Finance & Accounting

Sub-ledger transactions → account reconciliations → adjusting journal entries → trial balance → financial statements → management reporting package / flux commentary

• Automate
  Transaction matching & reconciliation — AI matches bank/sub-ledger/GL items, clears the 95% that tie out, and queues exceptions with suggested causes · high volume, rule-governed, and every match is verifiable against source records
• Draft
  Accrual & adjusting entry preparation — AI drafts recurring accruals from POs, invoices-in-flight, and prior-period patterns, with support attached · verifiable inputs but entries hit the books, so a preparer must own each one before posting
• Assist
  Estimate-heavy entries (reserves, impairments, revenue judgments) — AI assembles the data pack and prior methodology; the human makes the call · low verifiability, high stakes, and auditors will challenge the judgment, not the arithmetic
• Draft
  Flux analysis — AI computes period-over-period variances and drafts explanations by tracing drivers to underlying transactions · explanations are checkable against the GL but causal narratives are frequently plausible-and-wrong
• Draft
  Management reporting package assembly — AI populates the deck from the closed TB, formats, and drafts commentary in house style · mechanical assembly is safe but commentary shapes executive decisions, so the controller owns the final read
• Avoid
  Close sign-off & certification — reviewer attests reconciliations and statements are complete and accurate · SOX/attestation liability sits with a named human; delegating the certification defeats its purpose

Planning assumptions memo → department budget templates → consolidated model → variance-to-plan analysis → reforecast → board budget deck

• Draft
  Assumption gathering & prior-year analysis — AI summarizes actuals trends, seasonality, and last cycle's forecast misses per department · fully verifiable against actuals, but framing the assumptions steers the whole cycle
• Automate
  Template pre-population — AI pre-fills department templates with run-rate baselines and known contract changes · mechanical, checkable, high-volume, and errors surface immediately in review
• Assist
  Driver-based model building — AI writes and audits formula logic, flags broken links and hardcodes · the model architecture encodes business judgment; AI is a checker and pair-builder, not the author
• Assist
  Negotiation & target-setting with budget owners — AI preps talking points and scenario comparisons for each meeting · this is a relationships-and-politics stage; the artifact that matters is the agreement, not the document
• Draft
  Reforecast updates — AI rolls actuals into the forecast, reruns drivers, and drafts a bridge from prior forecast · repetitive and verifiable, but forecast changes trigger real resource decisions so FP&A owns the number
• Draft
  Board budget narrative — AI drafts the story arc from the consolidated model · high stakes and audience-sensitive; CFO voice and defensibility require human ownership

Invoice/receipt → coded & matched voucher → approval record → payment run / dunning sequence → aging report

• Automate
  Invoice capture & GL coding — AI extracts fields, codes to account/department/PO, and 3-way matches against PO and receipt · massive volume, deterministic verification against POs, cheap to spot-check
• Automate
  Fraud & duplicate detection — AI flags duplicate invoices, bank-detail changes, and out-of-pattern vendors for human review · AI as detector with human disposition; false positives are cheap, misses are what humans were bad at anyway
• Draft
  Exception resolution & vendor correspondence — AI drafts emails chasing missing POs, price discrepancies, W-9s · low stakes per message, but vendor relationships and edge-case terms need a human eye before send
• Avoid*
  Payment run approval & release — human approves the batch; AI pre-audits it for anomalies (Avoid for release itself) · irreversible cash movement plus fraud exposure makes autonomous payment release a segregation-of-duties violation
• Draft
  Collections & dunning — AI sequences reminders by customer risk profile and drafts escalating notices · high-volume and templated, but tone toward a strategic customer is a relationship decision
• Automate
  Aging & DSO/DPO reporting — AI generates aging analyses and cash-impact summaries on demand · pure computation over system-of-record data, fully verifiable

Business question → data pull → analysis workbook → scenario model → recommendation memo or deck slide

• Assist
  Question framing & metric definition — analyst pins down what's actually being asked and which definitions apply · context lives in hallway conversations and org politics; AI helps sharpen, can't source it
• Automate
  Data pull & cleaning — AI writes the SQL/queries, joins sources, and flags definitional mismatches between systems · verifiable by inspection and rerun; the classic 60% time-sink with near-zero judgment content
• Draft
  Exploratory analysis — AI runs cohort cuts, margin bridges, and sensitivity checks, surfacing what moved and why · fast hypothesis generation, but spurious correlations are fluent, so the analyst must interrogate every finding
• Assist
  Scenario modeling — AI builds the scenario toggles and stress cases on the analyst's model structure · which scenarios matter is the judgment; the mechanics are AI-friendly
• Draft
  Recommendation memo — AI drafts the memo from the analyst's conclusions in pyramid-principle structure · writing is AI-strong, but the recommendation carries the analyst's name and drives a real decision

Risk & control matrix → control test plans → evidence packages (PBC list) → testing workpapers → deficiency log → management responses

• Automate
  PBC evidence gathering — AI pulls, labels, and organizes requested documents against the auditor's list, flagging gaps · pure retrieval-and-index against an explicit checklist, fully verifiable
• Draft
  Control test execution — AI performs full-population testing (e.g., every user-access change vs. approval ticket) instead of sampling · results are verifiable, but a control tester must own the conclusion each control "operated effectively"
• Draft
  Workpaper documentation — AI drafts test procedure narratives and tick-mark explanations from the executed work · templated and checkable, but workpapers are what regulators re-perform against
• Avoid
  Deficiency evaluation & severity classification — is it a deficiency, significant deficiency, or material weakness? · a judgment with SOX and disclosure consequences; regulatory exposure and low verifiability both point the wrong way
• Draft
  Management response drafting — AI drafts remediation plans from the deficiency description and prior remediations · useful first pass, but management is accountable for commitments it can actually keep
• Automate
  Continuous controls monitoring — AI watches transaction streams for control breaks (duplicate approvals, threshold splitting) between formal test cycles · high-volume anomaly detection where every alert is human-dispositioned

Bank balance feeds → daily cash position → 13-week cash forecast → funding/investment decision memo → covenant compliance certificate

• Automate
  Daily cash positioning — AI aggregates multi-bank balances, categorizes flows, and produces the morning position across entities and currencies · verifiable against bank feeds, daily repetition, and errors surface within 24 hours
• Draft
  13-week cash forecasting — AI builds the forecast from AP/AR schedules, payroll calendars, and historical patterns, with variance-vs-actual learning each week · inputs are systematic but lumpy items — deal closings, tax payments, delayed customer receipts — need treasury's private knowledge
• Assist
  FX & interest-rate exposure analysis — AI quantifies open exposures and models hedge scenarios · analysis is computable, but hedging decisions commit real money under a board-approved policy
• Avoid
  Funding & investment execution — moving cash, drawing revolvers, placing short-term investments · irreversible movement of large sums; even "AI recommends, human clicks" needs dual control and payment-fraud paranoia
• Draft
  Covenant monitoring & compliance certificates — AI computes covenant ratios from the closed TB, drafts the certificate, and projects headroom under forecast scenarios · arithmetic is verifiable, but the certificate is a legal representation to lenders signed by an officer