Roles covered: CFO, controller, FP&A lead, accounts payable and receivable managers. Functions: finance, accounting, internal control. Linked workflows: support replies, onboarding, incident triage, order intake, inventory forecasting, document extraction.
The decisions you own
Finance has two jobs in an AI initiative. It reviews business cases across the company, and it runs finance workflows that could use assistance itself. The first job decides which proposals receive money and on what evidence. The second puts finance in the same position as every other process owner, with the added constraint that its outputs feed the books.
Three decisions belong to finance rather than to engineering, a vendor, or the sponsoring department:
- Which claimed benefits count, and when. A sponsor can describe released hours. Only finance can say whether those hours will change a payment, a hiring plan, a contract, or a revenue forecast, and on what date.
- What evidence releases the next tranche of money. Finance sets the gate between an evidence-gathering stage and a funded forecast, and the conditions under which a pilot stops.
- What an AI-assisted step may do inside the financial record. Who may approve a payment, post a journal, or release an invoice does not change because a model proposed the entry.
The measure that looks like progress but is not: the sum of "annual savings" across approved pilots. Released capacity, avoided expense, and revenue hypotheses are different quantities. Adding them produces a number no one can reconcile after the fact.
Separate five kinds of benefit
Released capacity means employees can perform more work with the same hours. Avoided expense means a payment or planned expenditure no longer occurs. Incremental contribution means additional revenue after the costs required to serve it. Working-capital improvement changes the cash tied up in operations. Risk reduction changes exposure to possible losses.
Each needs different evidence, and they cannot be treated as interchangeable annual savings. Ask the owner to identify the realization mechanism and the person with authority to carry it out. The economics chapter sets out these ledgers in full, and the economics worksheet gives the equations; this guide does not repeat them.
An illustrative capacity case: 100 hours released at $50/hour represents $5,000 of labor capacity. If only 40 hours replace overtime that would otherwise be paid at that same assumed rate, the demonstrated avoided expense is $2,000. Do not add $5,000 and $2,000; the avoided expense is one realization of part of the same capacity.
Inventory benefits show why the ledgers must stay apart. In the forecasting chapter, a lower average inventory releases working capital once, while the carrying cost it avoids recurs each year, and neither counts unless the agreed service level held. A case that books the released inventory as annual savings has confused a balance-sheet change with an operating one.
Map the work you manage
The table covers both jobs. Starting authorities are proposed, not observed practice; "unchanged" means ordinary software, a template, or a policy decision should come first.
Capability verbs in the table follow the capabilities chapter; starting authority uses the arrangements in assistance, approval, and autonomy.
| Task | Consequence of an error | Candidate capability | Proposed starting authority | Canonical chapter or first fix |
|---|---|---|---|---|
| Review a department's AI business case | Money committed to benefits that never realize | None; this is judgment | Unchanged | Economics worksheet |
| Reconcile pilot results with the approved case | Failed pilots continue; good ones lack evidence to expand | Classify, retrieve | Assistance | Pilot design |
| Extract fields from supplier invoices | Wrong amount, supplier, or tax; duplicate payment | Extract, classify | Approval before action; AP approves before posting | Document extraction |
| Match invoices to orders and receipts | Payment for goods not received | Rules first; extract for unstructured lines | Unchanged; assistance on exceptions only | This chapter, weeks 4–8; order intake for analogous staging and duplicate prevention |
| Release a payment | Loss of cash; fraud exposure | Act | Unchanged; human authorization only | This chapter |
| Draft variance commentary for the close | Misstated explanation to management | Generate, retrieve | Approval before action; preparer approves before distribution | This chapter |
| Post a journal entry | Misstated financial record | Act | Unchanged; human preparer and approver | This chapter |
| Support demand or cash forecasts | Excess inventory or shortfall; misdirected funding | Predict | Assistance; the planner owns the number | Inventory forecasting |
| Classify AI costs for reporting | Misstatement; audit finding | None; accounting determination | Unchanged | This chapter, classification |
Three of these tasks carry a line that should not move in a first program: releasing payments, posting journals, and determining accounting treatment. The first two are actions on the financial record. The third is a professional determination that depends on framework, facts, and policy. Assistance can prepare evidence for each; it should not perform them.
Several finance tasks do not need a model. Three-way matching of structured invoices, purchase orders, and receipts is a rules problem. Duplicate-payment detection starts with identifiers and amounts. Recurring variance explanations can start as templates. Fix those first; the remaining exceptions are the candidates for assistance.
