Pilot Spreadsheet AI Before Client Data
A spreadsheet rollout is a reason to run a careful pilot
Recent AI developments have put a familiar question in front of accounting and bookkeeping firm owners: can a new AI capability take real work off the team’s plate, or does it add another tool that needs supervision?
The answer won’t come from a polished demo. It comes from testing a recurring task against a sample that contains no client-identifying or confidential financial data, then checking the result against the firm’s own standards. Until you’ve reviewed the product’s data-handling terms, access settings, retention controls, and review process, keep client financial data out.
That’s especially important in accounting. A spreadsheet can look anonymous while still revealing a client through a company name, bank reference, tax identifier, unusual transaction, or combination of dates and amounts. Removing a business name isn’t the same as de-identifying the file.
The useful question isn’t simply, “Can it work with spreadsheets?” It’s, “Can it handle one well-defined step in our process, show its working, and leave a qualified person in control of the final decision?”
That’s a practical way to explore spreadsheet AI without making a risky commitment. Start with a repeatable task, set a clear pass-or-fail test, and keep the data boundary firm.
Pick one recurring task with a clear check
Month-end work is a good place to look because the same steps repeat across clients. Staff may compare bank activity with the general ledger, investigate unmatched transactions, check that totals tie, review unusual movements, or prepare a variance summary for a manager or partner.
Some of that work may already happen in spreadsheets assembled from exports. A staff member downloads a file, adjusts column names, fixes dates and number formats, adds formulas, and checks that the result agrees with a source report. Then they explain exceptions in notes or a close checklist. It’s routine, but it still requires care. A wrong formula or misunderstood category can flow into a close pack and create more work later.
A sensible first pilot might ask a tool to:
- Standardise column names and date formats in a synthetic transaction file.
- Categorise a small set of sample transactions using rules you supply.
- Flag rows that don’t match a stated reconciliation rule.
- Summarise changes between two versions of a mock report.
- Draft an exceptions list with a reference to the relevant row.
Keep the source data synthetic or thoroughly de-identified. Create it for the pilot rather than copying a client workbook and trying to clean it up afterwards. Give the tool a narrow instruction, such as “flag entries that don’t match these rules,” rather than asking it to “complete the close.”
Then define how you’ll judge the result. Does every flagged item point to the correct row? Are totals unchanged unless the instruction permits a change? Does the output distinguish a fact from a guess? Can a reviewer trace the result back to the source? Record the errors and the time spent correcting them, not just the time saved on the first draft.
If your team needs a grounding in the product before choosing a pilot, our guide to working with Claude can help frame the questions. The broader overview of Claude for business work is useful for separating general chat from a controlled workflow.
Keep financial data out until controls are checked
Before anyone uploads even a small file, have the person responsible for technology, security, or compliance review the relevant terms and settings. The details matter, and they can change by plan and configuration. Don’t assume that a business subscription automatically answers every question your firm has about data use.
Check what happens to prompts and uploaded files, who can access them, how long they’re retained, whether they may be used for model improvement, and what deletion options are available. Confirm whether the tool connects to other storage or business systems, and whether an administrator can control those connections. Make sure the firm’s policy on client data, confidentiality, and professional obligations covers the proposed use.
Review controls matter just as much. Can the user inspect the source and output side by side? Can the system show which rows support a result? Is there a record of what was asked and what was returned? Can a reviewer reject an output without it silently changing the underlying workbook? If the answer is unclear, the pilot stays synthetic.
A good de-identification check is concrete. Remove names, emails, account numbers, tax identifiers, addresses, bank details, file names, and document metadata. Replace dates and amounts where they could reveal a client or transaction. Check free-text descriptions, worksheet tabs, comments, hidden columns, and formulas. Have someone other than the preparer inspect the sample before it leaves the firm’s environment.
This isn’t bureaucracy for its own sake. A spreadsheet pilot should help the firm learn whether a task is suitable for automation, not expose client information before the controls are understood.
What an agent-led close could look like
A spreadsheet feature can help with a bounded task. An AI agent is a broader workflow that receives information, follows defined steps, routes exceptions, and prepares work for a person to review. It shouldn’t be treated as an unsupervised accountant.
Take a month-end close. The process could begin when approved bank, accounts payable, accounts receivable, and payroll feeds arrive through the firm’s authorised systems. The workflow checks whether expected files are present, confirms basic structure, and identifies missing periods or unfamiliar columns. It then applies firm-approved reconciliation rules and surfaces unmatched items or unexpected variances for staff attention.
