AI adds drafting, matching and first-pass checking to fund administration software: investor statement commentary, reconciliation break triage, exception summaries and review notes. It does not add judgement. Valuation sign-off, the NAV figure, cash movements and the four-eyes check stay with named people, because those are where an administrator's accountability sits.
The useful split is between work that produces a draft a reviewer will read anyway and work that produces a number someone relies on. AI is good at the first and unreliable as the only source for the second. Almost every sensible use below is a way of making the preparer faster while leaving the reviewer's job untouched.
What a small administrator's quarter actually contains
A small administrator, say a team of six looking after a dozen private funds and a few hundred investors, repeats the same cycle each period. Prices and accruals come in, the NAV is struck, capital account statements go out, cash and positions are reconciled to the bank and custodian, and every step has a preparer and a reviewer. The software you already use, whether a dedicated platform or a set of linked spreadsheets, handles the calculation. What eats the hours is everything around it.
That surrounding work is where AI earns its place. It is text-heavy, repetitive and exception-driven: writing commentary that explains why a statement moved, working out why two lists of transactions disagree, turning a reviewer's scribbled queries into something the preparer can action, and chasing missing documents. None of it needs the model to decide anything. It needs the model to draft, sort and summarise so a person can decide faster.
Some administration platforms now ship an assistant or a "copilot" panel. Whether you use one of those, a business plan of ChatGPT, Claude or Gemini, or a Copilot inside Microsoft 365, the questions are the same: what data goes in, who reads the output, and where is that recorded. The next sections go task by task.
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 nowFour places AI helps and one where it should not
Investor statement commentary
Most statements carry a short narrative: opening balance, contributions, distributions, fees, the period's gain or loss, closing balance. Preparers often write that paragraph by hand for every investor class. Give the model the already-calculated figures from the statement as a small table and ask only for wording. It should never recompute.
You are drafting the commentary box on a capital account statement.
Use ONLY the figures in the table below. Do not calculate, round or infer anything.
If a figure you would need is missing, write [MISSING: name of figure].
Tone: plain, neutral, no performance language beyond what the figures show.
Length: 60-90 words.
Opening balance: 250,000.00
Contributions: 50,000.00
Distributions: 0.00
Management fee: -1,875.00
Net gain allocated: 14,320.50
Closing balance: 312,445.50An illustrative output reads: "Your capital account opened the period at 250,000.00. You contributed 50,000.00 and received no distributions. A management fee of 1,875.00 was charged and a net gain of 14,320.50 was allocated, giving a closing balance of 312,445.50." The reviewer's check is arithmetic: 250,000.00 + 50,000.00 - 1,875.00 + 14,320.50 equals 312,445.50. Because the prompt forbids new figures, the reviewer is checking transcription, not mathematics, and that takes seconds.
Reconciliation break triage
The hard part of a reconciliation is rarely the matching of the 95 per cent that agree. It is working out why the last few lines do not. Export the unmatched items from both sides to a sheet, then ask for a categorised list of probable causes. Typical categories: timing (booked on different days), amount differing by a fee or rounding, duplicate, one side missing, wrong investor or reference.
Here is a worked example with illustrative numbers. A month-end bank reconciliation for one fund has 212 lines on the bank side and 208 in the ledger. After automatic matching, nine bank lines and five ledger lines remain. The preparer pastes those 14 lines with dates, references and amounts and asks the model to propose pairings and a reason for each unpaired line. Four of the rows that come back, with the preparer's decisions:
| Bank line | Ledger line | AI's proposed reason | Preparer's verdict |
|---|---|---|---|
| 12,500.00, ref SUB-0417 | 12,500.00, 2 days later | Timing, same amount and reference | Accept, clears next period |
| 9,980.00, ref DIST-22 | 10,000.00, ref DIST-22 | 20.00 difference, likely bank charge | Accept after checking charge on statement |
| 4,200.00, no reference | None | Unreferenced receipt, possible subscription | Query with the manager, do not book |
| None | 3,150.00, ref FEE-Q3 | Fee accrued but not yet paid | Accept, confirm against fee schedule |
Notice what the preparer did with the third row. The AI offered a plausible story ("possible subscription") and the preparer treated it as a lead, not a fact. A reconciliation note that says "AI suggested X" is not evidence. The evidence is the bank statement, the subscription document or the manager's reply. For more on keeping the match decisions checkable, how to automate bank reconciliation with AI and check the matches covers the controls in detail.