Who decides what
| Decision or task | Finance leader | Process owner (sponsoring department) | Engineering or IT | Security or risk | Controller or accounting | Internal audit |
|---|---|---|---|---|---|---|
| Whether a benefit is realizable | Decides | Names the mechanism and owner | Informs costs | Informs risk costs | Informs | Reviews after the fact |
| Stage gates and stop criteria | Approves | Proposes | Informs feasibility | Sets non-negotiable limits | Informs | Informed |
| Pilot measurement and baseline | Requires and reviews | Owns and runs | Instruments | Informs | Informed | Informed |
| Cost classification | Informed | Supplies contracts and usage facts | Supplies technical facts | Informed | Determines | Tests |
| Who may approve a payment or post a journal | Owns policy | Follows | Enforces in systems | Reviews access | Owns procedure | Tests |
| Retention of AI-assisted records | Sets policy with legal | Follows | Implements | Reviews | Specifies audit trail | Tests |
Operating the system is not owning the outcome. An AP team that uses an extraction assistant still owns whether the right supplier was paid the right amount. Engineering operates the integration and enforces permissions; it cannot decide whether a proposed entry is correct.
Make the business case auditable
Require the baseline volume, eligible share, expected adoption, effort before and after assistance, and the quality requirements. Include the people who review exceptions and maintain the system. Separate recurring costs from setup. Distinguish quantities measured in the company from assumptions borrowed for planning.
Proposed minimum contents of a case that finance will review:
- Unit of work and denominator. A resolved request, an approved catalog record, a verified brief. Not "drafts generated."
- Measured baseline. Observed volume and effort for the unit, with the observation window stated.
- Assisted effort including review. Review and rework belong inside assisted minutes, not in a separate line that can be dropped.
- Continuing work. Knowledge upkeep, access review, quality sampling, and integration monitoring, each with an owner and a funding source. If existing staff absorb the work, record it as consumed capacity rather than cash.
- Realization mechanism. The specific payment, hire, contract, or sale that changes, the person with authority to change it, and the date.
- Downside case and stop criteria. Set before results are seen.
Review the downside before the upside
| Question | Why it changes the decision |
|---|---|
| What if review takes twice as long? | Labor benefits may disappear |
| What if only half the eligible work uses it? | Fixed costs are spread over fewer cases |
| What if usage increases sharply? | Variable costs may overwhelm pricing assumptions |
| What expense actually changes? | Capacity may not convert into cash |
| What happens after a supplier or model change? | Revalidation creates continuing expense |
Each expanded workflow chapter carries these sensitivities for its own example. In the support chapter, a seven-minute assisted effort instead of five turns a $3,400 monthly capacity surplus into a $200 shortfall. Require a useful baseline and an agreed quality floor. A cheaper process that produces costly mistakes is not an improvement.
Plan a first quarter
The company, team, and decisions below are a fictional teaching example. Thistlemark is an invented B2B software company. Its three pilot proposals borrow the illustrative inputs from the workflow chapters so the arithmetic can be traced there; they are not measurements from any company.
Weeks 1–3: review three proposals against one standard
| Proposal | Capacity after recurring cost | Realization mechanism offered | Cash result under the chapter's own assumption | Decision |
|---|---|---|---|---|
| Support reply drafts (support chapter) | $3,400/month | 70% of capacity replaces planned contractor hours | $1,780/month; $12,000 setup repaid in about 6.7 months | Fund a bounded pilot |
| Onboarding guidance (onboarding chapter) | $680/month | 50% replaces contractor onboarding hours | −$160/month; needs about 60% realization to break even | Fund measurement only |
| Incident briefs (incident triage chapter) | $433/month | None identified; 25% offered as a planning figure | −$567/month in cash; even counting all capacity value, the $15,000 setup takes about 34.6 months to recover | Decline as a labor case |
Three proposals, two very different answers
- Capacity after recurring cost
- Monthly cash result
Show the data
| Proposal | Capacity after recurring cost | Cash result | Realization basis | Decision |
|---|---|---|---|---|
| Support reply drafts | $3,400 | $1,780 | 70% replaces planned contractor hours | Fund a bounded pilot |
| Onboarding guidance | $680 | −$160 | 50% realization | Fund measurement only |
| Incident briefs | $433 | −$567 | 25% offered as a planning figure | Decline as a labor case |
The support proposal names a planned contractor expense that the support leader has authority to cancel, and its downside case still shows the break-even minutes. Finance approves a pilot with a stop criterion tied to measured assisted effort, not to satisfaction with the drafts.
The onboarding proposal fails on cash under its own assumptions but has an unmeasured baseline. Finance funds four weeks of baseline measurement, which costs little, and defers the tool decision until eligible volume and realization can be observed.
