement's issue date? Statements often show a period end that differs from the date you need.
• Ownership. Does the title on the statement show the trust, the former trustee individually, or joint ownership? An asset in the wrong name may be outside the trust, and that is a legal question, not a tool question.
• Source. Does each row point to a page of a saved document, so a reviewer can verify it in one click?

A simple inventory layout (illustrative figures):

The last row is deliberately unfinished. A tracker that shows a gap is more trustworthy than an inventory that looks complete and is not. For help with the extraction mechanics in general, the tutorial on chat-with-PDF tools and client files sets out which kinds of tool suit sensitive documents.

Pass 3: Beneficiary notices drafted from your own template

Notices are usually the first formal communication beneficiaries receive from a new trustee. What the notice must say, who must receive it and by when are set by the trust document and the governing rules, which vary from place to place and by type of trust. Have an adviser confirm the required content and timing once, for each jurisdiction and trust type you work with, and save that as the practice's approved template. After that, AI's role is narrow: fill the variable fields and adjust the tone.

The practice's template might have these variable fields: trustee name, trust name, date of change or death, beneficiary label, summary of the beneficiary's entitlement as drafted by the adviser, contact details, and the date by which any follow-up is expected. The fixed legal wording stays untouched.

Fill the variable fields in the approved notice below using the data
sheet. Change NOTHING outside the square brackets. Do not rephrase
the legal paragraphs. If any field has no data, leave it as [MISSING]
and list it at the end.

APPROVED NOTICE: [paste template with [FIELDS]]
DATA SHEET: [paste labels and facts]

This is a mail-merge with a language model's help, and for a single notice a mail-merge is enough. AI becomes useful when the notices differ: a trust with seven beneficiaries, two of them minors with a parent as contact, one living abroad where the delivery method must differ, and one who has asked for plain English. It can draft the covering note that sits on top of the approved notice in the right register for each, leaving the legal paragraphs alone.

A covering note before and after (illustrative)

Draft: "As trustee I will make distributions at my discretion and will keep you informed of all developments."

After review: "I will write to you again with a summary of the trust's assets by [date]. Any decision about distributions will be taken in line with the trust document, and I will tell you when one is made."

The draft promised discretion and unlimited information, both of which are legal positions. The reviewed version says what will actually happen. Every promise in a beneficiary letter is something the trustee may be held to, so a reviewer reads those sentences with particular care.

Timing is the dangerous part

AI cannot know the notice period that applies, and it will not warn you if it does not. Build the timeline from the adviser's confirmation and put each notice date in the tracker with its source. The tutorial on answering client status questions shows how to keep routine follow-up questions consistent once the formal notices are out.

Pass 4: Receipts and payments ledger

From the date the trustee becomes responsible, every receipt and payment needs recording with a category and a document. Categories are the practice's own: interest, dividends, rent, sale proceeds, funeral or administration expenses, professional fees, taxes, insurance, repairs, distributions. Bank-line categorisation is the same job a bookkeeper does, and the tutorial on receipt capture and bank categorisation covers the setup. For a trust, add two columns the bookkeeper would not: whether the item is income or capital, and whether it was paid from trust funds or by the trustee personally and is awaiting reimbursement.

The income-or-capital split matters because it can change who is entitled to the money, and it is a classification the trust document and the governing rules decide. Use AI to propose a category and flag any line it cannot place. Never let it decide an income-or-capital question: a person with the trust document in front of them does.

Categorise each bank line using ONLY the category list below.
If a line does not clearly fit, label it QUERY with a one-line reason.
Do not decide income vs capital; leave that column blank.
CATEGORIES: [list]
LINES: [date | description | amount]

A realistic sample result: out of 140 lines, 121 are categorised, 19 are queries. Of the 19, six are card payments with cryptic merchant names, four are transfers between the trust's own accounts (which should not appear as receipts or payments at all), and nine are genuine questions for the trustee. The transfers are the common trap: counted twice, they inflate both receipts and payments and break the reconciliation.

Pass 5: Accounting checks that catch real errors

The accounting is the trustee's account of stewardship for the period: opening position, receipts, payments, distributions and closing position. Its form is usually fixed by the governing rules or by a court or beneficiary expectations. AI is more useful as an auditor of the draft than as its author.

