June 2, 2026
Hi Everyone,
A 2025 MIT study found that companies spent around 70% of their AI budgets on sales and marketing tools. But the best returns came from back-office work like finance and document processing.
Most companies build a customer-facing AI project first, like a chatbot or a sales copilot. That's the wrong place to start. The returns are smaller and the failures more common than in back-office work.
Today we're walking you through where the AI returns are showing up in operations and how to pick your first project so it pays for itself.
Where are the wins coming from
In 2025, Stanford and MIT Sloan looked at 79 small and mid-sized firms using AI in their accounting work. The teams using AI cut their monthly close time by 7.5 days and supported 55% more clients per week.
Adjacent operations are showing the same kind of result. A 2025 Hackett Group benchmark of accounts payable platforms found that companies with mature AI in Accounts Payable are now processing 60% of invoices without human review, at cycle times 59% faster than before.
In both studies, the work was repetitive and well-documented. The team saw results within weeks because the data was already clean.
Why customer-facing AI is harder
Customer-facing AI has to get every output right. The model becomes more exposed to edge cases and must handle frustrated tones - and a bad answer reaches customers before anyone catches it.
Back-office AI doesn't have that problem. If it miscategorizes an invoice, the team sees it before it affects anyone and fixes the model in private.
In early 2024, Klarna announced its AI assistant was handling two-thirds of customer service chats, the equivalent of around 700 agents, and saving an estimated $40 million a year. By mid-2025, CEO Sebastian Siemiatkowski said the company had gone too far. Quality dropped and Klarna started rehiring humans for complex cases.
Whereas with a back-office misfire, you typically won’t see it showing up in customer reviews.
Pick your first project
Before you sign a contract with a vendor, look for one process where four things are true:
- It's high volume and repetitive: Think invoice processing, expense categorization, vendor onboarding, or support tickets that look the same week after week and happen at high volume.
- The data is already clean: If you'd need a six-month data project before AI could do anything useful, that's not your first project.
- Someone outside IT can own the result: A sales team lead for a contract close-time project, a head of customer ops for a project that automates routine tickets. If only the technical team is excited, the project will get stuck after the pilot.
- You can measure the outcome in something the board already tracks, like days to close or cost per invoice. If the metric needs explaining, the project won't survive contact with the next budget review.
Walk over to your finance or operations team and ask which task they would pay to never do again. That's your first AI project candidate. And if your candidate process fails any of the above questions, pick a different one.
Go deeper
👉 MIT NANDA: The GenAI Divide — State of AI in Business 2025 – the original report on why most AI pilots fail and where the wins are showing up.
👉 Stanford GSB: Human + AI in Accounting — Early Evidence from the Field – the 79-firm study behind the 7.5-day close discovery.
👉 McKinsey: How finance teams are putting AI to work today – what 102 CFOs say about scaling AI inside finance functions in 2025.
👉 The Hackett Group: 2025 Accounts Payable Digital World Class Matrix – AP automation benchmarks across 15 software providers.
Coming up tomorrow
Tomorrow we're breaking down six warning signs that product teams typically miss, together with the one action that gets things back on track.
That's it for today!
P.S. If you're already running an AI project in operations, hit reply with the function and one lesson you've taken from it so far.