Review notes that the preparer can action
A reviewer's notes are often terse: "check fx here", "why is this different to last Q", "ref?". Converting those into a numbered query list with the cell or document referenced is a pure text job. Paste the notes and the working paper's headings, ask for a tidy list with one action per line and a place to record the response. The model does not need the figures at all, which also makes it the lowest-risk use in this tutorial.
Turn these review notes into a numbered query list for the preparer.
One action per line. Keep my wording where it is clear.
Add a column "Evidence required" with one of: document, calculation, explanation.
Do not answer the queries yourself.
Notes:
check fx on row 14 / why bigger vs last Q / ref? on 4,200 / who approved fee waiverAn illustrative result: "1. Row 14: confirm the exchange rate used and its source date. Evidence required: document. 2. Explain why the balance is larger than last quarter. Evidence required: explanation. 3. Provide the reference for the 4,200.00 receipt. Evidence required: document. 4. Confirm who approved the fee waiver and attach the approval. Evidence required: document." What to fix: the model turned "why bigger" into "explain why the balance is larger", but the reviewer meant the management fee line, not the balance. A preparer who answers the wrong question wastes a review cycle, so the reviewer reads the tidied list once before it goes back, and any note that was ambiguous in shorthand gets a cell reference added.
Exception summaries for the manager
At quarter end the administrator's manager wants one page: what is late, what is unreconciled, what is waiting on a client. Feed it the status columns from your tracker, not the underlying documents, and ask for a summary grouped by fund with the oldest items first. The value is consistency. The same layout every quarter makes slippage visible, and the manager reads a page instead of opening twelve files.
A filled-in excerpt, with invented funds and items: "Fund C: two unreconciled custody lines, 19 and 6 days old, waiting on the custodian's corrected statement. Fund F: one investor's capital call notice not acknowledged, 12 days. Fund H: manager has not approved the quarter's expense accrual, 4 days." The check is quick because every line maps to a row in the tracker: count the open items in the tracker per fund and compare with the summary. If the summary lists five items and the tracker has six, the model dropped one, and the dropped one is usually the oddly worded entry, which is often the item that matters.
Where it should not be the answer: the number itself
Do not ask a chat model to strike a NAV, allocate a gain across classes, or work out a performance fee crystallisation. Language models can produce confident, tidy and wrong arithmetic, and a mistake presented in a well-written paragraph is harder to catch than one in a blank cell. The calculation belongs in the platform or a locked, versioned spreadsheet. AI sits beside it. A good habit is to let AI generate a check formula or a test case, then run that inside the spreadsheet where the result is visible. The tutorial on ChatGPT with Excel for accountants shows how to keep that boundary clear.
The four-eyes check when AI is in the chain
Four-eyes means a second, independent person reviews work before it leaves the building. Adding AI does not change who the two people are. It changes what the preparer hands over, and that has to be visible to the reviewer. Three adjustments keep the control meaningful.
- Mark what AI touched. Each working paper carries a short line: "Commentary drafted with AI from table on tab 2; figures checked by preparer." The reviewer then knows where to look hardest.
- Keep the source beside the draft. The reviewer ticks the statement figures against the platform's output, not against the draft's own numbers. Ticking a document against itself is not a check.
- Record the reviewer's name and date. An AI check does not appear in the sign-off. Only a human name does.
