In brief:
-
MIT research found that 95 percent of enterprise generative AI pilots deliver no measurable P&L impact. The core issue is not model quality. It is a learning gap: AI that cannot see how the specific business works.
-
The stall is rarely a capability problem. Models are strong. What they lack is context: the institutional logic that lives in spreadsheets, and the verification layer that makes AI output trustworthy at scale.
-
A general-purpose AI can summarize a single workbook plausibly. It cannot tell you that this file, among four thousand near-identical siblings, is the only one that changes a tax valuation formula. That fact requires estate-level context.
-
AI-ready is not a tooling status. It is a property of the estate: legible logic, governed change, and verification built into the workflow.
The pressure on finance to show AI progress is no longer subtle. Boards want a story, executives want outcomes, and every vendor demo promises copilots and agents that will change how the function operates. Yet inside most finance organizations, momentum looks like pockets: a productivity gain here, an impressive prototype there, and a growing pile of pilots that never became operations.
The instinct is to blame the technology, or to wait for the next model release to fix it. Both miss the actual constraint.
The pilot plateau is real
According to MIT's NANDA initiative, whose GenAI Divide: State of AI in Business 2025 report drew on 150 executive interviews, a 350-employee survey, and 300 public AI deployments, 95% of enterprise generative AI pilots stall without measurable P&L impact.
Read carefully, that statistic is not an indictment of the technology. MIT's own diagnosis points away from model quality and toward what it calls a learning gap: tools that cannot learn from or adapt to how the enterprise actually works. The models are good enough that people keep starting pilots. Something downstream of the model keeps stopping them.
In a recent webinar with Coherent, Purav Desai, Head of AI & Product Life at Resolution Life, a closed-block life insurer running AI at scale across contracts and actuarial models, described the gap precisely: the technology handles content and code creation at volume, and organizations then hit the harder question of whether the output can be trusted enough to run the business on. Getting from a working demo to an authentic business decision is the chasm, and it is where most pilots stall.
Watch the full conversation: AI Trust at Scale, with Peter Roschke (Coherent) and Purav Desai and David Simmons (Resolution Life).
That chasm has two walls. One is context. The other is verification.
The context wall: AI knows the industry, not your business
Coherent CTO Peter Roschke describes today's AI as a well-trained temp worker: deeply versed in the industry, shows up Monday morning, and knows nothing about how your specific business runs.
Ask it about standard actuarial practice and it performs brilliantly, because it has read more than any human ever will. Ask what it knows about how underwriting is done inside your company, for your specific block of business, and the honest answer is close to zero. Worse than zero, in practice, because as Desai put it, "you're always going to get an answer." The model will respond confidently whether or not it has the context to be correct.
The same webinar offered a demonstration of what that means for anyone tempted to solve spreadsheet analysis by uploading files to a chatbot.
Take a medium-complexity financial workbook: ten tabs, seventy thousand formulas, a hundred thousand cells. A general-purpose model will produce a reasonable description of what it does, the equivalent of a week of human write-up, and if you are lucky it will be right.
But the question that actually matters is different.
Among the four thousand other files in the estate that look almost identical to this one, this is the only one that changes the tax valuation logic on sheet three.
That is the material fact about the file. No amount of model capability surfaces it from a single upload, because the fact does not live in the file. It lives in the file's relationship to the estate around it.
Roschke's shorthand for this problem has become something of a refrain:
Excel is the largest unmanaged codebase in any company.
Twenty years of decisions and institutional logic, scattered across drives and SharePoint in drips and drabs, encoding how the business actually works. That is precisely the context an AI initiative needs and precisely what no foundation model ships with. Making that layer legible is its own discipline, covered in depth in what estate-level analysis actually involves.
One CIO put the resulting position bluntly in a conversation recounted on the webinar: asked where the organization was on its AI journey, the answer was that they were still trying to figure out what was in their documents at all. So, not step one. Step minus one.
The verification wall: everyone can generate, almost no one can validate
The second wall gets less attention because it arrives after the pilot looks successful. Generative tools have handed everyone the ability to produce: reports, analyses, reconciliations, even new spreadsheets and code. What organizations have not built is the matching ability to validate, and ungoverned generation at scale is how a finance function trades one control problem for a bigger one. A figure that is confidently wrong moves faster through an AI-accelerated process than it ever did through a manual one.
This is where the finance instinct for control becomes an advantage rather than a drag. The organizations getting past the plateau are the ones treating trust as a workflow property, engineered in, rather than a hope applied at the end. AI does the first pass, evidence and reasoning are captured as the work happens, and a human approves before anything reaches a decision.
Resolution Life offered a concrete version of what that looks like, described by Desai on the webinar:
-
Interrogate at a scale humans can't. The team used AI to extract information from contract documents across roughly 1,200 products, at a depth manual review could never reach.
-
Refuse to take the output on faith. Rather than treating plausible-looking results as finished, they loaded the AI outputs into Coherent Spark, where the actuarial logic already runs as governed calculations.
-
Prove it against known results. The outputs were regression-tested against sixteen months of actuarial calculations, turning "this looks right" into "this reconciles."
The generative layer produced the answers; the deterministic layer proved which ones held. That pairing, generation checked by governed calculation, is the difference between an experiment and an operating capability.
What AI-ready actually means for finance teams
Underneath the tooling decisions, AI readiness in finance reduces to two questions:
-
Can the AI see how your business actually works? That means the spreadsheet layer, where the operational logic lives, is mapped: which models matter, which files are variants of which, where the logic diverged, and who owns what. Context is not a nice-to-have for accuracy. It is the difference between an answer about your industry and an answer about your company.
-
Can you prove the output is right? That means verification is part of the workflow, not a post-hoc audit: evidence attached to conclusions, changes tested against known results, human approval where materiality demands it.
Neither question is answered by procuring a better model. Both are answered at the estate layer, which is why the organizations making real AI progress in finance started somewhere that looks unglamorous: understanding what they have.
That is the specific ground Coherent Insights covers, making the spreadsheet estate legible enough that AI can be applied deliberately, with evidence, and at the points where it will actually hold up. And it is the same foundation that determines whether a broader modernization effort accelerates or joins the long list of transformations that stalled on the layer nobody mapped.
Finance does not need more automation ideas. It needs a clear line of sight, and a way to prove what it builds on that sight is right.