Give the assistant the finished figures as a table, with no names, and ask it to look for specific problems.

Below is a trust accounting summary as figures only. Check:
1. Opening balance + receipts - payments - distributions = closing.
2. Any category total that moved by more than 25% against the
   previous period, and any category with a single item over 10%
   of its total.
3. Any payment with no matching document reference.
4. Duplicate amounts on the same or adjacent dates.
5. Any transfer that appears as both a receipt and a payment.
List findings only. Do not correct the figures.

The thresholds (25%, 10%) are examples; set your own to suit the size of the trust. The instruction "list findings only, do not correct" matters because a tool that silently adjusts figures to make a total tie is the worst outcome.

The checks a person must do, whatever the tool says

• Reconcile each bank and custodian balance to the statement at the period end, by hand or in your own spreadsheet.
• Confirm the closing assets agree with the inventory plus movements.
• Confirm every distribution has authority in the trust document and a record of the beneficiary's receipt.
• Check fees and expenses are ones the trust document or the rules allow the trustee to charge.
• Check the period dates and the date of the last entry.

Keeping the trustee's duty intact

A trustee can usually use agents and technology, but remains responsible for supervising them. Practically, that means three habits. First, keep a short record of how AI was used on each trust: which steps, which tool, and who reviewed the output. Second, keep every source document, so any figure can be traced without the tool. Third, keep conclusions out of the tool: it may sort, extract, draft and flag, and it does not decide.

The practice's own governance matters too. Use a business plan that does not train on your content by default (ChatGPT Business, Claude Team and Microsoft 365 Copilot, which Microsoft's documentation now calls Microsoft Copilot, say this on their own pages, as checked on 4 November 2026), keep the account under practice control, and record the tool in your information-security procedures. The tutorial on keeping customer data private when a team uses AI has a policy you can adapt, and the list of legal tasks never to hand to AI is a good basis for the "not here" column of your own procedures.

A full example: a modest trust, one afternoon and one week

Say a small practice supports a family member who has become successor trustee of a trust holding two bank accounts, a fund platform account, a share of a rental property and a life policy, with four adult beneficiaries. The trustee's adviser has confirmed which notices are needed and when. The practice's paralegal does the following over a week.

On Monday she masks and extracts six statements into the inventory, using the extraction prompt with the printed-total check. Two totals do not match: one because of a page break, one because the tool misread a decimal point on a unit price. She fixes both and notes them in the working paper. On Tuesday she fills the approved notice for each of the four beneficiaries from the data sheet, writes covering notes for two who prefer plain English, and hands the pack to the adviser, who changes one sentence about a beneficiary's entitlement. On Wednesday and Thursday she categorises the first quarter's bank lines (96 lines, of which 14 are queries) and clears the queries with the trustee by phone. On Friday she runs the accounting check prompt on the first-quarter summary; it flags a 5,000 transfer from the savings account to the current account that appears as both a receipt and a payment. She corrects it, the adviser reviews the pack, and the notices are posted with the date logged against the tracker.

The quick sum shows why the balance check alone missed that transfer (illustrative figures). As drafted: opening cash 80,000, plus receipts of 13,400, less payments of 8,150, gives a closing 85,250, which agrees with the bank. Corrected: 80,000 plus 8,400 less 3,150 gives the same 85,250. The closing figure was right both times, but receipts and payments were each overstated by 5,000, and a beneficiary reading the accounting would have seen a quarter with 8,150 of spending that never happened. Check 5 in the prompt, not check 1, is what caught it.

No timesheet is needed to see where the effort went: the extraction and the categorisation, the two jobs most likely to produce errors, took the tool minutes and the checking an afternoon, and that checking is exactly what cannot be dropped.

Failure modes to look for in this particular work

• Stale values. A statement date after the date of valuation needed. Show-up: the inventory total disagrees with the adviser's figure. Fix: write the valuation date on the sheet before extraction.
• Co-mingled assets. The former trustee's personal account appears in the pile. Show-up: an account with a person's name only. Fix: refer to the adviser before adding it to the inventory.
• Notice for the wrong beneficiary class. A template written for one type of interest reused for another. Show-up: a beneficiary queries the wording. Fix: the adviser confirms the template per class.
• A tool that rounds. A chat assistant presented 18,240.55 as 18,240. Show-up: pennies differ at reconciliation. Fix: the exact-copy instruction and the print-total check.
• Silent assumptions. The draft accounting assumes a fee was allowable. Show-up: a beneficiary objects. Fix: the trust-document check on every fee line.

