A year or two ago, most writing about AI in finance was a forecast. Enough time has passed to check the forecast against what finance teams actually run. The honest picture is uneven. Some processes have handed real decisions to software, others use AI as an assistant to a person who still decides, and a few have barely moved.
This post goes process by process. For each one it answers three questions: what the AI now decides, what a person still owns, and how far along adoption really is.
How widely has AI been adopted in finance?
Gartner has surveyed finance leaders on this each year. Its 2024 survey found that 44% of finance functions use intelligent process automation, meaning AI added to existing automation tools such as RPA to handle information processing. The 2025 follow-up, as reported by CFO Dive, found adoption essentially flat year over year, with accounts payable automation and error and anomaly detection among the most common uses. Respondents named data literacy, technical skills and data quality as the main obstacles.
Definition: in this post, a process has moved to AI when software makes a decision a person used to make, such as which invoice a payment belongs to or which customer to contact first, and a person reviews exceptions instead of every item. A process where AI only drafts text or summarises a report for a person to act on is AI-assisted, which is a smaller step.
Collections: what does the AI decide, and what does the collector still own?
Collections was one of the first finance processes to move, because the inputs are structured (open invoices, payment history, contact records) and the decisions repeat daily across hundreds or thousands of accounts.
What the AI decides now: which accounts to work first, ranked by how likely each is to slip rather than by invoice age alone; when to send outreach and through which channel; and a first draft of each message shaped by how that customer has paid before.
What a person still owns: disputes, negotiated payment plans, credit holds on strategic accounts, and any conversation where the relationship matters more than the invoice. The collector reviews the queue the software builds and handles the exceptions.
How far along: widely adopted for scheduling and reminders; less common for genuine prioritisation, where the software decides the order of work. Many teams still run rule-based dunning, which sends the same sequence to every customer on a fixed day count. That is automation, and it works for a predictable book, but it does not decide anything.
Tesorio customers have reported an average DSO reduction of 33 days and 3x collector productivity, figures from Tesorio's own customer base rather than an industry benchmark.
Cash application: can AI match payments without a person?
Cash application is matching incoming payments to open invoices.
What the AI decides now: which invoice or invoices a payment settles, including partial and bundled payments, by reading remittance details and learning from how past payments were applied. When confidence is high it posts the match; when it is low it proposes one.
What a person still owns: short payments, deductions and unidentified cash, plus the decision about what counts as high enough confidence to post automatically.
How far along: mature where remittance data is clean, uneven where it is not. As one data point, in a proof of concept for a customer processing over a million invoices a year, Tesorio reached a 78 percent automatic match rate across 1,150 ACH, wire and lockbox payments before any manual configuration. The remaining share is exactly the exception work a person still owns.
Accounts payable: which parts of AP have moved to AI?
On Gartner's reporting, AP automation sits among the most common AI uses in finance.
What the AI decides now: extracting header and line data from invoices in any format, coding them to the right accounts, matching them to purchase orders and receipts, and flagging likely duplicates or unusual amounts for review.
What a person still owns: approving payment, resolving mismatches with suppliers, and vendor master changes, which are a common fraud target and deserve a human check every time.
How far along: one of the furthest along of any finance process, largely because invoice capture was already digitised before generative AI arrived.
There is a related process on the receivables side that is often confused with AP. Many large customers require suppliers to submit invoices through the customer's own AP portal, and payment status lives there too. AP portal monitoring means software logs into those customer portals, submits invoices and reads back status and rejections, so the AR team stops checking each portal by hand. It sits in order-to-cash.
Cash forecasting: does AI replace the forecast or the spreadsheet?
Forecasting used to mean collecting inputs from several teams into a spreadsheet, then adjusting by judgement.
What the AI decides now: expected payment dates per invoice and per customer, based on each customer's actual payment behaviour, which rolls up into a receipts forecast that updates as payments arrive.
What a person still owns: the assumptions the model cannot see, such as a large customer's known budget freeze, an acquisition, or a change in payment terms, and the final forecast that goes to leadership.
How far along: receivables forecasting from payment behaviour is established; full cash forecasting that blends AR, AP, payroll and treasury remains mostly spreadsheet-led with AI inputs. The gap is usually data: AR, AP and bank data sit in different systems and rarely share a customer or vendor key.
Close and reconciliation: where does anomaly detection fit?
Error and anomaly detection is one of the most common AI uses Gartner has reported in both surveys cited above. In the close, it means software scans journal entries, account balances and reconciliations and surfaces what looks wrong.
What the AI decides now: which entries and variances are unusual enough to review, and in many tools a proposed explanation or matching entry for routine reconciliations.
What a person still owns: the judgement about whether an anomaly is an error, the accounting treatment, and sign-off.
How far along: AI-assisted rather than moved. The software narrows the search; a person still decides.
What do the processes that moved have in common?
Across the five, the processes that moved furthest share three traits: structured data, a decision repeated at high volume, and a clear exception path back to a person. Collections, cash application and AP invoice capture have all three. Forecasting and the close have the data but carry decisions where the cost of a wrong answer is high, so people keep the final call.
The other common thread is connection. The order-to-cash cycle runs from credit through collections, cash application and the receivables forecast, and each step feeds the next. A team that automates one step in isolation, such as reminders, gets a faster reminder. A team whose software sees the whole cycle can let a failed payment change the collection priority and the forecast in the same pass. For many teams a single-step tool is the right fit; for others the handoffs between steps are where the time goes.
Key takeaways
Finance AI adoption has plateaued in breadth and is now about depth: which decisions software is trusted to make. Collections prioritisation, cash matching and AP invoice processing are the processes where AI now decides and people handle exceptions. Forecasting and the close use AI to inform a person who still decides. The main obstacles are data quality and skills, which is why connected data across a process matters more than any single feature.
If you want to see what it looks like when software ranks your receivables book and drafts the outreach itself, see how Tesorio's AR agent works.
