TL;DR. Vendor AI productivity numbers are denominated in inputs (hours saved). Financial statements are denominated in outputs (revenue, EBIT, free cash flow). The gap is unexercised optionality, not measurement error. Run every vendor claim through four questions: where does the hour leave the system, what is the conversion rate, what is the realisation timeline, what is the dependency. The number that survives will be smaller than the one that came in. It will also be defensible.
Every vendor deck closes with a number
Every AI vendor deck closes with a number. Twenty per cent fewer admin hours. Forty per cent faster reports. Three weeks of capacity returned per FTE per quarter.
The CFO writes it down. Then nothing happens to the P&L.
That is the pattern.
The accepted wisdom
The current orthodoxy in AI buying is that productivity gains are gains. Microsoft's 2025 Work Trend Index reports 82% of leaders treat AI as necessary for the next twelve months. McKinsey's State of AI 2025 finds 78% of organisations now use AI in at least one function. Multiply hours saved by salary, multiply by FTE count, and you have a number large enough to anchor an investment committee.
Reasonable people accept this. The savings are real. The hours are measurable. The vendor used a third-party benchmark. The number lands in the deck and lives there.
The crack
The crack shows up six to twelve months after deployment.
The hours are saved. The headcount has not moved. The revenue per FTE has not moved. The cost-to-income ratio has not moved either. Operating leverage, the thing the CFO actually cares about, looks identical to the prior year.
A McKinsey 2025 follow-up survey found fewer than one in five firms running GenAI at scale could attribute material EBIT impact to it. EY's 2025 Wealth and Asset Management survey found 95% of firms had scaled GenAI to multiple use cases, and only a quarter reported substantial business impact. The hours are leaving compliance, advice and operations on time. The dollars are not arriving downstream.
Something is not adding up.
Pull the thread
The thread runs cleanly into corporate finance.
Productivity gains are denominated in inputs. Hours, tasks, drafts, calls handled. Financial outcomes are denominated in outputs. Revenue, gross margin, EBIT, free cash flow. The vendor benchmark measures the input side. The CFO measures the output side. The gap between the two is the question nobody asks out loud.
Three things have to happen for an input gain to become an output gain.
First, the saved hour has to be removable. Either it leaves payroll (lower headcount), or it is reallocated to a higher-value activity (more revenue per FTE), or it is removed from a queue that constrained throughput (capacity release). One of those three. Not "freed up." Not "redirected to strategic work." Removed.
Second, the saved hour has to be saved consistently. A scattered 20 minutes a day across six tasks rarely consolidates into a removable hour. Aidan Toner-Rodgers' 2024 MIT study of AI-assisted R&D scientists at a large materials firm showed 12% to 15% productivity gains in patent filings. The gains landed unevenly: top performers gained, bottom performers gained little. An average of 11 saved hours says nothing about whether those hours arrive in the same place every week.
Third, the saved hour has to convert at a rate the firm can actually realise. A senior adviser at $250 per hour has different conversion economics than a paraplanner at $80 per hour. A vendor's "savings per FTE" calculation typically uses a blended rate that does not match how the firm actually staffs work.
When all three conditions hold, the productivity gain converts. When any one breaks, the gain stays trapped on the input side of the ledger.
The inversion
A productivity gain is not a gain. It is optionality. The firm has acquired the right, not the obligation, to remove cost or add revenue. Until the firm exercises that right, the gain has zero value on the financial statements.
Vendor productivity numbers are option grants disguised as savings. They tell the buyer what could happen. They do not tell the buyer what will happen, because what will happen depends entirely on what the firm does next.
That is why CFOs who run AI investment cases on vendor productivity numbers consistently miss target. The number was never a forecast. It was a ceiling on a forecast that nobody built.
The four-question CFO read
A better mental model for AI productivity claims, designed for CFOs and principals who own the investment case.
Four questions. Apply them to every vendor number before it goes into the model.
1. Where does the hour leave the system. Pick one. Headcount reduction, capacity reallocation, throughput release, or "we are accepting the hour stays in the system as cushion." That last one is a legitimate answer. Most firms accept it implicitly without writing it down. Writing it down is the work.
2. What is the conversion rate. For headcount: the actual fully loaded cost of the role, including superannuation, technology, real estate and indirect overhead. For reallocation: the marginal revenue per hour of the activity the time is being moved into, and whether demand exists for that activity. For throughput: the constraint downstream of the saved hour, and whether it absorbs the additional volume.
3. What is the realisation timeline. A 20% efficiency gain over six months, on a base of 50 FTE, in a function with a 12-month notice cycle for restructure, realises differently from a 5% gain in a contractor-heavy operation. Investment committees want NPV. NPV requires timing. Vendor decks rarely include either.
4. What is the dependency. What other change has to happen for the gain to land. CRM cleanup. Process redesign. Compliance approval. Vendor integration. New role definitions. The dependency list is the gap between the vendor case and the CFO case. Most firms run a six-month dependency list against a twelve-month investment thesis. The arithmetic does not survive contact with the calendar.
This is the free consulting bit. Run any vendor productivity claim through these four questions before the next investment committee. The number that survives will be smaller than the one that came in. It will also be defensible.
What this changes
A firm that adopts this read shifts its AI conversation from procurement to portfolio management. AI tools become option contracts. Some are exercised, some expire. The CFO holds the strike price (the dependency list and the conversion rate) and decides which options are worth the investment to convert.
That shift has implications.
The vendor relationship changes. The buyer stops asking "how much time will this save" and starts asking "what dependencies do we own to convert the saving." The vendor that can answer the second question wins the deal. The vendor that only answers the first question loses the second-year renewal.
The board pack changes. AI investment cases stop being presented as "expected savings" and start being presented as "options held, options exercised, conversion rate, residual value." The board can compare AI investments to the rest of the capital stack on a like-for-like basis.
The operating model changes. Productivity gains that go unexercised become a recurring agenda item. Either the firm converts them or it stops counting them. There is no third state in a healthy financial system.
The uncomfortable implication
Most "productivity gains" reported by Australian financial services firms in 2025 and 2026 are not gains. They are unexercised options. The firm has the time. The firm has not chosen to monetise it. That is a strategic choice, not a measurement problem, but the financial statements treat it as if no choice has been made.
Two CFO-shaped consequences follow.
If the productivity gains do not show up in the next two reporting cycles, the investment committee that approved them will be asked to explain why. The vendor benchmarks will not survive that meeting. The dependency list will.
If the gains do show up, the firms that exercised the options early will hold the operating leverage. The ones still talking about "freeing up time for higher-value work" will run the same cost base on a smaller revenue line, because the higher-value work was never operationalised.
Read it like a CFO
Every AI vendor deck closes with a number. The number is real. The hours are real. The dollars on the slide are arithmetic, not invention.
The dollars in the financial statements are different arithmetic. They require the firm to do something with the time the AI returned. The vendor's number is the ceiling. The firm's choices are the floor. The CFO's job is to close the gap.
The number does not lie because it is wrong. It lies because it is incomplete.
Read it like a CFO. Then decide whether to exercise the option.