Ninety-two percent of technology executives say their organization tracks the financial impact of AI-generated work. Two percent say more than half of that work is actually recorded as a business outcome.
That gap, from Lanai’s 2026 AI Labor Report released on 9 June, is not a rounding error. It is an accounting problem sitting inside every large company that has scaled AI faster than it has built the systems to account for it. Lexi Reese, co-founder and CEO of Lanai and former COO of Gusto, calls the pattern “AI labor orphaning”: AI systems perform work, but that work never formally enters financial systems, performance records, or systems of record. The output is real. The ledger is incomplete.
“Those aren’t just quirky statistics; they raise basic questions about the numbers executives rely on,” Reese said in the report’s release. “If AI is doing a meaningful slice of the work but never shows up in the ledgers, how confident can you be in your P&L, your headcount plan, or the org chart you use to run the organization?”
That is a boardroom question, not an IT question. And it arrives at an awkward moment.
Adoption Outran Accountability
Enterprises did not adopt AI slowly. They adopted it everywhere, often without a single owner responsible for proving what it produced.
Lanai’s survey of 200 U.S. technology executives at organizations with 1,000 or more employees found that 90% lack a single, dedicated function responsible for tracking how AI delivers return on investment. Accountability is scattered across finance, IT, operations, and business units. No one clearly owns the answer when the CFO asks whether AI spending is creating measurable value or simply adding another layer of cost.
The numbers get worse under pressure. 79% worry their AI budgets will be cut because spending cannot be tied clearly to revenue or profit. Reese calls this faith-based budgeting: spend that survives on belief rather than evidence. Ninety-six percent have already lost at least one ROI opportunity because they lacked visibility into how AI made decisions. This is not a future risk. It is happening in budget cycles right now.
Independent research points the same direction, if less sharply. McKinsey’s 2025 State of AI survey found nearly 88% of organizations now use AI regularly, yet only 39% attribute any EBIT impact to it, and just 6% qualify as high performers capturing significant enterprise value. BCG’s 2025 research found just 5% of companies achieving AI value at scale, its “future-built” leaders, with 60% reporting little to no material impact. The technology works. The institutional architecture for capturing value does not.
The Invisible Workforce
The standard shadow-AI story treats unauthorized tools as a security problem: employees pasting proprietary data into consumer models, compliance breaches to be hunted and destroyed. That framing misses the signal.
Lanai’s report found that 53% of executives estimate most automated work runs through unmonitored shadow applications. In many firms, the AI that finance and IT approved is not the AI doing most of the work. Employees route around bureaucracy because the tools deliver the productivity the board demands. By attempting to ban the behavior rather than capture its telemetry, enterprises blind themselves to their highest-performing workflows.
“Right now, the story inside big companies is not that the robots took over,” Reese said. “It’s that AI quietly took on pieces of the job, writing the first draft, sorting the queue, flagging the anomalies, while people stayed responsible for the final call. The accounting and governance systems just haven’t caught up to that split.”
The performance management system makes the distortion visible. 87% of organizations credit AI-assisted output entirely to the human employee, sometimes or always. Performance reviews, promotion decisions, and bonus pools are built on work where the machine contribution is invisible, not because anyone is hiding it, but because no system was built to record it any other way. 88% have no formal methodology for attributing business outcomes to AI. Forty-three percent assume that if AI was involved, it contributed. Thirty-eight percent rely on educated guesses. Only 12% have a defensible answer when finance asks.
That is not shadow IT. It is shadow labor: real work, no owner, no line on the books.
Accountability Without Power
The structural problem is not visibility alone. It is authority.
Boards hold technology leaders accountable for AI return on investment while denying them the levers to deliver it: the power to set standards across business units, reallocate spend from failed pilots, and redesign workflows where AI actually changes how work gets done. Reese has described this as accountability without power. The CIO is asked to prove value from a budget they do not fully control, over tools they cannot see, inside processes they cannot mandate.
