FinOps went from invisible to indispensable in two years. The model transfers to people enablement teams.

Two years ago, almost no one outside finance could tell you what a FinOps team did. Cloud cost management. Spreadsheets about cloud server bills. (Riveting, I know.) Then AI spend landed on the cloud bill, the team that already owned the spreadsheet on calculating overages got handed the AI version of it too, and is now a mission-critical business stakeholder. They are the team answering the big question: is the juice worth the squeeze?

I have never heard of this team, what are you on about, FinWhat?

I know, it took me some digging too, but I wanted to find out which part of the enterprise typically owns the BIG impact number, or can at least help us figure it out. So after a month of online sleuthing and a lot of conversations (a month I am not getting back, and a browser history that now thinks I have pivoted industries), I landed on FinOps as a rising (darling) category in the org chart.

A quick definition, since I needed one too: FinOps, short for financial operations, is the practice (and the cross-functional team) that brings financial accountability to cloud and infrastructure spend. When the spend is per-seat licenses for tools like Microsoft Copilot or ChatGPT Enterprise, the same job often sits with a different team, usually IT Asset Management or SaaS operations. Different team, different tools. I am going to keep saying FinOps because it is the cleanest example of the thing I am pointing at, but when you go do this, know which door to knock on.

And the rise is on the record. Flexera's annual cloud survey put FinOps teams at 51% of organizations in 2024 and 63% now. The FinOps Foundation's practitioner survey says 98% of those teams now manage AI spend, up from 31% two years ago. The Foundation even rewrote its mission this year, from the value of cloud to the value of technology, so it is expanding its mandate.

Now the reckoning itself

The data is consistent on one thing: It isn't the tech that is breaking. RAND notes that more than 80% of AI projects fail, roughly twice the rate of IT projects that do not involve AI, and pins the blame on leadership and organizational factors, not the technology. S&P Global found the share of companies abandoning the majority of their AI initiatives before they reach production surged to 42% in 2025 (2026 numbers still outstanding), up from 17% a year earlier, even as that same survey shows adoption outrunning the targets companies set the year before. And Gartner predicted at least 30% of generative AI projects would be killed after the proof of concept by the end of 2025, then said it actually came in over 50%.

A fair reader will push back here, so let me beat them to it: a high kill rate after a proof of concept is not proof the technology is broken. Killing weak pilots fast is what a disciplined portfolio is supposed to do, and the survivors post real gains. The honest read is not "AI does not work." It is that most enterprises cannot yet tell which of their AI bets are working. This is looking to me like a measurement problem.

Ok great Josh, thanks for this, should I point my sales team for my new AI Spork learning fidget spinner at the FinOps team? Ehmm. Sure. But I keep coming back to this, because the majority of enablement teams I have spent time with spent the same two years answering the same AI mandate with needs analyses, course libraries, and completion dashboards. (I have had the dubious honor of driving some of those builds. I am especially proud of my digital happy-sheet dashboarding skillz.) Nobody was being lazy. That playbook is what the business asked of the field for two decades, or at least our interpretation of what they asked, arrived at through a very advanced triangulation method of listening to whoever was yelling loudest, plus satisfaction surveys and NPS scores. Winning.

Ok sure, another L&D-is-wrong hit job, $hit post.

Well, kinda. More of a call to action, with some ideas on the how. FinOps is a new-ish function (more of a mashup, really). They do not have 9,000 playbooks or four decades of baggage as a compliance-reporting function. They had one gnarly asset, spend data. Nobody else wanted to own it (more likely they were volun-told), and they built a whole function on it. L&D has some really great data too, if we hunt it down and define what it actually is, and it connects straight to the business the moment we look at it through the right lens.

Ok and….

Mercer's 2026 Global Talent Trends found only 8% of executives see HR as strategically embedded in the business. But the deeper problem is not who "owns" training. It is that companies buy enterprise-wide licenses for tools employees do not know how to use, and then do not use. The shelfware piles up. LayerX, a browser-security vendor, looked at its own enterprise telemetry and found only about 18% of users touch an AI tool in a given week; the median employee has logged about a dozen AI conversations total, while the top sliver of power users runs roughly twelve times that volume at nine times the prompt depth. Read it as directional, not gospel, it is a company that sells AI-usage monitoring, but the shape matches what everyone is seeing. IT can track the login, FinOps can track the seat cost, and neither can manufacture the daily habit that turns an idle license into a capability.

These numbers are catnip. You could read them as the obituary for L&D. I read them as the opening. In nearly every large company there is a capability-data conversation happening right now: what the AI rollout is doing to performance, where skills actually sit versus where the org chart says they sit, what time-to-competence looks like for the roles that matter. IT owns the tools, finance owns the spend, HR owns the headcount, and the question in the middle, are people actually executing better, mostly goes unanswered.

Every budget in the org is getting challenged right now. Mission-critical roles, functions, and tools have to be defended. Gartner's CFO survey puts enterprise-wide cost optimization near the top of finance's 2026 priorities, and a separate Gartner survey shows CFOs planning overhead to grow slower than revenue. If you run an L&D budget, you have already sat in the meeting that data describes. Capex or Opex, the question is the same: what did we get for it, and how much can be cut and reallocated to AI?

