Before the news lineup, here is my own news: Balanced AI Insights turned two last month, and I think it's time for a change. The first issue went out in September 2024, and since then the newsletter has grown alongside how I've worked with finance teams on AI.
As a first step, I'm removing the paid tier. I value my paid subscribers, but the paid sections took a lot of effort to produce every week, and only a small group of readers ever saw them; I'd rather put that work where everyone can use it. This is the last edition with a subscriber-only section, and starting next week the newsletter will be free for everyone.
I'm also rethinking the format, and I'll keep you updated as it takes shape.
If you would like to share anything (what you read, what you skip, what you wish I wrote about), just reply to this email; I'd be happy to hear your thoughts.
Blend2Balance is now a Chapter Partner of the FFA
I'm joining the Fractional Finance Association as a Chapter Partner and Sponsor across all FFA chapters nationwide.
Fractional finance can be a lonely seat, and the people in it deserve a community of their own. Membership is free, and chapters run both virtual and in-person events; I'm working with Mark Tranter on shaping what this partnership will look like, so there will be more to share soon. If you work as a fractional CFO, controller, or FP&A lead, register here!
What Happened in September
Companies like their AI; the P&L hasn't noticed yet
McKinsey's State of AI 2026 surveyed 1,719 people: 44% of companies now scale AI across the enterprise (up from 38% last year), but the share reporting any EBIT contribution from AI is flat at 37%, and only 6% attribute 5% or more of EBIT to it. Grant Thornton's Q3 CFO survey of about 230 US finance leaders shows the same split from the finance side: 84% say AI ROI is meeting or exceeding expectations, while only 33% see higher revenue from it.
So what: the two numbers fit together once you look at how most companies measure AI; they track usage and time saved, and very few trace it to a line on the P&L, so "meeting expectations" says as much about the expectations as about the results. One more number deserves attention: last year 32% of McKinsey's respondents predicted AI-driven headcount reductions and only 14% saw them, yet 39% expect them next year.
What to do:
Take your three largest AI use cases and write down which P&L or balance sheet line each one should move, and by how much. If you can't name the line, treat it as a productivity tool and budget it that way.
In any AI business case, model avoided hiring rather than headcount reductions, unless the reductions are already in a plan with names and dates.
Gartner: AI literacy is now the biggest barrier in finance
Gartner surveyed 160 senior finance leaders and found that data extraction, AP and AR automation, and report creation pay back in 9 to 10 months, while data management, insight generation and forecasting take longer. Low AI literacy ranked as the biggest barrier, ahead of finding talent.
So what: I have been saying for a while that AI literacy matters more than the tool you pick, and it's good to see Gartner put a number behind it. The second point is less obvious: the quick wins are real, but if they take all the budget, the forecasting and decision-support work never starts.
What to do:
Split your AI list into quick-payback items and longer bets, and keep at least one longer bet funded this year.
Give your team hands-on time on real work (Gartner suggests project assignments, sandbox time and short on-the-job tasks) instead of a one-time demo.
Skills come to ChatGPT, Gemini and Copilot
OpenAI is retiring custom GPTs and replacing them with plugins built from skills: Enterprise workspaces can't create new GPTs after October 26, custom GPTs retire on December 11, and some Enterprise customers can defer to February 11, 2027. Sharing settings don't carry over, and OpenAI warns that migrated versions may respond differently. Google started replacing Gems with skills on September 30, with personal accounts switching in November and Workspace in March 2027. Microsoft added reusable Copilot skills to Excel in June, built on open-standard markdown files.
So what: both building blocks the other tools are now adopting started at Anthropic: MCP for connecting AI to your data, and skills for packaging how you work. That is one of the reasons I keep building my own workflows in Claude and teaching it. The other reason matters just as much for your team: because both are open standards, a skill you write today isn’t locked into one vendor, so the time you put into it stays useful even if your company changes its tool choice.
What to do:
List your team's custom GPTs and Gems now, with an owner for each and what it's connected to.
Migrate and test them on real work before the deadline, then reset sharing by hand.
Write any new reusable workflow as a skill, so it can move with you.
Finance and AI vendors
Reuters reviewed Anthropic's draft IPO prospectus, filed confidentially with the SEC in June, and TechCrunch, Fortune and The Next Web followed with the details. 2025 revenue was nearly $4.6 billion, up twelvefold, with an operating loss of about $8 billion; the reported $42 billion net loss includes an accounting charge of roughly $34 billion tied to future share conversions. About a quarter of 2025 revenue came from two customers, the company plans $518 billion in compute commitments, and nearly a third of the document covers risk factors.
So what: I build and teach on Claude, so I read this one closely. Your AI vendor is now a company whose numbers you can read like any other supplier's, and revenue concentrated in two customers plus very large compute commitments mean pricing and terms may change once public investors are watching. I don't see that as a reason to switch tools; I see it as a reason to treat your AI vendor like any other critical supplier.
What to do:
Add your main AI vendor to your critical-vendor review, next to your ERP and your bank.
Check your contract for price-change notice periods, data terms and what happens to your data if you leave.
Keep your workflows portable: instructions, skills and prompts stored in your own files, not only inside one tool.
The Big Four now test every transaction
The Financial Times reported that KPMG and EY now use AI to analyze entire transaction populations instead of samples; one KPMG executive compared it to "filtering an entire river rather than collecting a bucket of water." The UK Financial Reporting Council flagged the risks: overreliance on systems nobody can see into, junior auditors losing basic skills, and thin oversight.
So what: when the auditor sees 100% of your transactions, every manual journal, odd vendor coding and round-number entry becomes visible, and the questions come earlier. The deskilling point applies to finance teams too, if juniors stop doing reconciliations themselves.
What to do:
Before year-end, run your own full-population check on a GL export: manual entries, weekend postings, round amounts and duplicate payments.
Ask your auditors in the planning meeting what their AI tools flag, so you see it before fieldwork.
Keep juniors doing some reconciliations by hand, so they still know what "wrong" looks like.
Agents start moving money, and a person still releases
Danske Bank and Mastercard completed Denmark's first AI agent payment, a pilot in which a consumer had an agent book a coffee tasting, with consent confirmed by passkey. At Sibos, banks described live agent workflows: BNY's agent handles over 10% of its global payment repairs, and a person still releases each payment. Six banks, including Bank of America, Capital One, ING and NatWest, published a paper on agent liability warning that instant payments can't be pulled back and that banks often can't see an agent's decision log.
So what: I have said I won't let agents touch the GL or cash yet, and the banks are clearly moving faster than that. Look at where they draw the line, though: agents prepare, repair, and check, and a person releases the money. That's the line I would draw for a finance team too; agents can prepare, but they don't post or release.
What to do:
Map your payment process into three stages: prepare, approve and release.
Look for agent candidates only in the prepare stage, such as invoice matching, payment file checks and exception flags.
If a vendor pitches agentic payments, ask who is liable when the agent goes beyond its authority, and whether you get the full decision log.
That's the news for September.
Paid subscribers: one more section below for you (last one).
Closing Thoughts
Thank you for reading, and for two years of replies, questions, and pushback; they shaped this newsletter more than any plan I started with. Next week, the newsletter will be free for everyone, and I'll share more about the new format as it comes together.
Until next Tuesday, keep balancing!
Anna Tiomina
AI-Powered CFO