A 90% forecast is not the point

A recent Business Standard report on EY’s view of agentic AI in cash forecasting put a compelling number into the conversation. The report suggests agentic AI could help corporate cash forecast accuracy reach 90%.

That number should make an accounting firm owner curious. It shouldn’t make you promise a client 90% accuracy.

Cash forecasts are only as good as the underlying data, the collection assumptions, open purchase commitments, payroll timing, seasonality, and the judgment applied to unusual events. A forecast for a clean, stable business with reliable AR history is a different job from a forecast for a construction client managing variations, retentions, and uneven subcontractor payments.

The useful part of this trend is not a headline percentage. It’s the chance to rethink a familiar service model.

Too many bookkeeping and accounting firms still create cash forecasts as an occasional spreadsheet exercise. Someone exports bank balances, aged receivables, aged payables, and payroll data. They copy figures into a workbook. They ask the client about significant payments. Then they build a 13-week view that is already stale by the time it reaches the client.

That work has value. The manual method does not have to remain the default.

Agent-assisted cash-flow forecasting can gather current information, apply agreed rules, identify gaps, draft scenarios, and produce a client-ready briefing. Your accountant still owns the review, the assumptions, and the recommendation. The agent takes on the repetitive assembly and monitoring that keeps senior people trapped in spreadsheets.

For a firm in the $1 million to $25 million range, that distinction matters. The annual leakage band we often see across accounting and bookkeeping operations is $60,000 to $180,000. It usually isn’t one dramatic failure. It’s small chunks of partner time, rework after bad source data, slow close cycles, and advisory opportunities that never get scheduled.

The spreadsheet forecast has a hidden operating cost

A manual cash forecast often begins with a good intent. A client says cash feels tight, the lender wants visibility, or the owner is thinking about hiring. Your team builds a forecast to answer the immediate question.

Then reality changes.

An invoice expected on Friday is paid three weeks late. A customer disputes a milestone. Quarterly tax falls due. The client makes an equipment deposit. Payroll timing shifts because of a public holiday. The forecast needs to be rebuilt, yet nobody has budgeted time for ongoing maintenance.

The result is a document that looks useful but isn’t used to make decisions.

Here is the usual manual chain of work:

1. Export bank transactions and balances from the accounting platform.
2. Extract AR and AP aging reports.
3. Check payroll, tax, loan, card, and recurring supplier dates.
4. Clean duplicate, missing, or misclassified transactions.
5. Move data into a cash forecast template.
6. Guess collection dates based on aging buckets or a conversation with the client.
7. Ask about large purchases, owner drawings, sales pipeline, and committed work.
8. Build a base view, then revise it after someone finds an exception.
9. Produce a report or email summary.
10. Repeat from scratch next month, or after the client asks an urgent question.

None of these steps is intellectually difficult on its own. Combined, they turn senior accounting capacity into data administration.

The timing is particularly painful around month-end and year-end. In many firms, 30% to 50% of staff time can become concentrated in roughly four weeks of the year. Forecasting gets delayed because close work has to be completed. Yet a late forecast is often least useful when the client needs it most.

The bigger issue is commercial. Advisory work commonly earns two to three times the billable rate of standard compliance work. If partners and managers spend their best hours chasing source data and updating workbook tabs, there is no room left to have the conversation the client would actually pay for.

That is the operating problem agentic forecasting should solve.

What an agent-assisted forecast actually does

“Agentic” is a loaded word, so strip it back. In this context, an AI agent is a controlled workflow that can gather data, apply instructions, flag uncertainty, and route decisions to a person. It is not an unsupervised bot moving client money or sending financial advice without review.

A practical cash forecasting agent follows a defined sequence.

1. It collects from approved sources

The agent pulls read-only data from the systems you approve. That might include the general ledger, bank feeds, accounts receivable, accounts payable, payroll calendar, debt schedules, and a client-maintained sales or project pipeline.