Most functions answer that question badly. FinOps answers it for a living: unit costs, utilization, forecast accuracy, value per dollar. (Read that list back and count how many L&D data points have ever headlined a board deck. I'll wait…) In my opinion this is the beginning of our opportunity.

The part L&D treats as a chore

L&D has historically treated data as the vegetable course (usually very boiled asparagus), something to choke down before the real work of building spectacular courses. I get the instinct, data is the least interactive, least engaging corner of the learning world for a lot of people. And most of what the field gets asked to measure, completions and satisfaction scores, genuinely is busywork (our fault, we trained executives on those metrics, then watched them stop meaning anything the second they became targets, which is just Goodhart's Law wearing a lanyard). Twenty+ years of reporting the same data set has taught everyone that "data" means dashboards that are quickly covered but not aligned to business impact. That is now violently changing.

FinOps built its entire standing on server billing data, which might be the dullest dataset in any company. Learning data is so much richer: behavior, performance, capability, mobility. Frederick Taylor timed factory workers with a stopwatch a century ago and built a management revolution on it; we have behavioral data he could never have dreamed of and mostly point it at completion percentages. What we could track and feed back to the business is the coolest opportunity we have in this space, maybe in the whole company.

Nobody is coming to hand L&D this. The work is unglamorous and politically difficult to gather, which is most of why it is still unclaimed. It is hard. It means talking to a lot of different departments (IT, Finance, HR, and whoever quietly built the data lake nobody will admit to owning). FinOps is not easy either. It has also become mission critical.

What the data work actually looks like

To pull this off, L&D has to grow up about data. The mature version maps a skills architecture straight to productivity, not to isolated learning environments.There are true privacy risks, try to link individual learning records to CRM and performance data at the employee-ID level and Legal and HR Compliance will shut you down, fast especially in countries that have strong individual data privacy laws.

So let's get creative. We could start with anonymized cohort tracking. We do not track John Doe; we track the time-to-competence of the "Q1 Mid-Market Sales Cohort" that went through an AI upskilling sprint, against a control group. The business gets actionable data, and we get around the GDPR and privacy roadblocks the honest way, by never touching an individual record, instead of spending 6 months trying to get them only to be told no.

In our terms it is Data-Driven Learning Design run as a live loop, not a report: discover what the merged picture says, monitor what people do after something launches (especially where they quit), and respond by fixing it while it is live instead of filing it for next year's redesign. Formative evaluation with a responsive SLA.

Time-to-competence for a revenue-critical role is where I love to start, because it has visibility and a lot of trackable moments, and because the business results so far are all over the map.

In one Fortune 500 deployment, generative AI lifted customer-support output 15% on average, but 30% for the least-experienced workers and almost nothing for the veterans, who actually dipped a little in quality. It is a leveling, not a rising tide. Which means the cohort finance most wants to cut is often the one a single good intervention would convert.

You cannot blindly map a call-center study onto analysts or developers, though. Here is the one that should keep an L&D leader up at night. When Harvard and BCG ran a field experiment with 758 elite consultants, AI made them dramatically better on tasks inside its range and quietly worse on one that sat just outside it. Then the researchers sat one group through a prompt-engineering crash course first, on the theory that a little coaching would help. It did the opposite. On the out-of-frontier task the control group was right 84.5% of the time, the AI-only group 70%, and the coached group worst of all at 60%. The line between where AI helps and where it quietly wrecks the work is really hard to see, even to experts at one of the most capable firms on earth. You cannot train people to feel their way across a line they don't know about. We can start to measure which side they landed on, by role and by cohort.

Even the scariest headline, MIT's GenAI Divide finding that 95% of enterprise AI pilots show no measurable impact on the income statement, says something different than a lot of the doomscrollers like to post. The same report found workers at roughly 90% of companies already reach for personal AI tools to do real work. The pilots appear to be failing on procurement, integration, and workflow, the operating-model layer. (Caveat emptor: it is preliminary, not peer-reviewed, and the group behind it sells the agentic plumbing it says would fix the problem, so weight it accordingly.) Nobody built the analytics dashboard from people-are-using-it to people-got-measurably-better, the only version the income statement notices.

And across 51 enterprise deployments that eventually worked, 61% had hit a failed AI project first, and the hard costs were the invisible ones, change management and data quality, not the licenses. So the job is to check whether people actually got better at the work, instead of counting logins and calling it a win.

This is a time-to-competence curve we can track, by role and by cohort. Put it in FinOps language. They can already tell you who is burning tokens and who is sitting on idle licenses, that is their data, not ours. What they cannot tell you is what the training actually enabled people to do, who your high and low performers are now, how fast each cohort reached the bar, and whether the program moved the performance numbers the business set, and why or why not.

The test for every data question is simple: does it make money, save money, or reduce risk (the main benchmarks we use to measure L&D initiatives at 8Levers)? If it does not trace to one of those three, do not start there. But L&D cannot just abandon its compliance mandate either, drop that ball and the company gets fined. So you rebuild the intake, Lever 1 in our framework, into tiers: Tier 1 is automated, self-service compliance (low touch); Tier 2 is standard role onboarding; Tier 3 is the strategic growth work, where the agile budget and the FinOps partnership live.