Where the same pattern repeats across many trusts, the practice has a case for a more structured workflow with its own templates, not just prompts. That is the kind of job an AI implementation consultation is built for; you can reach me through the contact page if you want to talk it through.

Records to keep for each trust

If a beneficiary ever challenges an accounting, the practice's defence is the paper trail. For each trust, keep a single folder or workspace with the same layout every time so a colleague can pick it up cold:

• Authority: the trust document, amendments and evidence of the trustee's appointment.
• Inventory: the final table with every source document linked, plus the earlier drafts showing corrections.
• Notices: each notice as sent, proof of delivery and the adviser's confirmation of timing.
• Ledger: receipts and payments with document references, and the query log showing how each unclear line was resolved.
• AI use note: one paragraph per pass saying whether a tool was used, which account, what data was sent in masked form, and who reviewed the output.
• Sign-off: the reviewer's name, date and what they checked.

The AI use note takes two minutes and is the item most often skipped. It is also the one that answers a later question such as "who produced this number and how was it checked?".

Choosing what to try first

If the practice handles only a few trusts a year, start with extraction on one trust and keep the printed-total check. It is the easiest win to measure, because the totals either match or do not. If it handles dozens, start with the ledger categorisation, because the volume of bank lines is where hours go. Leave notices until the adviser has approved the templates, and leave accountings until you have two or three clean ledgers behind you. Run each new step on a closed trust first, comparing the AI-assisted result with the one already signed off, and note every difference. That costs a few hours and tells you more about the tool's real accuracy than any vendor claim.

A short procedure to pin up

1. Confirm the notice content, deadlines and valuation date with the adviser before starting.
2. Mask identifiers. Only approved business accounts for any AI use.
3. Extract with the printed-total check; fix mismatches before moving on.
4. Fill notices from approved templates only; the legal wording does not change.
5. Categorise bank lines; queries go to the trustee; income-or-capital is a human decision.
6. Run the accounting checks; reconcile by hand; log the reviewer's name and the date.

Questions about AI and trust administration

Can AI prepare a trust accounting?

It can help assemble and check one: reading statements into a ledger, flagging transactions that do not fit a category, and spotting totals that do not reconcile. The accounting itself is the trustee's statement, so a person must reconcile it to source records and own every figure before anyone relies on it.

Are beneficiary notices safe to draft with AI?

The drafting is low risk if you work from your own template and supply the facts. The risk is in the content and timing: required wording and deadlines depend on the governing law and the trust document, so a qualified adviser should confirm those before the first notice goes out.

What should never be pasted into an AI tool?

Full account numbers, identity numbers, original death certificates, passwords and anything that identifies a beneficiary by more than a first name or label. Work from extracted values and masked identifiers, and keep the unredacted documents in your document system.

Do I need to tell beneficiaries I use AI?

Not necessarily, but the trustee owes duties of care and, usually, of communication. If a beneficiary asks how the accounting was prepared, you should be able to describe the process and show that a person checked it. Saying so in the engagement terms stops the question arising later.

Further reads

• How to Summarise Bundles and Transcripts With AI Safely — Summarising large bundles of statements and letters safely.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Drafting routine letters for professional review.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Test whether the practice is ready to adopt AI.
• How to Anonymise Client Data Before You Paste It Into AI — A method for masking identifiers before any prompt.
• Is AI Worth It for a One-Person Bookkeeping Business? — Whether AI pays off for a very small practice.
• How Accountants Use AI to Spot Errors in Client Books — Using AI to flag errors in ledgers.
• Trust Account Reconciliation for Law Firms: What AI May and May Not Do — The three-way trust reconciliation step by step, with exactly where AI may flag problems and which checks must stay with a named person.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: vendor privacy and plan pages for ChatGPT Business, Claude Team and Microsoft 365 Copilot (checked 4 Nov 2026); trust-administration practice guides; the governing trust document and applicable local rules must always be checked.