The incident-brief proposal is capacity only. Responders' time is not a line that changes, and no planned expense depends on it. Finance declines it as a labor case and tells the engineering leader what would change the answer: evidence that the brief improves response quality, measured as the incident chapter describes, and a sponsor willing to fund it as a quality investment rather than a saving. That is a legitimate decision, but it should be presented as one.
Weeks 4–8: finance's own first workflow
Thistlemark's AP team proposes assisted extraction of supplier invoices that arrive as PDFs. The proposed boundary: the assistant proposes header and line fields with a source location for each, validation rules check supplier identity, purchase-order reference, totals, and tax against master data, and an AP specialist approves every staged invoice before it enters the payables system. The assistant has no permission to create suppliers, change bank details, or release payments.
Before any tool is selected, AP measures the current unit: an invoice from receipt to approved entry, including exceptions and supplier queries. The document-extraction chapter covers field-level versus document-level evaluation and why confidence scores need local calibration; the order-intake chapter covers staging, duplicate prevention, and retries for an equivalent document flow. Finance reuses those designs rather than inventing its own.
Weeks 9–13: offline replay, supervised pilot, decision
Replay a held-out set of past invoices, including credit notes, invoices with changed bank details, duplicates submitted twice, and invoices whose purchase order was later amended. Then run supervised use on one supplier segment with manual entry available throughout. The controller decides expansion on counts with denominators: invoices approved without correction, field errors caught before approval, duplicates stopped, and total AP effort including exceptions.
The counterexample: do not automate payment release or journal posting
A vendor offers to let the assistant post matched invoices and schedule payment automatically when confidence is high. Finance declines. The PCAOB defines internal control over financial reporting to include procedures that provide reasonable assurance that "receipts and expenditures of the company are being made only in accordance with authorizations of management and directors." PCAOB AS 2201, paragraph A5 That definition sits in an auditing standard for US public-company audits, and it does not prohibit automation. It does make authorization the point that has to be designed, documented, and tested. A confidence score is not an authorization, and a change of bank details is the kind of instruction a fraudster could place in an invoice. Keep release and posting with named human approvers until an auditable control design has been reviewed by the controller and, where relevant, the auditors.
Treat classification as a separate question
Record what the business buys and what activity it supports. A customer-facing capability, internal sales tool, and development project may have different accounting implications. The controller should determine treatment under the applicable reporting standards, transaction facts, and compliant company policies. The field guide identifies economic mechanisms; it does not assign universal COGS, OPEX, or capitalization rules, and it gives no accounting opinion.
Two targeted source readings illustrate why the facts matter. For US GAAP, FASB's 2018 update on implementation costs for hosting arrangements that are service contracts addresses when such costs are capitalized and how they are presented; it does not classify every implementation invoice. FASB ASU 2018-15 A 2025 update changes internal-use software recognition guidance and applies to annual periods beginning after December 15, 2027, with early adoption permitted, so adoption status must be established first. FASB ASU 2025-06 Under IFRS, a 2021 agenda decision on configuration and customisation costs in a cloud arrangement turns on whether the customer controls a qualifying asset. IFRS agenda decision These are targeted readings of the source texts, not an accounting opinion, and no accounting specialist has reviewed them. Take the contract terms, the adoption status, and the facts of each cost to your auditor or technical accounting adviser.
This separation prevents the label from substituting for the investment case. An expense can be classified correctly and still be a poor investment.
Measure the right things
| Measure | Why it matters | How it can mislead |
|---|---|---|
| Realized cash effect per approved case | Shows which benefits actually changed a payment or plan | Timing lags; a hiring freeze can be attributed to AI when it had other causes |
| Capacity released, reported separately | Shows operational change even when cash does not move | Easily summed with cash into a single "savings" figure |
| Effort per completed unit, including review and rework | Tests the assisted-minutes assumption behind every case | Excluding exceptions or follow-ups flatters the result |
| Error rate on financial outputs, caught before and after approval | Separates model quality from reviewer diligence | Strong reviewers hide a weak assistant; count both stages |
| Continuing cost against budget | Captures upkeep, access review, and revalidation | Upkeep absorbed by existing staff disappears unless recorded as capacity |
| Pilots stopped at a gate | Shows that gates are real | A portfolio with no stopped pilots may have gates that do not bind |
Reconcile expected and observed adoption, effort, cost, and quality after every pilot. Record why results differ. A stopped experiment that prevents a larger bad investment is useful information, not a reason to hide the findings.
Your team: work, judgment, and skills
In the proposed AP design, keying shifts to checking: comparing each proposed field against its source location, resolving validation exceptions, and querying suppliers. The judgment that remains is the same judgment that matters today: whether this supplier, this amount, and this purchase order belong together. Specialists need to learn how the staging display presents uncertainty, which exceptions indicate a supplier problem rather than an extraction problem, and how to report a systematic error.