A realistic failure: a preparer asks the model to "tidy" a statement narrative, and it rounds 14,320.50 to 14,321 in the prose while the table still shows 14,320.50. Nobody notices for a quarter because the reviewer only checked the table. The fix is procedural: prompts forbid rounding, and the reviewer reads the narrative against the table line by line. The cost is thirty seconds per statement. The cost of a restated statement to an investor is far greater.
A quarter-end worked example with numbers
Say an administrator runs 12 funds with 340 investor statements per quarter. Today the preparers spend roughly 12 minutes per statement on commentary, assembling attachments and a first self-check, which is about 68 hours. The reviewer spends 5 minutes per statement, about 28 hours. These are illustrative figures for planning, not benchmarks.
With the drafting approach above, the commentary becomes a prompt filled from a table, taking perhaps 4 minutes including the preparer's own tick against the platform. Preparer time drops to roughly 23 hours. The reviewer's time does not fall by much, and should not: say 5 minutes becomes 4 because the layout is now identical and the AI-touched line points to the checks. That is roughly 45 hours saved on preparation and about 6 on review, so something like 50 hours a quarter, more than a working week for one person.
Now the cost side. A business plan of a major assistant is typically a modest per-user monthly fee. Confirm the current price on the vendor's page rather than relying on this tutorial. Add the one-off set-up: two or three hours writing prompts, an hour on a template, and the pilot on a closed quarter. A fair estimate is two working days of the most senior preparer, once. That is paid back in the first live quarter if the figures above are even roughly right. If the savings are half that, it still pays back within two quarters. If the error rate in the pilot is high, you stop, and you have spent two days, not a quarter's reputation.
Running a pilot on a closed quarter
Pick a fund whose quarter has already been issued. Gather the same inputs the preparer had, strip the investor names to codes if you want an extra margin, and have the AI-assisted route produce the commentary and the exception list. Then compare with the real issued pack and score each item.
| Measure | How to count it | Pass mark to go live |
|---|---|---|
| Figure transcription errors | Any number in narrative that differs from the platform | Zero across 30 statements |
| Wording that needed rewriting | Statements where the reviewer changed more than a phrase | Under 1 in 5 |
| Reconciliation reasons right | AI proposed cause matched the verified cause | 7 in 10 or better, and no false "cleared" |
| Time per statement | Stopwatch for preparer, end to end | At least a third lower |
The pass mark for figure errors is deliberately zero. A tool that sometimes changes a number is a tool that cannot draft figures without a full reread, and you want to know that before it is in the live process. If the model fails on transcription, switch the prompt to produce wording with placeholders such as {closing_balance} and merge the numbers from the table yourself. That removes the model from the numbers altogether, and it works well.
Client data, confidentiality and what to ask the vendor
Investor names, account details and holdings are confidential, and your service agreements probably say so explicitly. Before anything goes into an assistant, settle five points. Does the plan train on your content by default? Who in your firm can see conversation history? Where is data processed, and can you pick a region? Is there an audit log you can export? Can you delete content, and how fast? The tutorial on what to check in an AI tool's privacy policy and terms turns that into a reading list, and classifying business data before using AI tools helps you decide which fields may be pasted at all.
A pragmatic rule many small administrators adopt: investor identifiers are replaced with codes before text leaves the platform, and the mapping stays in the platform. A statement draft that says "Investor 0417" and gets the real name merged back in at the end is easier to defend than one containing a name and an account number. Treat bank account numbers as never pasted, whatever the plan.
Also look at your clients' side. Fund managers send due-diligence questionnaires, and "how do you use AI" is now a common line. Have a one-paragraph answer ready: the tasks, the data, the review step, the tools. If your agreements have a clause on subcontractors or technology, ask your lawyer whether an AI vendor counts. The page on mentioning AI use in client contracts gives practical wording to start from.
Other jobs worth a look, with their checks
- Chasing missing documents. Draft polite, specific reminders from a list of outstanding items per investor (for example, a tax form or a refreshed identity document). Check: a person sends them, and the tracker, not the draft, is the record of what is outstanding.
- Fee schedule cross-checks. Paste a fee clause and a computed fee line and ask the model to describe the basis in words, then recompute in the spreadsheet. Check: the sheet's number, never the model's.
