Finance
Close the books faster, catch invoice errors before payment, and run real-time cash flow scenarios.
Updated 4 days ago · 12 sources
Proof over hype
- NewEnterprise
HSB
HSB processed 1.5 million invoices a year through more than 300 accountants, and every invoice got entered twice: once into an OCR (optical character recognition) system, then again by hand into a separate coding platform, because the two systems could not talk to each other.
Over 25,000 hours a year saved, 72% of invoices processed with no human touch, at 96% coding accuracy — on the vendor's account.
The takeaway The lesson isn't the hours-saved figure, it's the sequencing: the system earned autonomy gradually, invoice by invoice, as it proved itself against HSB's own data — it wasn't switched to full autopilot on day one.
Reported by Vic.ai · checked Aug 2026
How they did it
HSB moved onto Vic.ai in June 2020 and collapsed the two-step process into one: invoices are ingested automatically and coded and classified by the system itself, with staff validating the output instead of typing it in from scratch. As Vic.ai built up confidence in how HSB's own invoices typically looked, it started letting the highest-confidence ones go straight from ingestion to approval with no human touch at all — a feature called Autopilot, switched on in the tool's first full month live. A year in, coding accuracy sat at 96%, 72% of invoices moved through with no human touch, and the average handling time per invoice dropped from over two minutes to 45 seconds.
Tool used Vic.ai
- NewEnterprise
Georgetown University
Georgetown processed $400 million a year in supplier invoices through just six full-time AP staff, entirely by hand. Purchase orders got created reactively, after AP chased departments down for missing paperwork, and invoices sat 30 to 90 days behind schedule.
AP cycle times down 76% (roughly 30 days to 7) and 1,600 hours saved in the first six months, with 66% of invoices auto-approved — on the vendor's account.
The takeaway The real change wasn't speed, it was turning a data-entry team into people who could actually prioritize — because the system did the triage that used to eat their day before they ever opened an invoice.
Reported by AppZen · checked Aug 2026
How they did it
The university deployed AppZen's Autonomous AP in August 2022 to capture, standardize and prioritize invoices arriving mostly as PDF email attachments, rather than leaving that sorting to staff. AP director Richard Marea had described the old process as one that turned his team into 'a team of clerical workers who were used to keying data rather than a team of analysts' — the new system let them work by due date, amount or supplier instead of whatever order invoices happened to arrive in. Within the first six months, cycle times fell from roughly 30 days to about 7, and two-thirds of invoices were approved with no human touch at all.
Tool used AppZen
- NewEnterprise
Westbury Street Holdings
Westbury Street Holdings runs £1.7 billion a year through 20-plus subsidiary businesses across 7 countries and 5 currencies, with no shared version control or audit trail across the group — exactly the conditions that let the same invoice get paid twice without anyone noticing until much later, if ever.
£774,000 in duplicate invoices stopped in the first four months, processing time down 38% within three months, and a 2.4x increase in fully autonomous processing — on the vendor's account.
The takeaway Duplicate payments are the easiest category for a tool like this to catch, because the check is mechanical — does this invoice match one we already paid. Start an AP pilot there before anywhere judgment-heavy.
Reported by AppZen · checked Aug 2026
How they did it
The group put AppZen's AP Inbox Service Center in front of every invoice, automatically labelling and categorizing what arrived by email before a person opened it, and checking each one against prior payments before release rather than sampling after the fact. UK and Ireland financial controller Colin Yates said the system gave the finance team something the group had never had: 'an audit trail — you can see the original email and each part of the invoice, whether it's been actioned by automation or manually.' In its first four months, the system caught £774,000 in invoices that would otherwise have been paid twice.
Tool used AppZen
- Enterprise
Hewlett Packard Enterprise
Every Monday, 40 to 50 HPE finance and sales leaders sat through an operational review built on roughly 100 PowerPoint slides that took days to compile — a call CFO Marie Myers called 'a very manual-driven, rear-vision mirror look-back.'
Financial reporting cycle cut by about 40%.
The takeaway The deterministic-answers design choice is the transferable lesson: if you want a finance tool people will trust in a room with the CFO, engineer it to give the same answer to the same question every time, not just a plausible one.
