Identifying AI use cases in the close cycle

Identifying use cases in the close cycle means walking systematically through the finance cycle — close, reporting, AR, VAT, cash — and at each step asking where AI can usefully take over work. For finance teams this reverse approach (process first, AI second) works better than "where can we use AI?" — it produces a longlist rooted in actual time sinks, not in a vendor pitch.

"Where can we use AI in finance?" rarely produces useful answers. It's too abstract: AI is a tool, not a goal. The more useful question is the reverse: what work does the finance team do today, where are the time sinks, and which of those steps lend themselves to AI support. Start there and you end up with a longlist rooted in the actual close cycle, not in a vendor pitch.

Process mapping is the tool to move from that abstract "AI in finance" to a concrete list. It works the same way as a classic process analysis, with one addition: at each step you explicitly ask whether an AI capability adds value here.

Why start with the finance process

Deploying an AI tool on a one-off task ("let ChatGPT write the board summary") is fine, but it stays a discrete productivity boost. It only becomes substantial when an AI step is embedded in a chain the finance team already runs: the month-end close, accounts receivable, VAT reconciliation, the quarterly board pack. Those chains are already organized, measurable, and labor-intensive. That's exactly where the volume sits to build ROI.

Thinking in processes also forces a realistic view. A use case that sounds fantastic in isolation — "AI does our close" — turns out, in process context, to depend on a sloppy chart of accounts, a non-integrated data flow, or an authorization matrix that doesn't lend itself to automation. The process context makes those dependencies visible before you start the build.

Step 1 — scope the process

Don't start with the entire finance function — start with one process. A process is a chain of steps that together produce one outcome: a published month-end report, a closed VAT return, a collected invoice, an executed payment. For an SME or scale-up finance team, one process is usually graspable in six to twelve steps.

Typical candidates in finance:

• Month-end close (from bank reconciliation to board pack)
• Accounts receivable (from invoice to collection)
• Accounts payable (from inbound invoice to payment)
• VAT reconciliation (from data pull to filing)
• Cash forecasting (from data to report)
• Variance analysis and management reporting
• Annual-report compilation

Prefer a process that recurs often (volume = ROI), is structured enough to describe in steps, is document- or text-heavy (where AI genuinely adds value), and has an owner willing to think along.

Step 2 — break the process into sub-steps

Per main step, write the sub-steps in active verbs. "Month-end close" is not a step — "pull bank transactions", "generate match proposals", "chase unexplained items", "book draft journals", "intercompany reconciliation", "write variance commentary" are steps.

Per sub-step, record:

• who does it (role, not name — AP, AR, controller, finance manager, CFO);
• how long it takes (average per month or week);
• what input is needed (Exact export, bank statement, open-items list, tacit knowledge);
• what output results (booking, email, report);
• what can go wrong or often requires rework.

A table in Excel or a document will do. The discipline is in granularity: a step that takes more than 30 minutes can usually be broken down further.

Step 3 — look at each step for an AI fit

Walk through each sub-step and ask four questions:

1. Is this text-in, text-out work? Reading, writing, classifying, summarizing. Core territory for language models.
2. Does the step repeat in a recognizable pattern? If 80% of iterations look the same (think: bank-transaction match), an AI can learn the pattern.
3. Is the required context available in digital form? If the knowledge only lives in the controller's head, no workable use case will come out. If it sits in Exact, in an existing manual, in old reports: it can.
4. What is the cost of an error? High (booking, payment, external communication) → human-in-the-loop. Low (internal helper) → can run more autonomously.

Every step where you answer yes on three of four is a candidate. Note it concretely: not "AI in the close" but "agent proposes, for every bank transaction, a match from open invoices; controller approves in a queue."

Step 4 — collect broadly, filter later

Don't filter too early in this phase. A longlist of 20-40 candidates across the whole finance cycle is normal. Writing down only "realistic" ideas excludes the gold nuggets that looked odd at first glance. You'll be strict on impact and feasibility in the next phase — see scoring use cases for finance.

A practical technique: have two or three people who run the process daily (AP clerk, controller, finance manager) walk the list independently and add their own ideas. Every process has blind spots, and the people doing the work see different opportunities than the manager who has it on paper.

Step 5 — cluster by capability and by value chain

Order the longlist on two axes.

AI capability matters because use cases sharing the same capability share the same infrastructure. Five different "extract X from a PDF" cases run on the same building block — build once, apply many. See the AI capabilities that matter for controllers.

Finance cycle shows whether the opportunities sit mostly in close (operations), in reporting (interpretation), in AR/AP (cash cycle), or in VAT/compliance. A healthy finance portfolio isn't skewed: if 90% of your ideas are in reporting, you've probably ignored the cash cycle.

Where this lands in practice

• A 80-FTE scale-up maps the month-end close in 9 steps and finds 14 candidates: bank-match proposals, draft journals, intercompany reconciliation suggestions, variance commentary, board-pack sections, KPI extracts, and more.
• A 50-FTE trading company walks the AR process and finds 8 opportunities: payment-behavior classification, personalized reminders, escalation routing, payment-plan call prep, and automatic CRM updates after collection.
• A 35-FTE professional-services firm maps VAT reconciliation and finds 6 candidates: balance-vs-return reconciliation, intracommunity checks, input-VAT correction, bad-debt VAT recovery, draft corrections.

What these three have in common: nobody started with "where do we put AI in finance." They started with the process, and AI became the answer to "what can make this step simpler."