Reese is careful about where that accountability actually sits. In her experience this year, the CIO supports AI transformation but is often not the directly responsible individual the CEO holds to account for revenue growth or operating efficiency from AI spend. The executives who feel the squeeze most are the COO, optimizing for revenue per head while headcount stays flat, and the CTO, when engineering owns both the customer-facing AI and the internal productivity layer. The CIO, in this telling, carries the bill without authority over the workflows generating it, which is why the leaders of go-to-market and R&D functions matter as much as any single title in the C-suite.
Lanai’s answer points toward a role evolution: the Chief Information Officer becomes a Chief Intelligence Officer, responsible not for keeping servers online but for governing intelligence flows, human and machine, across the enterprise. The company launched its AI @ Work Operating System to provide portfolio-level visibility across sanctioned tools, shadow AI, and autonomous agents. The product pitch is observability. The organizational claim is decision rights.
The question worth asking is whether a new title solves an accountability problem or relocates it. A Chief Intelligence Officer with telemetry but without workflow authority is governance theater: policy without the power to change how the company operates. A Chief Intelligence Officer with standards authority, spend reallocation, and cross-functional redesign mandate is a different job entirely. Most enterprises have not decided which one they want.
“These are not just measurement gaps; they’re cracks in how companies describe themselves on paper,” Reese said. “If AI’s share of the work never shows up in the accounts, then the P&L, unit costs, and even the organizational chart start to drift away from how the business actually operates. You can’t plan hiring, investment, or restructuring on numbers that leave out an entire category of labor.”
The Other Half of the Agent Story
The accountability gap has a mirror image in the agents enterprises deploy on purpose.
Autonomous procurement agents at companies like Walmart and Maersk negotiate supplier contracts inside defined guardrails, update enterprise systems, and leave audit trails. That is AI labor with an owner, a budget line, and a governance layer. Shadow AI is the inverse: unauthorized tools doing authorized-looking work with no record, no attribution, and no one accountable when the CFO asks what the spend produced.
Together, the two problems form one boardroom argument. Enterprises are scaling machine labor on two tracks simultaneously. On one track, agents they can see and govern. On the other, AI they cannot measure and cannot defend in a budget review. The roughly one in ten who can already give finance a defensible answer are the ones treating AI execution cost as a labor line rather than a generic IT expense, building attribution methodology before they are forced to, and recording AI contributions in the systems executives actually use.
Everyone else is running a company on numbers that omit an entire category of worker.
The report’s survey also found that 100% of organizations still require human review after AI generates work. None reported fully autonomous workflows. The dominant model is supervised machine labor: AI drafts, classifies, and flags; humans check, edit, and approve. That is the reality behind the autonomy hype. And it makes the accounting failure more urgent, not less. If humans remain responsible for the final call, but machines do the first draft, the performance system must record both contributions or it lies.
Reese offers a concrete illustration from a customer using Lanai’s Token Tuner. The company measured two groups inside the same finance team running identical workflows for 30 days, same work and same output quality. The only variable was the default model each group happened to use, a choice no one had made deliberately and no one had seen until it was measured. One group’s bill came to $52,015. The other’s came to $13,007, for the same work product. Fixing that single default saved roughly 5% of the team’s annual token spend, straight to the bottom line. “The right question isn’t how many tokens each team consumed,” Reese said. “It’s what outcomes they produced and what those outcomes cost.” The discipline she draws from it is to translate token spend into labor economics: the hours of work AI produces, the cost of that work per hour, and its value in language operators can actually manage against.
What Boards Should Ask
Three questions cut through the vendor framing.
First: who owns AI ROI, and what decision rights do they actually have? Not the title. The authority to kill pilots, standardize tools, and change workflows.
Second: what percentage of AI-assisted work can finance attribute to a business outcome today, not in theory? Lanai’s 2% figure is a vendor survey result. The direction of travel is what matters. If the answer is close to zero inside your own company, the budget conversation is already broken.
Third: is shadow AI a compliance problem or a productivity signal? If employees are paying for tools out of pocket because they work, the enterprise has a product-market fit problem in its own IT stack, not a discipline problem in its workforce.
Reese’s report lands as AI budget scrutiny intensifies across large enterprises. The companies that treat AI as unrecorded labor will discover the gap in the next planning cycle. The companies that build attribution first will have something rarer in this market: a defensible answer when the board asks what all that spending actually produced.
Leave a comment