How they did it
HPE built an internal agentic AI tool, now called CFO Insights, with Deloitte and Nvidia. It pulls live data across HPE's finance and supply chain systems and generates the analysis a team used to assemble by hand, including the shipment-data calculations and accuracy checks analysts used to run manually. The team deliberately engineered it for deterministic outcomes — the same question returns the same answer every time — specifically to avoid the unpredictability that makes people distrust a typical AI chatbot. That freed the Monday call to spend its time on what the company plans to do next instead of re-litigating what already happened. HPE has since started extending the same approach into transactional finance work: credit and collections, and accounts payable and receivable.
Tool used Deloitte
- Enterprise
Uber
Uber's finance team processes enormous invoice volume plus regulatory notices arriving in dozens of languages — the kind of high-volume, rules-based work that used to justify large back-office teams.
Over 96% of invoices processed by AI at over 95% accuracy, regulatory notice turnaround cut about 70%, and more than 90% of finance staff using AI tools routinely.
The takeaway Invoice processing is where finance AI is furthest along — start there, not with forecasting.
How they did it
Uber put AI through invoice processing, contract review for revenue recognition (when a sale counts as income), and regulatory notice handling. Finance director Tiho Nedkov says more than 90% of the finance team now uses AI tools routinely, and credits finance with an early lead inside the company: 'Finance actually led this within Uber for a couple years… at least until agentic coding came along' — engineering has since overtaken finance as Uber's heaviest AI user, burning through the company's full 2026 AI budget in four months on coding tools alone. Nedkov's own caution is worth keeping: even with these adoption numbers, he says finance is still in the 'early innings,' and unlike engineering, 'there's no token-maxxing in finance' — the team isn't chasing usage for its own sake.
- Enterprise
Alphabet
Paying and reconciling invoices, and treasury work (managing the company's cash), ate time in Alphabet's back office even though the teams doing it were small.
The takeaway When a company discloses detailed AI numbers everywhere except finance, that gap is informative on its own — the transactional wins here look real, just earlier and smaller than the engineering headlines suggest.
How they did it
Alphabet's finance team put AI agents to work on paying and reconciling invoices and inside its treasury function. CFO Anat Ashkenazi said the push reaches 'all the way from the engineering team to small teams within our back office, even with my finance team' — part of the same agentic push that now has roughly half of Alphabet's own software code written by AI agents on the engineering side. Alphabet has not disclosed hours saved or accuracy figures for the finance deployment specifically, which is itself worth noting: a company this open about its engineering AI numbers stayed quiet on finance-specific ones.
- Learn from thisEnterprise
Deloitte Australia
Australia's Department of Employment and Workplace Relations paid Deloitte roughly A$440,000 (about US$290,000) for an independent review of the IT system that automates penalties in the country's welfare compliance program.
Refunded roughly A$97,000 (about US$64,000) of the A$440,000 contract, and added an AI-use disclosure the original report never had.
So: Before you pay for a report, from Deloitte or anyone, ask in writing who checked its sources against the originals, and hold back part of the fee until you get a straight answer.
How they did it
Deloitte used generative AI, later confirmed to be Microsoft's Azure OpenAI service running GPT-4o, to help produce the review. University of Sydney researcher Chris Rudge read the published report and found the consultants had invented the references: nonexistent academic papers attributed to real researchers, and a quote attributed to a Federal Court judgment that was never written, misspelled judge's name included. Once the errors became public, Deloitte published a corrected version months later that added a disclosure of AI use the original report never carried, while maintaining that 'the substance' of its findings and recommendations stood.
- Enterprise
EY
EY's assurance (audit) teams spent hours per engagement summarising documentation and hunting through accounting guidance for the rule that applied to the transaction in front of them.
The takeaway The embed-don't-bolt-on choice is the transferable move: EY put the AI inside the tool auditors already live in every day, instead of asking them to open a separate app — the same test worth applying to any tool you're evaluating for your own team.
How they did it
EY embedded AI agents directly inside EY Canvas, its own audit platform, and rolled them out to assurance professionals worldwide rather than shipping a separate opt-in tool. Global assurance transformation leader Marc Jeshonneck said the design choice was deliberate: 'to embed AI directly into the audit platform, so our auditors aren't having to navigate several separate tools.' Today the agents handle project management and administrative work, summarise audit documentation, and search accounting guidance. EY has already scheduled the next step for later in 2026: having the agents reconcile documents to outside evidence, with a goal of supporting the full end-to-end audit by 2028.
Try this today
Pull the money terms out of a vendor contract
15 minA one-page list of every price rise, auto-renewal and notice date buried in the agreement.