FP&A analysts reviewing drafted variance commentary need the discipline to check every figure against the ledger and every causal sentence against something they know. Automation bias, described in the authority chapter, is the reason review targets must reward corrections rather than throughput.
For the business-case job, finance staff need enough fluency to question an AI proposal: the difference between capacity and cash, why assisted minutes must include review, what a held-out evaluation set is, and why a vendor's accuracy figure on its own documents says little about yours.
Whether released AP or FP&A capacity leads to fewer contractor hours, redeployment to analysis, or a change in planned hiring is a decision for finance leadership to make explicitly and measure. This guide does not predict it. Protect practice: rotate specialists through unassisted processing so that the team can still process invoices by hand when the assistant is unavailable, and so that new staff learn what a correct invoice looks like.
Who to bring in. For finance's own first workflow, the AP extraction pilot, outside help can fill three roles: a builder for the extraction and staging step, integration work on validation against master data, duplicate prevention, and the handoff to the payables system, and an independent evaluator to replay held-out invoices, hard cases included, and report counts with denominators. Security or risk reviews access. Keep in-house the judgment of whether a benefit is realizable, cost classification with the controller, approval of every staged invoice, and authority over payment release and posting. See Staffing an AI pilot and Choosing and working with an AI partner.
Change management and governance
Proposed practices:
- Tell the team what the pilot is for and what it is not. If the goal is to absorb volume growth without adding contractors, say so. If a staffing decision may follow, say that it will be made separately, on measured results.
- Change targets before the pilot starts. An invoices-per-hour target during an extraction pilot rewards approving without checking. Measure corrected errors and exceptions resolved alongside throughput.
- Give reviewers a simple way to report failures and make clear that reporting a bad extraction is part of the job, not an admission of slow work.
- Record AI-assisted outputs as part of the audit trail. Keep the source document, the proposed fields, the reviewer's changes, the approver, and the configuration in use. Agree retention with legal and the auditors rather than defaulting to the vendor's setting.
- Hold access boundaries in systems, not prompts. Supplier-master changes and payment release stay behind existing role permissions.
Warning signs
| Signal | What it may indicate | Proposed response |
|---|---|---|
| Approved cases report annual savings but no named realization owner | Capacity is being counted as cash | Return the case; require the mechanism and the date |
| Approval rate rises and correction rate falls to near zero | Reviewers may have stopped checking | Sample approved items independently; revisit targets |
| Continuing costs appear in another department's budget | Upkeep is unfunded or hidden | Record it against the initiative, as cash or capacity |
| A vendor or model update arrives without revalidation | Tested performance no longer applies | Re-run the held-out set before resuming |
| A classification is chosen before the contract is read | The label is driving the case | Route to the controller with the facts listed in the brief |
| No pilot has been stopped in a year | Gates may not be binding | Review whether stop criteria were set before results |
Questions to ask vendors and internal teams
- What is the unit of work, and what was the baseline effort per unit, measured how and over what window?
- Do assisted minutes include review, correction, exceptions, and follow-up?
- Which specific expense, hire, contract, or sale changes, who can change it, and when?
- What continuing work does the system require, who does it, and is it new spending or existing staff time?
- What happens to cost if usage doubles? What does a model or version change require?
- For finance workflows: what can the system write, create, or release? Where is that enforced? What does the audit trail record for each item?
- Was the vendor's accuracy figure measured on documents like ours, at the document level or the field level, and with what denominator?
When to wait
Defer an AI business case when the baseline has not been measured, when the only benefit is capacity with no one prepared to change a plan, or when the process itself is unsettled. The simpler investment is measurement: a few weeks of observed volume and effort can decide the case more cheaply than a pilot.
Defer a finance workflow when master data is unreliable, when duplicate and exception handling is informal, or when approval rights in the ERP do not match policy. Fix supplier records, enforce matching rules, and tighten access first. Assistance built on weak controls makes errors faster, not rarer.
What external research establishes
For your analyst
The sources cited here are definitions and standards, not evidence about AI in finance. The PCAOB definition of internal control over financial reporting frames authorization as a control objective for US public-company audits; it does not address AI. The FASB and IFRS readings are targeted readings of the source texts, not reviewed by an accounting specialist. The NIST risk profile cited in the authority chapter names over-reliance as a risk without measuring it.
We did not find, and do not cite, attributable evidence on the effect of AI assistance on finance-team effort, invoice error rates, or close duration. No finance team was interviewed and no deployment observed for this chapter.
Related reading
Economics, economics worksheet, document extraction, order intake, inventory forecasting, pilot design, and operations leaders. Prepare pilots with the pilot kit.