Six months ago, pulling a clean answer out of a messy trial balance was a skill you paid for. Now a language model does it in nine seconds, and mostly gets it right. If your finance function’s value was ever “we can produce the number,” the floor just fell out. The number was never the product. The question behind it was.
The methodology has said this for decades
Lean and Six Sigma have argued this point for years, in less fashionable language. DMAIC — the backbone of every Six Sigma project — front-loads two entire phases, Define and Measure, before anyone is allowed to fix anything. Define is nothing but disciplined question-asking: what is the actual problem, who is the customer, and what does “good” look like in numbers? Teams that skip Define rarely fail on execution. They fail because they execute flawlessly against the wrong question.
W. Edwards Deming, who built the modern quality movement, put it flatly: “If you do not know how to ask the right question, you discover nothing.” He also gave us the harder companion line — “Without data, you’re just another person with an opinion.” Put the two together and you get the 2026 bind: AI has made data and answers nearly free, so the scarce, defensible skill is knowing which question turns that data into a decision.
Here is the number that should worry you. A true Six Sigma process runs at 3.4 defects per million opportunities. Most finance close processes I diagnose run at 20,000–30,000 DPMO — roughly 3.5 sigma. AI will happily generate confident answers on top of a 3.5-sigma dataset all day long, at speed. Speed on the wrong question is just a faster way to be wrong.
A worked example every buyer recognizes
Take an EBITDA bridge in diligence. Ask a model “what was EBITDA last year?” and it will hand you a number off the P&L — instantly, correctly, and uselessly. The buyer does not care about that number. The buyer’s question is a different one: what is EBITDA after we normalize the owner’s above-market compensation, strip the one-time insurance recovery, re-cut the December revenue that was billed but not yet earned, and hold flat the customer that just churned?
That reframing is where the money lives. On a recent lower-middle-market engagement, the reported number and the defensible number differed by roughly $600K of EBITDA — on a business trading around 5x. That is ~$3M of enterprise value hanging entirely on which question got asked first. The AI could compute either answer. It could not tell you which one the buyer was going to underwrite. A person who has sat on the other side of a Quality of Earnings process could.
The same trap sits on the balance sheet. Ask a model for net working capital and it will average twelve month-end snapshots and return a peg. Ask what those month-ends are hiding — the receivable that was collected on the 2nd every quarter, the payables stretched right before close — and you get a different peg, and a different purchase-price adjustment. One of those numbers gets argued in a data room. The model does not know which.
The questions that do the real work in these scenarios are deliberately un-Googleable: What in this number won’t repeat? What got pulled forward to hit the number? What breaks if the largest customer walks? None has a lookup answer. Each one forces the expert to interrogate the data instead of accepting it — which is exactly what the Define phase demands before anyone touches a model.
What this means for diligence
So when you evaluate a CFO, a controller, or an outsourced finance partner in 2026, stop testing the thing AI is now good at. Everyone can produce the report. The differentiated, compounding skill is diagnostic: do they know which question exposes the risk before that risk shows up in a buyer’s model?
There is a simple tell. Hand a candidate or a firm a messy set of books and watch the first move. The weak ones start answering. The strong ones start asking — about cutoff, about concentration, about what changed and why. A junior analyst with a state-of-the-art model can now produce a polished answer to almost anything. What they cannot do is notice that the question in front of them is the wrong one. That noticing is judgment, and judgment is still scarce.
Good looks like a finance team that shows up to your quarterly review with better questions than you brought. Bad looks like a team that only ever answers the ones you already thought to ask. The first team compounds your understanding of the business every quarter; the second just renders it.
The commodity is the answer. The expertise is the question. That gap is widening, and it is the whole ballgame.
This is the discipline we apply to every Kaizen CFO engagement — questions first, numbers second, both defensible under diligence. Talk to sales →