Copy the prompt
You are a contract analyst working for a controller. Below is one vendor agreement. Find these items: the total fee and how it is worked out, any automatic renewal, the notice period to cancel, any price rise clause and its cap, late payment interest, and any minimum spend. Return a table. Columns: item, clause number, exact quote, what it means for us. If the contract does not say, write 'Not stated'. Do not fill gaps with what is normal for this kind of deal. Put anything unclear in a row marked 'Ask the vendor'. Contract text: [PASTE CONTRACT TEXT]
One check first. Check every quoted clause in the original. It misses exhibits, side letters, and scanned pages. Confirm your firm allows you to paste that contract.
Draft board commentary from your month-end numbers
20 minA first draft of your variance commentary — budget against actual — for readers who skip spreadsheets.
Copy the prompt
You are a finance business partner. Your reader is a board director who is not an accountant. Below is a variance table — budget against actual, line by line. Write no more than 200 words of commentary. Cover only the five largest differences by value. For each one, give the line, the size of the gap, and the driver. Use a driver only if my notes below state it. Where the cause is not in my notes, write 'Cause not yet confirmed'. Do not guess. Write plain English and avoid accounting shorthand. Then list the questions a director would probably ask. Variance table: [PASTE TABLE]. What I know about the causes: [PASTE YOUR NOTES]
One check first. It will invent a plausible cause if your notes leave a gap. Every driver in the final version must be one you can evidence.
Stress-test the assumptions in a business case
25 minThe questions your board will ask about a forecast, before the meeting instead of during it.
Copy the prompt
You are a skeptical audit committee chair. Below are the assumptions behind a business case. For each assumption do three things. Say what has to be true in the real world for it to hold. Name the input that would hurt the answer most if it is wrong. Rewrite the assumption as a downside case. Then show the payback period (how long before the case earns back its cost) if the three weakest assumptions each miss by 20%. Show your arithmetic line by line so I can check it. Do not add any assumption I have not given you. If an input is missing, ask me for it. Assumptions: [PASTE ASSUMPTIONS]
One check first. Redo the arithmetic yourself. These tools reason well and add badly. The downside case is only as good as the inputs you gave it.
Specialized tools, and what to ask vendors
| Tool | What it does | Setup | Best fit |
|---|---|---|---|
| Rampfree to mid-market | Runs corporate cards and expenses, and flags spending that looks wrong: duplicate subscriptions, price rises, out-of-policy purchases. | WeeksIT sign-off | More than a handful of departments buy software on their own cards.Skip it ifYour card programme is already centralised and you can see every purchase before it happens. |
| Trullionmid-market to enterprise | Reads contracts and source documents, then pulls out the numbers your accounting treatment depends on. | MonthsIT sign-off | Heavy lease and contract work, with documents scattered across systems your auditors have to trace.Skip it ifSimple contracts where a spreadsheet still does the job without anyone complaining. |
| Microsoft Copilot in Excelincluded in existing licensing | Explains, builds, and checks spreadsheet work in plain language, inside the file you already use. | Under a week | You already pay for it, so try it before you buy anything else.Skip it ifYour spreadsheets have outgrown themselves and you need a real planning system, not a better spreadsheet. |
| Vic.aiquote only | Reads incoming invoices, codes them, matches them to purchase orders and routes them for payment without a person touching most of them. | MonthsIT sign-off | You process enough invoices that a full-time person is doing data entry, and they arrive as PDFs and email attachments.Skip it ifA few hundred invoices a month. The setup and the integration work cost more than the clerk's time you would save. |
| MindBridgequote only | Scores every transaction in the ledger for risk, rather than testing a sample, and shows its reasoning for each one it flags. | MonthsIT sign-off | You own internal audit or financial control and want the whole population looked at rather than a sample.Skip it ifYour ledger data is inconsistent between entities. It will flag your data quality, at length, before it finds any fraud. |
| AppZenquote only | Audits expense claims and invoices against your own policy, in the background, and acts on the clear-cut ones. | MonthsIT sign-off | Large travel and expense spend across several countries, where nobody can read every claim.Skip it ifYou have no written expense policy. There is nothing for it to check against, and it will not write one for you. |
| Your existing accounting system vendorpossibly already paid for | SAP, Oracle, NetSuite, Workday and Sage all ship AI inside the ERP (your core accounting system) you own. | Talk to them | Before you evaluate anything new. The call is free and it often kills a purchase. |
Questions to ask before you buy
Ramp — 7 questions to ask them
- How do you make money when my team spends on your free card?
