AI Reality vs. Hype
You’ve seen this movie before. And so has your budget.
When RPA was the hot technology five years ago, the pitch was irresistible: deploy software robots, eliminate manual work, and watch headcount costs drop. Vendors, consultants, and business media all amplified the same story straight into the C-suite. The ROI projections looked great on paper. Most of them didn’t survive contact with reality.
I lived that experience firsthand — I started and led a process improvement group inside corporate finance, where I researched, piloted, and put real RPA deployments to work in the messy reality of finance operations. The AI pitch today sounds remarkably familiar. Different technology, same narrative arc. And if you’re an operations leader or process improvement practitioner being handed responsibility for AI adoption right now, I think you’re uniquely positioned to see through it — if you know what to look for.
There are really only two ways AI creates value
Before you can evaluate any AI investment, you need to separate two things that the market consistently conflates.
The first is new capability — AI doing things your team couldn’t do before, or doing them so much faster that it changes what’s possible. Think of an accountant running a quick data analysis that would have required a data scientist yesterday, or a controller asking a natural language question against a complex dataset without writing a single formula.
The second is automation — AI replacing work currently done by humans, reducing the labor cost of existing processes.
Both are real. But the hype machine treats them as interchangeable, which inflates expectations in ways that lead to real budget damage. You need to evaluate them separately, because the risks are completely different.
New capability is real — but don’t skip the controls
The new capability case is genuinely compelling in accounting and finance. AI can help your team do data clean-up, spend categorization, and light analysis that would have been out of reach before. That’s real value — and it’s available now.
But the failure mode is just as real. I recently used AI to consolidate several tax spreadsheets: pulling seven or eight data sets together, standardizing the formatting, and then taking a first pass at assigning spending categories to thousands of line items. The ETL piece — the consolidation and formatting — worked well. The categorization? It got me partway there, but required significant review and correction before I could trust it.
That experience is the norm, not the exception. AI performs impressively on average. It does not perform reliably on every individual output. The lawyers who got sanctioned for citing hallucinated cases in court didn’t do anything unusual — they just trusted the output without verifying it. In accounting, the consequences of that mistake are just as serious.
New capability has genuine value when you build verification into the workflow. Without that, it’s a liability dressed up as productivity.
The automation bet is where companies get burned
The harder conversation is about automation — specifically, the expectation that AI will let you reduce your existing workforce or stop backfilling roles. This is where the gap between projected and actual ROI lives, and where the parallels to RPA are most direct.
In the RPA programs I ran inside corporate finance, we found that most of the processes leadership wanted to automate weren’t ready. They had assumed the bots would reduce headcount quickly and skipped the prerequisite work. The data was inconsistent. The process steps varied depending on who was doing them. Exception handling was undocumented. The bots that actually worked — the ones that delivered real savings — were narrow, well-defined, and data-rich. They were a small fraction of total work.
AI is more capable than RPA. It can handle more variation and ambiguity. But “more capable” is not the same as “no prerequisites.” The same underlying problem that killed most RPA ROI projections is alive and well in AI: you cannot automate a process that isn’t ready to be automated. You just do it faster and at greater expense.
The unlock: process first, automation second
The organizations that got real results from RPA — including the one I built — shared a common pattern: they had invested in a process improvement function before they started automating. They knew which processes were standardized, routine, and backed by reliable data. They didn’t go looking for headcount savings; they went looking for process candidates. The headcount math followed naturally, on the handful of processes that qualified.
That prerequisite hasn’t changed for AI. The finance and accounting leaders who will get genuine returns from their AI investments over the next two to three years will be the ones who do this work first. Everyone else is going to spend the money and wonder what happened.
Before you commit budget, run the readiness test
Here’s a practical filter to apply to any process you’re considering for AI automation:
- Is the process standardized and documented, or does it vary by person or situation?
- Is the underlying data available, reliable, and consistent enough to feed to a model?
- Can you verify outputs before and after to catch errors before they cascade?
If you can answer yes to all three, you likely have a legitimate AI candidate. If not, fix the process first. Automating a broken or inconsistent process with AI doesn’t solve the problem — it accelerates it.
The good news for accounting and finance leaders is that you already know how to do this kind of process evaluation. It’s the same discipline that separates the teams that got results from RPA from the ones that wrote it off. The technology has changed. The prerequisite hasn’t.
Which of your processes would actually pass this readiness test? I’d genuinely like to know — reply with what you find, or what you’re wrestling with. That’s the kind of real-world data that makes this platform worth reading.