Next, it could prepare draft journal entries and a close pack containing the reconciliations, exceptions, and supporting references. A staff member checks the source transactions, confirms the accounting treatment, and approves or edits the drafts. The partner sees a review-ready summary, rather than a workbook with unexplained changes. Nothing posts to the ledger without the appropriate human approval.
That’s one possible shape for a Month-End Close Agent: a workflow designed to organise and prepare close work for review. Its job is to organise and prepare. The accountant remains responsible for review, judgement, and approval.
The same principle applies to onboarding. A Client Onboarding Agent could guide a new client through document collection, track what’s missing, set up a chart of accounts using approved templates, and prepare a clean opening trial balance for review. The agent can keep the handoffs visible, but it shouldn’t decide on an accounting treatment without a person checking it.
These workflows connect to the firm’s broader operating model. Our article on why accounting AI pilots stall at workflows explains why starting with a tool before mapping the work often leads to disappointing pilots. And AI accounts payable for accounting firms shows how the same review-first approach can apply to another recurring process.
Measure the business case without promising savings
The financial case starts with workload, not a vendor’s estimate. Staff time can become concentrated around close and reporting cycles. That kind of compression can burn out staff, squeeze margins, and push advisory conversations off the calendar.
A pilot won’t remove the need for professional review. It may show that staff spend less time reformatting files, chasing exceptions, or drafting repetitive summaries. It may also show that the data is too inconsistent, the exceptions too complex, or the review burden too high for the task to make sense. All three are useful findings.
Track the time spent preparing the input, prompting or configuring the workflow, checking the output, correcting errors, and documenting the final result. Compare that with the current process for the same task. Include the time of the reviewer, not just the person who prepares the file. If the pilot shifts work from a junior team member to a partner, that isn’t an efficiency gain.
Firms may want to estimate their own annual leakage from avoidable rework, deadline pressure, poor handoffs, or advisory work that never gets scheduled. That’s a planning range, not a promised return from a spreadsheet tool. Leakage may come from avoidable rework, deadline pressure, poor handoffs, or advisory work that never gets scheduled. Your own workflow and capacity will determine where the opportunity sits.
The opportunity cost can be significant when compliance crowds out higher-value conversations. Advisory work may be billed at a different rate from compliance work, but the firm can’t bill for advice it never has time to prepare or deliver. A carefully designed close workflow may help staff get to review sooner and give partners a cleaner starting point for client discussions.
If you want to see how these tasks fit into a wider operating model, Omni Ops covers the operational workflows we build around accounting teams. For a firm-specific view, see Omni for accounting and bookkeeping.
Use the pilot to decide what happens next
A pilot should finish with a decision, not a vague feeling that the technology looked impressive. Agree on a small set of pass criteria before the test starts. For example, the output must preserve the source rows, identify exceptions accurately against the test rules, and give reviewers enough evidence to verify each result. Also decide who can run the test, what data is allowed, and who signs off on any proposed move beyond synthetic files.
If the task passes, the next step still isn’t uploading live client books by default. Confirm the product’s current data-handling and review controls, document the approved configuration, and test a controlled process with the appropriate internal approvals. If it fails, keep the notes. They’ll help you identify whether the barrier is the tool, the data, the process, or an unclear control.
To prepare for the next close, download the Month-End AI Close Map for Accounting Firms. It’s a practical worksheet for mapping inputs, handoffs, review steps, and exceptions before you choose a pilot. You can also download the close map directly.
Make the next step specific
An Omni Audit is a working session to identify where a workflow is costing the firm time or money, what’s ready for a controlled pilot, and what should stay with a person.
If you’re considering spreadsheet AI, bring one recurring task and a description of the data it uses. We’ll help you separate the safe synthetic-data test from the questions that need to be answered before client information enters any tool. Book a 60-min Omni Audit.
A spreadsheet pilot is a useful first test. It can also expose the deeper work needed to make close, onboarding, or advisory processes consistent. If you want to map that opportunity across your firm, the AI audit for accounting and bookkeeping is a practical place to start.
Bring the workflow, not client data. We’ll map what can be tested safely, where a human needs to review, and what a successful pilot would mean for the firm’s capacity and margins. Book your 60-min Omni Audit.
Your guide is ready
Check your downloads folder. If it did not open automatically, use the button below.
Download the GuideYour guide is ready
Check your downloads folder. If it did not open automatically, use the button below.
Download the GuideTalk it through
Talk it through with Sam
30 minutes on what a Command Centre would look like for your business.
Book a call