- Meeting and call notes. Summarise a manager call into actions and owners. Check: the participant who made the commitment confirms the wording of any action touching cash.
- Policy and procedure upkeep. Turn a changed process into an updated checklist. Check: the compliance lead approves before it replaces the live version.
- Query answers to investors. Draft a reply to a "why did my balance change" question from the statement table. Check: reviewer reads it against the platform before it is sent.
The common thread is that every output has a human checkpoint with something concrete to check against. If you cannot name the document the reviewer ticks to, you have not found a good AI task yet.
When the platform's own assistant is enough
If your administration software already includes an assistant, try it first on the same closed-quarter pilot. It sits inside the permissions and audit trail you already have, which is worth a lot. A general chat assistant may do better on wording, but it is a second place where data lives. Decide on evidence from the pilot, not on brochures. Ask the platform vendor how its assistant handles your data, whether it is trained on it, and whether outputs are logged against the user.
For the spreadsheet-based parts of the process, a Copilot or Gemini inside your spreadsheet tool is convenient, with the same rule: figures come from cells, and AI writes about them. For a deeper look at reconciliation-style checks in a spreadsheet, reconciling a rent roll to the bank with AI works through a different industry with the same logic of matched, timing and unexplained items.
Keeping a record a regulator or client could read
Whatever you adopt, write it down on one page: the tasks AI is used for, the data allowed, the tool and plan, the review step, the owner, and the date last reviewed. When an auditor, a client or a regulator asks how you control AI use, that page is the answer. Keep a log of pilot results too. "We tested on two closed quarters, found zero figure errors across 60 statements and two wording rewrites" is a far stronger statement than "we believe it is accurate".
Finally, set a trigger for review: a change of model or vendor, a complaint about a statement, or any restatement. After any of those, rerun a small sample from the pilot. Models and plans change on their own schedule, and your control should not depend on nothing having changed. If you would rather map your own quarter-end steps with someone who has seen how these workflows go, that is what my AI implementation consultation covers, at the contact page.
Questions small administrators ask about AI in fund work
Can AI calculate a net asset value?
It should not be the calculation engine. Your administration platform or a locked spreadsheet calculates the NAV from priced holdings and accruals. AI can explain movements, draft the commentary and flag lines that look odd, but a human reviewer signs off the figure and the source behind it.
Do I need to tell fund managers I use AI?
Check your service agreements and your clients' due-diligence questionnaires. Many ask how you use automation. Being able to describe the tasks, the data handled and the human review that follows is better than being asked and having no answer ready.
Which AI plan suits client investor data?
Use a business-grade plan with no training on your content by default, admin controls and an audit trail, and confirm those terms on the vendor's current page. Avoid free consumer accounts for anything containing investor names, holdings or bank details.
Should AI replace the second reviewer?
No. A second pair of eyes works because it is independent. An AI check can be a helpful extra layer before the human review, but it cannot be the preparer's reviewer, because nobody is accountable for it.
How do I trial this without risking live files?
Run it on a closed quarter. Let AI draft, then compare against what was actually issued and count the differences. Two quarters of that gives you a real error profile before anything touches a live deadline.
Further reads
- AI Readiness Checklist for Accountants, Solicitors, Consultants — A broader readiness check before you change any workflow.
- How to Speed Up Month-End Close in a Small Business With AI — Close-cycle steps where AI saves time without touching figures.
- Claude for Excel for Accountants: Reconciliations a Manager Can Sign Off — Reconciliations a manager can sign off, in Excel.
- 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven AI slips that cost professional firms client trust.
- How to Stop AI Tools Training on Your Business Data — Switch off training on your content before uploading anything.
- AI for Insurance Brokerage: Reconcile Insurer Commission Statements — Another statement-to-ledger reconciliation worked through.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: vendor business-plan privacy and admin pages checked in 2026; the tutorial's figures are illustrative and not drawn from any fund.