The economics of enterprise AI.Six shifts reshaping technology strategy.

I spent an hour last month with a CFO who couldn’t explain why his AI costs tripled in Q1. His team had deployed a handful of copilots. Nothing dramatic. But nobody was watching the meter.

This is the new normal, and most organizations are flying blind. As enterprises scale generative AI, the fundamental unit of compute - the token - has become a primary economic discipline.

Traditional SaaS costs were predictable: Cost = Headcount × Price per Seat. Today, spending is decoupled from human users and tied to token consumption. One employee summarizing email might burn 10,000 tokens; another running reasoning tasks in the same seat burns 10 million - a 1,000× spread for identical “seats.”

The agentic multiplier

While conversational AI scales linearly, agentic AI scales geometrically. A single trigger initiates planning, sub-agents, and recursive loops - agents calling agents calling agents - producing thousands of hidden token transactions the original user never sees.

One healthcare enterprise racked up $6 million in unplanned costs in six months from unmanaged agentic loops before finance could intervene. Inside Microsoft, Meta, and Amazon, “tokenmaxxing” - feeding entire codebases into reasoning engines to hit adoption mandates - has pushed individual monthly developer bills past $150,000.

→ Try the widget: pick a task to see the cost multiplier.

In 2024, the build-vs-buy split was nearly balanced (47% vs. 53%). By late 2025 the ratio flipped hard: 76% of all AI use cases are now purchased rather than built internally (Menlo Ventures, 2025).

I call it the Great AI Flip - and it’s an admission of failure for custom builds that turned out to be academic projects, not production tools. But the decision was never really binary. It’s where you sit on a spectrum (I unpack each posture in Build vs. Buy: the spectrum is dead ↗):

• Buy. Vendor product as-shipped. Your process conforms to the software.
• Configure. Live inside a platform (Salesforce, ServiceNow, Workday). Their framework, your logic.
• Extend. Keep a trusted core; build differentiating capability on top via APIs and agents.
• Compose. Orchestrate multiple systems where no single vendor owns the truth.
• Build. Own the full stack. The capability is the business.

The hidden cost of moving right

AI-assisted “vibe coding” enables rapid prototyping but creates a dangerous illusion: the final 20% of a production release - security, governance, reliability - is 80% of the engineering effort. Teams celebrate a working prototype, then spend six months making it production-ready. The demo looked 80% done. It was 20%.

AI reduces the cost of writing code. It does not reduce the cost of managing complexity. Eight traps emerge - and they compound the further right you move on the spectrum. I take each one apart in The 8 traps of vibe coding ↗.

I’ve sat in budget meetings where finance had no idea how to model next year’s software costs. The pricing meter changed mid-contract. Nobody knew what to put in the forecast.

The old playbook - negotiate seats, forecast by headcount - doesn’t work anymore. When agentic AI completes a workflow end-to-end, the value isn’t tied to human engagement time. An agent can automate 100% of a workflow while requiring zero seats.

Two forces collide: value now comes from automated outcomes, but high inference costs force vendors to monetize usage. Layering AI fees on top of legacy seats is the trap everyone’s trying to escape.

A worked example: case management, 100 users

Run the math on action-based credits and the monthly number is small and predictable - about $1,800. That’s the comfortable part.

What’s not predictable is the spike: a product recall, a service outage, a seasonal rush. Consumption pricing means your software cost now moves with business volume. That’s a fundamentally different risk profile than fixed seats - and it belongs in the forecast as a range, not a line.

The four IT archetypes (McKinsey)

When I assess an organization’s AI readiness, this is where I start. Plot how much you spend keeping the lights on against how much you free up for change, and the quadrant you land in tells you almost everything about your trajectory.

Two of these quadrants feel like progress and aren’t. One compounds; one quietly sinks AI spend into legacy. The work of the next two years is moving toward the low-Run, high-Change corner before your competitors get there.

Where value actually shows up

Employee value makes teams faster - but time saved is not money earned. At best you’re optimizing expenses. Clear ROI embeds intelligence into core products: revenue lift, cost reduction, margin. Strategic value creates new products and business models - highest risk, highest reward.

The turning point in my own career: I stopped reporting “30% faster ticket resolution” and started showing “$20M added through AI-driven products.” Same technology. One gets a pat on the back; the other gets a seat at the strategy table.

Focus 80% of investment on ROI and Strategic value. That’s why the 81% are stuck - they’re optimizing tasks instead of targeting P&L.

Workflow debt - and the four-stage fix

Two forms of debt will undermine AI before it starts. Manual process fragmentation: work moving between systems by email, spreadsheets, and human memory - each handoff a failure point. Broken automation layers: brittle old integrations held together by tribal knowledge that snap when agents expect clean data flows.

Automating a broken process generates unstable “AI technical debt” that compounds faster than the legacy debt you already have. Refactor the business logic before deployment, not after.

This is the same lesson I learned at GE twenty years ago: lean before digitize. Fix the workflow first, then automate it. Different technology, identical principle.

Done in sequence - discovery, refactoring, test generation, self-healing - AI becomes a debt-reduction tool, not just a coding assistant.

I’ve had this conversation with dozens of executives. The math seems obvious on a whiteboard: AI does the work, we need fewer people. But the whiteboard isn’t where work actually happens.

AI is making individuals more productive - but “more productive” doesn’t mean “fewer people needed.” In practice: it compresses routine tasks but expands the surface area of judgment work; faster execution creates more iterations, not fewer hours; automated first drafts need more sophisticated review. The time saved on production shifts to validation, exceptions, and coordination.

The organizations capturing real value aren’t cutting headcount. They’re redesigning what humans do - from routine execution to judgment, oversight, and accountability.

The elevation framework

“Elevate before you eliminate” is an economic strategy, not an HR talking point. Role redesign should reallocate time into four value-weighted categories - and notice that AI expands oversight work while leaving judgment, relationships, and accountability firmly human.

The elevate-first audit

Before declaring headcount redundant, run six checks - each with a tempting layoff-first shortcut that quietly destroys capability. Task-level proof over role-level slogans. Paid upskilling over “reskill on your own time.” Internal mobility over external hiring. Apprenticeship preservation over ladder destruction. Worker consultation over top-down mandates. Financing transparency over using payroll as the AI piggy bank.