It should record when each source was refreshed. If a bank feed is two days behind or the AR ledger has not been reconciled, the forecast should say so. False precision is worse than an explicit warning.

A sensible pilot starts with a limited data set. Bank balances, open invoices, open bills, recurring payroll, taxes, and debt payments will answer most 13-week cash questions. You can add CRM and project data later when the core process is dependable.

2. It applies collection and payment logic

The agent does not simply use invoice due dates. It compares historical payment behavior with the current AR ledger.

For example, it may identify that a customer with 30-day terms typically pays in 42 to 48 days. It can group invoices by customer, age, dispute status, and size. It can then propose a collection date range rather than treating every invoice as certain cash.

On the payment side, the agent groups bills by due date, supplier pattern, category, and commitment status. It can identify recurring costs, tax obligations, payroll runs, loan repayments, and unusually large outflows.

The accountant defines the rules. A client with a seasonal retail cycle needs different assumptions from a professional services client billing monthly retainers. The agent works within those approved assumptions and highlights where the data does not support confidence.

3. It produces scenarios, not one fragile answer

A good cash forecast gives an owner choices.

The agent can draft a base case using standard collection behavior and committed payments. It can also build a conservative case where specific overdue invoices slip by 14 or 30 days, and an upside case where expected project revenue lands on time.

The point is not to fill a dashboard with charts. It is to answer practical questions:

• Which week does cash reach its lowest point?
• What customer payments are critical to staying above the agreed cash buffer?
• Can the business fund the planned hire without drawing further on a facility?
• What happens if the top two overdue invoices are paid 30 days late?
• Which supplier payments can be renegotiated without creating a new problem?

Those are advisory questions. They need an accountant’s judgment. They don’t require an accountant to keep copying rows between files.

4. It creates an exception queue for review

The agent should not quietly decide that an invoice will be collected or a payment can be delayed. It should put uncertain items into a review queue.

That queue might include:

• Invoices over a set threshold with no payment promise
• Customers whose payment pattern has deteriorated
• Bills coded to unusual accounts
• New recurring payments not seen in the prior three months
• Large cash movements with unclear descriptions
• A forecast cash low point below the client’s agreed buffer
• Differences between forecast and actual cash from the last cycle

Your review team confirms, adjusts, or rejects the suggestions. The final forecast is then published with a clear audit trail showing assumptions, data refresh timing, and reviewer changes.

This is where firms protect both quality and trust. The agent prepares the work. A qualified person remains accountable for the advice.

Data controls aren’t a side issue

Cash information is sensitive. Clients will rightly ask what data is being accessed, where it is stored, who can see it, and what the system is allowed to do.

Build the pilot around these controls from the start:

• Use read-only connections for source systems wherever possible.
• Restrict the agent to named clients and approved data fields.
• Separate one client’s data from another client’s workspace.
• Require human approval before any forecast is released externally.
• Keep a record of data sources, assumptions, prompts, outputs, and reviewer changes.
• Set retention rules for files and working data.
• Disable any ability to initiate payments, change bank details, or submit filings.
• Define an escalation path for missing data, unusual transactions, and low-confidence predictions.

A client doesn’t need a technical lecture. They need a clear explanation of what is happening and what isn’t. “We use your accounting and bank data to prepare a weekly forecast draft. Our team reviews every version. The system cannot make payments or alter your books.” That is plain language, and it sets the right expectation.

If your firm has not already documented practical operating controls, our material on Omni operations is a useful place to frame the workflow before choosing tools.

Start with a small client cohort

Don’t announce a new AI cash forecasting service across your entire client base. Choose three to five clients for a structured pilot.

The best candidates tend to have:

• A cloud accounting system with reasonably current reconciliations
• Regular invoicing and a visible AR ledger
• Repeatable payroll and supplier payment patterns
• A genuine need for cash visibility
• An owner or finance contact willing to provide business context
• Enough transaction volume that manual updating is already frustrating

Avoid starting with the messiest file in the practice. You want to test the workflow, not spend six weeks fixing historical data before the agent can do anything useful.

