5% of traffic to your docs is now AI agents, not humans. If your documentation isn't structured for machine readability, your product is invisible to Claude, Cursor, and every other coding agent your buyers use daily. Mintlify is built for both audiences.

See how it works

Last time I argued that pilot purgatory is a strategy failure, not a technology one that the 90% of enterprises seeing zero ROI from agentic AI are stuck because they never answered three basic questions: what business outcome, who owns it, and what kills it. But even CTOs who nail those questions hit a wall the moment they try to get budget approval. The financial model is wrong.

Here's what I mean. Most enterprises are still budgeting AI the way they budget cloud infrastructure: usage-based, metered by consumption, reconciled quarterly. For cloud compute, this works. For AI, it's poison.

The pattern I keep seeing is CTOs presenting AI business cases built on token pricing projections. They estimate prompt volumes, multiply by per-token costs, add a buffer, and present a number to the CFO. The CFO asks what happens if usage doubles. The CTO says the cost doubles. The CFO kills the project.

This isn't irrational. Usage-based pricing for AI creates a cost structure that is fundamentally unpredictable at the exact moment you need executive confidence to scale. The more successful your AI deployment becomes more users, more queries, more agentic workflows the more your costs spike in ways nobody forecasted. Token math punishes success.

The granularity is a trap too. When you're modeling costs at the token level, you've already lost the strategic conversation. Nobody in the C-suite cares about tokens per request. They care about cost per outcome. The unit economics of AI need to map to business units, not API calls.

The emerging pricing model that actually works for enterprise buyers is what I'd call the Fractional FTE model. Instead of pricing AI by consumption, you price it by the human-equivalent work it displaces or augments.

Think about it this way: if an agentic workflow handles 40% of your Tier 1 support tickets, you don't model that as X million tokens at Y cost per thousand. You model it as 0.4 FTEs across Z support agents, with a known cost basis that finance already understands. The AI vendor charges you a predictable monthly fee pegged to the scope of work, not the volume of inference calls.

This isn't just a pricing trick. It fundamentally changes how the CFO evaluates the investment. AI spend moves from the "infrastructure" line item where it competes with cloud, networking, and storage to the workforce augmentation line item, where the ROI conversation is completely different. Infrastructure spend gets scrutinized for efficiency. Workforce spend gets evaluated for leverage.

The vendors are starting to figure this out. The ones that survive enterprise procurement in 2026 will be the ones offering outcome-based or capacity-based pricing, not the ones clinging to per-token metering. Buyers are demanding predictability, and predictability is becoming a competitive moat for AI vendors themselves.

Here's how I think about building the financial case that actually survives a board review.

Start with the pain, not the model. Map every proposed AI investment to exactly one of three categories: revenue generation, cost reduction, or risk mitigation. If it doesn't fit cleanly into one of those, it's a science project. Science projects are fine just don't pretend they're business cases.

Price against the counterfactual. The question isn't what does this AI system cost? It's what does it cost to keep doing this with humans, or to not do it at all? When you frame AI spend as a displacement cost rather than a new cost, the math changes dramatically. A $200K annual AI deployment that displaces $600K in manual process costs doesn't need a complex ROI model. It needs a signature.

Build in the kill switch from day one. This connects directly to the kill criteria I talked about last time. Every AI investment should have a 90-day checkpoint where you measure actual business outcome against projected outcome. Not model accuracy. Not user adoption. Business outcome. Revenue moved, cost avoided, risk reduced. If the numbers aren't tracking, you redeploy the budget. No sunk cost attachment.

The counterintuitive thing about this approach is that it actually makes it easier to get aggressive AI investments approved, not harder. When finance sees a framework with clear outcome mapping, predictable cost structures, and built-in exit ramps, they stop being the bottleneck. The bottleneck shifts to execution speed.

Which creates its own problem. An ROI-first framework gives you permission to move fast. Teams that were stuck in pilot purgatory suddenly have budget, mandate, and urgency. But the organizations that adopted AI fastest are now discovering something uncomfortable: 59% of enterprises have deployed AI tools, but only 43% have governance policies covering them. The most aggressive adopters are the most exposed. The question becomes how do you govern AI at scale without rebuilding the bureaucracy that killed your pilots in the first place?