The partnership sitting one floor away

Everything so far has been about copying the FinOps playbook and hoping it works. The faster route is to walk down one floor (virtually or physically), ask them what jobs they need done, and figure out how a team that actually knows how people learn could help them hit their numbers. This is stakeholder work, and it means dropping the L&D evangelism (nobody in IT cares about our "learning culture") and learning to talk in the rigorous language of the business. Please, no L&D jargon. Study their world a bit, use their terms and the problems they actually complain about in their own forums (In L&D terms this would be a needs analysis).

FinOps can price AI usage down to the token. ITAM knows exactly which licenses sit idle, and the usage numbers above say a lot of them do. But ITAM's mandate when it sees that is simple: de-provision the idle seats and pocket the savings. It does not care why people are not using them. The trouble is, yanking licenses starts a fight with business-unit leaders who say IT is killing their productivity.

This is where L&D earns a seat. But the fix cannot run on goodwill and manual CSV exports. You wire the HRIS and the learning platforms into the IT asset-management system, so an idle license automatically triggers an intervention. (Yes, that is ecosystem plumbing, Lever 7, the boring part nobody claps for.) When ITAM flags a seat as inactive for 30 days, instead of auto-deleting it, the user drops into a short, automated upskilling sprint. A license-offboarding buffer.

Do not assign them a 45-minute SCORM module on "The History of Large Language Models." The sprint has to be short, tactical, and in the flow of work, show them how to save 30 minutes today or do not bother. If they do not finish it and show some usage inside 14 days, the seat gets reclaimed automatically. FinOps finds the waste, L&D runs the remediation, IT keeps the political headache off its desk, and the company finally drives up adoption on the seats it already paid for.

Ok, how do we do this? What do we bring to the table?

We put performance data next to spend data, and we stop talking about "learning outcomes" and start talking in their units: cost-per-active-user versus cost-per-allocated-seat. Those underused license cohorts, the ones finance is already side-eyeing, become our capability targets. And the research says that is exactly where the return hides: AI lifts the least-experienced users the most, so the cohort finance most wants to cut is often the one a single good intervention would convert.

Their dashboards become our scorecard, which means we get measured inside a system finance already trusts, instead of one we built for ourselves. And it runs both ways: cost-per-outcome is exactly the muscle corporate finance is trying to build, so we work with the team that owns the spend how to read the value. Nobody else is volunteering.

And once the sprint works and people start using the tools well, do not let their best prompts rot in some random Teams channel. Capture what the power users figured out, clean it up, and push those high-value workflows back out to everyone else, so the habit scales instead of staying tribal knowledge. (That is knowledge management and a little internal marketing, Lever 6, doing real work.)

The FinOps and IT infrastructure teams are becoming some of the most mission-critical operations in the building, and almost nobody in L&D is talking to them. Nobody is going to hand us that introduction. That is most of why it is still available.

One honest caveat, because L&D has promised to measure capability for its entire existence and mostly has not. (We do some really, really good work; we usually just do not scale it well across the business.) So is this the team anyone is going to lean on, or do we keep rolling the rock up the hill with Sisyphus? The difference this time is access. The merge is the hard part, and plenty of teams will not win the data-access arm-wrestling match without some creativity around how to report on impact.

Where I would start with this

Start with the version that needs nobody's permission. Pull your own completion data for one AI rollout, your Copilot or ChatGPT/Claude/Gemini Enterprise cohort, and put who finished the training next to who actually uses the license. You already own half of that. The gap between them is your whole argument, and you built it without asking anyone for access.

Then book 30 minutes with whoever runs ITAM, FinOps, or SaaS operations, or their second in command. Lead with the offer: your cohort performance data next to their spend data, proving you can lower cost-per-active-user for one AI use case. In most organizations nobody from L&D has ever done that.

Run one merge you have never run. A single business metric a leader already cares about, learning data next to it. One cohort, one metric, one page.

Deliver it as a finding, not a deck. "We looked at X, found Y, we think it is costing us Z" is the kind of sentence that gets head nods.

Our happy path and potential to start the transition to strategic

BCG's AI Radar has 72% of CEOs now calling themselves the main decision-maker on AI, double the share a year earlier, and yet 83% of them say AI success depends more on people's adoption than the technology, and yet 59% of them expect the CHRO's influence to grow. They are running point on AI personally, and they still want the people function to matter more. They are not shutting us out. They are waiting for us to show up with something better than a completion rate. Map AI adoption to the skills the business actually needs, prove that one specific intervention moved someone from a dead login to a user with real impact, say it in the language the C-suite already speaks, and L&D is suddenly tied to something the business cares about.

This is work Lori and I do at 8Levers, so I have an obvious incentive to believe the argument. So test it. Beat it up in the comments and show me where it breaks. And if you take the cost-per-outcome idea to your ITAM or FinOps team, tell me how it went. I am collecting those stories as I have these conversations myself.