Set a 60- to 90-day pilot window. For each client, establish a starting position:

1. How many staff hours currently go into each forecast cycle?
2. How long does it take to obtain required data?
3. How often is the forecast updated?
4. How far does actual cash vary from forecast, and why?
5. How many advisory meetings result from the work?
6. What decisions did the client make differently because they had better visibility?

Review forecast variance every week. If the error comes from unrecorded bills or stale bank feeds, don’t blame the agent. Fix the operating process. If it comes from unrealistic collection assumptions, update the client-specific rule. That feedback loop is what makes the forecast improve.

For a clearer view of the broader opportunity, see Omni for accounting and bookkeeping. The aim is not to automate every task. It is to find the workflows where controlled AI creates more capacity for your people.

Connect forecasting to the rest of the practice

Cash forecasting becomes much more valuable when it is not a standalone spreadsheet replacement.

The Month-End Close Agent can pull bank, AP, AR, and payroll feeds, reconcile transactions, flag variances, draft journal entries, and prepare a partner-ready close pack. A cleaner and faster close gives the cash forecasting workflow better source data.

The Advisory Insights Agent can then read each client’s monthly numbers, surface three things to discuss, and draft partner talking points before the meeting. A cash low point is no longer just a cell turning red. It becomes a prompt for a structured conversation about collections, margins, funding, or spending plans.

The Client Onboarding Agent also matters. It can collect documents through a guided workflow, help set up the chart of accounts, and produce a clean opening trial balance. That gives your team a more consistent starting point for forecasting new clients, instead of inheriting months of uncategorised transactions and unreliable opening balances.

You can download the Month-End AI Close Map for Accounting Firms as a practical worksheet for mapping the handoffs between close, forecast, review, and client advisory. If you want the file directly, use this download link. It will help you identify where data becomes available, where it stalls, and who should approve what.

This is also why I would not start with a generic chatbot. You need workflows that understand operational sequence, ownership, and exceptions. You can see how those pieces fit across Omni, including the work that turns recurring firm processes into controlled agent workflows.

The financial case is usually capacity first

A firm does not need hundreds of forecasting clients to justify a pilot.

Imagine a team has 20 clients receiving some form of cash forecast. If each forecast takes three to five hours a month to gather, update, check, and explain, that is 60 to 100 hours of recurring work. Some of it needs accounting judgment. A meaningful share is data retrieval, formatting, and chasing context.

If controlled automation and better process design remove even a portion of that effort, the recovered time can go into higher-quality review and actual advisory meetings. It can also reduce the scramble during close, when experienced staff are most expensive and most likely to burn out.

The dollar impact will differ by client mix, labour model, and service pricing. That is why leakage in the $60,000 to $180,000 range is a useful investigation point, not a promise. You may find the value in lower rework. You may find it in more advisory engagements. You may find it in retaining a manager who would otherwise leave after another brutal year-end cycle.

The first job is to measure the workflow honestly.

Where does the data come from? Who touches it? How many times does it get checked? What creates avoidable delay? What decisions can only a partner make, and what is simply preparation work disguised as professional judgment?

If you want an outside view of those questions, Book a 60-min Omni Audit. We use the time to identify the workflow, the controls required, and the potential economics. There is no deck to sit through.

What to do next

Treat agent-assisted cash forecasting as a service design test, not a technology purchase.

Choose a small cohort. Document the current manual steps. Agree the source systems and rules. Set human review points. Compare forecast to actual cash each week. Record the time taken and the client decisions that follow.

If the pilot gives your accountants better information earlier, it has done its job. If it also frees enough capacity to create regular advisory conversations, it can change the margin profile of the service.

You can find the scope and approach behind the AI audit for accounting and bookkeeping. It is built for firm owners who want to move beyond broad AI discussions and identify a workflow worth implementing.

When you are ready to map your own pilot, Book a 60-min Omni Audit. You will leave with three outputs, the highest-value workflow to target, the controls it needs, and a practical next-step plan.