- What credit limit will you underwrite for us, and what happens to it if our cash balance drops?
- Which spend flags fire most often for a company our size, and how many turn out to be wrong?
- How does the sync into our accounting system handle multiple entities and currencies, and where does it break?
- If we leave, what transaction and receipt history do we get back, in what format, and how fast?
- What does your product not cover, so we would still need a separate tool for it?
- Which customers moved off Ramp in the last year, and what did they say on the way out?
Trullion — 7 questions to ask them
- Run ten of my ugliest scanned contracts through it live, and show me what it got wrong.
- What happens when the model is unsure — who sees it, how fast, and does anything post without review?
- How do you handle amendments and side letters that change a lease after the original was read?
- Show me every step from a page in a source PDF to the number posted in my accounts.
- Has my external audit firm signed off on your documentation before, and on which engagements?
- When you update the model, can my extracted numbers change afterwards, and how would I find out?
- Which customers have my document types and my accounting system, and can I call one?
Your existing accounting system vendor — 6 questions to ask them
- Which AI features are in my current licence today, and which ones cost extra?
- Which of those are generally available now, and which are early access or roadmap?
- What version do I have to upgrade to, and what does that migration cost me in time?
- Name three customers on my version running these features in production, not in a pilot.
- When your fraud and error checks flag something, what evidence do they hand my auditors?
- What do these features do badly, and where would a specialist tool genuinely beat you?
What everyone is asking
- NewJul 2026
Best AI for closing the books faster?
No product wins outright. Reviewers rate FloQast for tracking who owes what and BlackLine for matching records, but the real delay is waiting on data from other teams.
What to watch for
Reviewers split on setup cost. BlackLine users say configuration takes a long time; Adra users find it quicker but slower to load.
Also worth a look. FloQast, BlackLine Financial Close Management, Workiva, Adra by Trintech, OneStream
- NewJul 2026
Is AI safe to use on numbers my auditors will see?
Yes for drafting and matching, as long as a person checks the output before it posts. Regulators are warning that finance teams trust the machine too readily.
What to watch for
Audit firms are shipping AI agents into every engagement while the PCAOB (the US audit regulator) warns the same tools erode professional skepticism.
Also worth a look. AI for drafting and summarising, AI for matching records, Fully automated posting with no human review
- NewJul 2026
Will AI cut finance headcount and junior roles?
Transactional work carries the risk. About half of finance chiefs expect AI to reduce headcount in accounting, paying suppliers and collecting cash. Tax, treasury (managing the company's cash) and investor relations look steady.
What to watch for
The surveys disagree. Only 30% of finance chiefs expect any cut at all, and most of those expect under 10%.
Also worth a look. accounting and month-end roles, supplier payment roles, cash collection roles, tax, treasury, investor relations
- NewJul 2026
Best AI for invoices and accounts payable?
Nothing has clearly won. Uber runs most of its invoices through AI. Workday's fraud and error checker is still in limited release, and buyers are arguing about audit evidence.
What to watch for
Buyers argue about audit evidence more than accuracy. A tool can get the right answer and still fail an audit because it cannot show its work.
Also worth a look. Workday Financial Test Suite, Medius, Uber's in-house build
Worth following
CFO Dive
Daily news desk covering the finance function.
daily · publication
Why them
It names the companies behind the AI stories, which most trade coverage does not.
CFO Brew
Daily finance and accounting briefing from the Morning Brew team.
daily · newsletter
Why them
Short, readable, and honest about AI projects that have not paid off yet.
Francine McKenna — The Dig
Accounting and audit journalist who covers public company reporting.
weekly · substack
Why them
Skeptical and technical, and the counterweight to vendor optimism about audit technology.
The Accounting Podcast
Blake Oliver and David Leary run a weekly accounting news show.
weekly · podcast
Why them
They test the tools on air and say plainly when something does not work.
The Footnotes Analyst
Steve Cooper and Dennis Jullens write on financial reporting and valuation.
occasional · publication
Why them
Careful writing on what the numbers mean, which keeps AI output in perspective.
CFO.com
Long-running publication for finance leadership.
daily · publication
Why them
Practical coverage of month-end work, controls, and finance technology.