I’ve been watching a pattern play out across every major technology wave I can remember. Companies deploy the new thing. The demos look great. The headline numbers are impressive. And then… nothing changes on the income statement.
That’s the “last mile” problem. And according to a new Harvard Business Review article from researchers at Harvard Business School and Microsoft, AI is no different.
The study is based on a closed-door summit at Harvard where senior leaders from a dozen global organizations — healthcare, banking, manufacturing — gathered to compare notes on why their AI transformations have stalled. These weren’t laggards. One global investment bank had over 250 active AI applications. A global apparel firm had automated more than 18,000 finance processes. One payments network had 99% of employees actively using AI copilots.
And yet none of them could point to where the gains showed up on their financial statements.
We’ve seen this before
In the 1990s and 2000s, companies spent millions on SAP and Oracle. Employees kept running parallel spreadsheets because the workflows didn’t actually change. The technology got deployed. The operating model didn’t.
In the 2010s, everyone migrated to the cloud. Infrastructure improved. Value creation didn’t.
AI is the latest chapter — but with one twist that makes it different from those prior waves. AI is a diagnostic tool. It surfaces broken processes faster than organizations can fix them. At one healthcare insurer in the study, AI exposed fragmented workflows faster than they could be resolved. A global services firm found the same process being executed dozens of different ways across 170 countries. The technology held up a mirror. Organizations didn’t like what they saw.
Every major technology wave promises efficiency. But the accounting impact — the part that actually shows up on financial statements — only materializes when the operating model changes to capture it. AI is no different.
7 frictions stalling the last mile
The research identifies seven specific structural problems — not technology problems — that prevent pilots from scaling.
- Pilot proliferation. You can build a proof of concept in days now. Getting it to enterprise-grade, with the right security and tech stack? Months. Most pilots never make the jump.
- The productivity gap. Saved time gets reabsorbed. If you can draft 15 things instead of 10, you end up in 5 more meetings. The efficiency never escapes the workflow.
- Process debt. Decades of acquisitions and workarounds mean AI is automating chaos. If your data and workflow foundation is shaky, you don’t even understand the current process well enough to improve it.
- Tribal knowledge. The expertise that lives in people’s heads is valuable — and protected, because it confers status. Getting that knowledge into AI systems requires an identity shift, not just a reskilling program.
- Governance breakdown. Human-in-the-loop controls work for isolated cases. They collapse under multi-agent architectures. One bank in the study is already operating 100+ agents and planning for tens of thousands. Where do humans step in?
- Architectural complexity. Getting SAP, Microsoft, and Google environments to talk to each other reliably is genuinely hard. One company spent months on it.
- The efficiency trap. When you frame AI as a cost-cutting tool — the way offshoring was framed — you trigger defensive behavior from middle management and narrow what the C-suite thinks is possible. The biggest gains come from rethinking value creation entirely, not shaving minutes off existing tasks.
Two problems accountants need to own
There are two places where this lands directly on the finance and accounting function.
First, productivity gains that don’t hit the P&L. This isn’t just a management problem — it’s a measurement problem. Finance leaders at these organizations can’t find the AI ROI in their numbers. That points to a gap in KPIs, budget models, and possibly cost center classifications. If your job is to measure business performance, and AI is producing value that your current framework can’t see, that’s your problem to solve.
Second, agentic AI as an internal controls issue. Earlier AI tools recommended. Agentic AI acts — it updates ledgers, routes approvals, coordinates across ERP systems. Traditional control frameworks weren’t built for this. When a multi-agent system makes a financial error, who owns it? Is the audit trail complete? These are questions auditors and controllers need to be asking now, not after something goes wrong.
Key takeaways
- Most large enterprises are pilot-rich but transformation-poor — AI tools are widely deployed but enterprise-wide value hasn’t materialized
- The bottleneck is organizational design, not AI capability — the seven frictions are all structural, not technical
- AI productivity gains don’t flow to the balance sheet automatically — they require deliberate role redesign and budget reallocation
- Agentic AI creates new internal control challenges that existing governance frameworks aren’t designed to handle
- Finance and accounting professionals need updated measurement frameworks to capture AI ROI — and new skills in process design and governance to stay relevant
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply