The Hard Part of AI Isn’t AI

If you spend any time on LinkedIn, your feed has already decided this is easy. Thirty prompts every CFO should be using in 2026. A hundred Claude prompts for finance work, copy and paste. Somebody cut their close from ten days to two, and all it took was a weekend and a subscription.

I’ve spent eighteen years inside the finance functions of high-growth companies, and I want to offer a public service announcement: be skeptical. Not because the tools don’t work. They work better than most finance leaders realize. The trouble is that the company in the demo has one legal entity, one bank account, clean historical data, and a founder who approves everything. The company that actually raises institutional capital has six cash accounts across two banks, four sales channels with different settlement logic, payroll in eleven states, approval workflows the board requires, and revenue recognition that will be tested by auditors. Then add the two problems nobody posts about: the data is dirtier than anyone admits, and almost none of what you assumed was a process was ever written down (a checklist is not a process). It lives in people’s heads, in the exception-handling and fast-thinking they do without noticing.

Automating a process requires the “process” to exist as a process, rather than as a set of habits.

The adoption data backs this up, loudly. McKinsey’s latest State of AI survey found 88% of organizations now using AI somewhere, while nearly two-thirds haven’t begun scaling it and only 39% can attribute any bottom-line impact to it. BCG puts the share of companies capturing substantial financial gains at roughly 5%, and prescribes a 10-20-70 split for the winners: 10% of the effort on algorithms, 20% on data and technology, 70% on people and process. Anthropic’s own economic data shows the gap between what models could do and what they’re actually used for is widest in business, finance, and administrative work, the exact categories where theoretical capability is highest.

And if you want the plain-English version, it reached the biggest business podcast in the world last week. Mark Cuban, on All-In, made the point with a question that should be famous: if AI were what the hype says, you would just ask AI how to implement itself. Instead, Microsoft is hiring 6,000 forward-deployed engineers, and the frontier labs are building out services arms to install their own product inside enterprises. The richest companies on earth, with the best model access on earth, are solving deployment with armies of humans. AI makes a lot of things easy. Changing a century of accumulated business/accounting practice is hard.

What We Built

Propeller has lived on the hard side of that line for eighteen years. In 2008, we led the movement to bundle finance and accounting into a managed service, back when it didn’t have a name, embedding senior finance talent inside companies that couldn’t yet justify the full-time hires. Since then: 1,500+ companies served, 24 of them now unicorns, 250+ companies exited and a team of 250+ full-time professionals doing the work. A decade after we pioneered it, the market eventually called the category “fractional” finance.

We built it, grew up inside it, and then grew out of it.

Most Firms Like Ours Can Fast-Follow. We Can’t.

I want to be fair to the rest of our industry here. Most firms that get grouped with us are smaller, earlier-stage, and more transactional, and they are built (appropriately) to a different standard. If your clients are pre-seed companies that need a clean set of books and a monthly close, individual AI adoption is a perfectly rational strategy. Let your best people experiment, adopt what works, fast-follow the market. Nothing breaks if a workflow hiccups.

Our situation stopped looking like that years ago. Propeller moved decisively upmarket: later-stage companies that stay with us for years, audit readiness, ERP migrations, sell-side diligence, exits, and the buyer integrations that follow. Our current roster of 300+ active clients generates roughly $5 billion in combined revenue. At that altitude, the work product doesn’t get graded on effort. It gets tested by auditors, picked apart in diligence, and relied on by boards making irreversible decisions. A company preparing for a sale cannot run its finance function on an enthusiast’s prompt library, and neither can the firm it hired. What that tier of work demands is Finance-Grade AI, as opposed to consumer AI: every output traceable, every judgment reviewable, and a named human accountable for whether the number can be trusted.

The Accidental Head Start

The honest version of this story includes some luck. We didn’t foresee AI when we designed this firm. But several decisions we made for entirely different reasons, turn out to matter enormously right now.

We built systemically instead of heroically: one way of working, refined across hundreds of engagements, rather than a collection of talented freelancers each doing it their own way. We built with W2 teams instead of a contractor marketplace, which means the firm’s knowledge stays in the firm. And above all, we standardized: common processes, common chart-of-accounts logic, common close disciplines across every client, with the genuinely unique parts handled as deliberate exceptions rather than accumulated habits. For eighteen years that standardization was how we managed quality and continuity. It turns out it’s also the precondition for agentic work. You cannot automate a process that exists in three hundred slightly different versions.

Common structure buys something else, too. When hundreds of companies run finance on the same underlying architecture, what happens inside one of them can inform all of them: how a pricing change actually flowed through margin, what working capital did at a given stage, which patterns precede trouble. A finance function built this way doesn’t reset when someone leaves, and it doesn’t plateau when someone stays. It gets smarter over time, because it’s learning from more than one company’s history.

Eighteen years ago, we led the movement to bundle finance and accounting as a managed service. The movement we intend to lead now is building institutional AI into service delivery itself.

Institutional AI vs. AI Guy

That word, institutional, is doing precise work. George Sivulka published an essay through a16z this spring that gives the distinction its best framing yet. When factories first replaced steam engines with electric motors in the 1890s, output barely moved for thirty years; the gains only arrived when they redesigned the factory floor around the new technology. His argument is that AI has made individuals dramatically more productive while their companies have captured almost none of it, because we’ve swapped the motor without redesigning the factory.

The corporate version of swapping the motor is one I suspect you’ve already met: the AI guy. The controller who started experimenting on nights and weekends. The analyst who rebuilt the forecast around a workflow nobody else understands. The gains are real, and so is the fragility, because everything that person built lives in their head and their personal account. The judgment about when to trust the output lives nowhere at all. I’ve written before about the talent paradox: key finance hires leave and institutional knowledge walks out with them. The AI guy is the talent paradox on fast-forward, because these are now the most recruitable people in the market. Your finance function never actually got faster; one person did, and you were renting the difference.

Cuban added the detail that should worry even the companies whose AI guy stays. Agentic workflows decay. As the underlying models change, automations built against last quarter’s behavior quietly stop matching it; in his words, “agents get bored and they drift.” Drift isn’t a bug to be patched once. It’s a permanent property of building on a substrate that keeps moving, which means the maintenance obligation never expires. An AI guy who visits, fixes your fails, and moves on leaves you with workflows that are already decaying. The only durable answer is an institution that stays and stays accountable.

The Iron Man Suit

All of which brings me to the part the automation crowd gets backwards: where the humans go.

The right mental model for AI in a serious finance function is the Iron Man suit. The suit doesn’t fly itself, and nobody wants it to. It amplifies a human who remains responsible for every decision made inside it. That’s the design principle for everything we’re building: our people orchestrate the system design, run the agentic processes, review what comes out, and attest to whether it can be trusted. We are the last mile of accountability to our clients. When a number leaves our hands for a board, a lender, an auditor, or a buyer, a named human has stood behind it. Finance-grade means someone signs.

And this human layer is permanent, which is worth saying plainly because so much AI commentary treats people as scaffolding to be removed later. Drift alone guarantees the opposite: on a moving substrate, somebody has to notice when a pattern starts misfiring, and models don’t interrupt themselves. Moving people out of data entry and into judgment, review, and attestation isn’t the transition phase of agentic finance. It is agentic finance.