Here's a question I ask every enterprise client before we scope an AI workflow:

How much budget do you want to allocate to this specific use case?

Almost every time, the room pauses.

Then someone gestures toward the broader picture:

• The annual platform commitment
• The Microsoft 365 AI add-on
• The discretionary IT budget

They're solving for the wrong question.

Traditional software gets cheaper as usage scales.

AI often gets more expensive.

Every interaction with a large language model carries a discrete cost. AI behaves more like labor than software. ^[1] That changes the economics of every deployment decision — and most enterprise budgeting processes were never designed for it.

The Per-Workflow Cost Model

The unit of analysis in AI budgeting is not the platform.

It is the workflow.

A document review process costing $5 per run, executed 100 times per day, spends:

• $500 per day
• $15,000 per month
• $182,500 per year

The number most organizations fail to model correctly is production-scale volume.

A process used 10 times per day during a pilot frequently runs 200+ times daily once embedded into operations.

Whether the investment makes sense depends entirely on one question:

What happens to the capacity that was freed?

This is where most AI budget conversations collapse.

Organizations approve spend at the portfolio level without assigning cost and return targets to individual workflows.

The result is already visible in the data:

• 84% of IT leaders say AI returns are beating expectations
• Only 43% actually require teams to track impact ^[3]

Confidence without measurement is not ROI.

The subscription model trained organizations to think of AI as flat-fee infrastructure.

That framing is borrowed time.

Anthropic users are reportedly consuming up to $8 in compute for every $1 of subscription revenue. ^[2] GitHub Copilot enterprise accounts hit hard usage ceilings as power users exhausted monthly allocations in under 90 minutes. ^[2]

As frontier labs move toward profitability pressure, subsidy economics disappear.

Budgeting at the platform level while costs accumulate at the workflow level is how organizations get blindsided six months after deployment.

“Most companies struggle to find ROI in AI because they are looking for ROI from an investment in AI itself. The right question is: what business outcome did AI enable?” ^[9]

— Max Chan, CIO, Avnet

Automation ROI: The Cleaner Case

When AI completely replaces a discrete task, the math is relatively straightforward.

1. Measure the Baseline

Before deployment, document:

• Time required
• Quality level
• Fully-loaded labor cost

Not estimates. Actual measured work.

Salary alone is insufficient. Knowledge-worker cost is typically 1.25–1.4× base compensation once benefits and overhead are included. ^[7]

Without a measured baseline, ROI becomes unverifiable.

2. Model Production Cost Properly

AI economics change dramatically at scale.

A process running:

• 10 times/day at $2/run = ~$600/month
• 500 times/day at $2/run = ~$30,000/month

Both may be justified.

Neither justifies itself.

3. Track Redeployed Capacity

Savings only exist if the freed time gets redirected toward something more valuable.

Two hours saved per employee that disappears into lower-priority work is not ROI.

It is availability.

McKinsey found that 60% of organizations deploying generative AI in at least one function still saw no enterprise-wide EBIT impact. ^[5]

The activity exists.

The value capture does not.

Augmentation ROI: The Harder Case

Augmentation is where most enterprise AI ROI claims quietly break apart.

A salesperson writes proposals faster.

A developer produces code in 20 minutes instead of two hours.

An analyst drafts reports in half the time.

The assumption is immediate:

Faster work equals more output.

Usually, it does not.

Early enterprise usage data suggests generative AI users save roughly 5–6% of weekly working hours on average. ^[4]

Yet aggregate productivity impact across organizations often remains near 0–1%. ^[4]

The savings disappear somewhere between the employee and the P&L.

The reason is structural.

When one stage of a system accelerates, the bottleneck simply relocates to the next slowest step. ^[8]

A developer writing code 40% faster does not increase shipped features if code review remains constrained.

The constraint moved. The output did not.

Johnson & Johnson encountered this exact pattern when AI accelerated invoice generation faster than downstream finance teams could absorb the increased volume. ^[6]

Local productivity gain.

System-level congestion.

The real augmentation ROI test is not:

Did the task get faster?

It is:

Did the organization produce more?

If employees complete the same amount of work faster without increasing throughput or scope:

You purchased speed, not output.

Speed absolutely has value:

• Reduced cognitive load
• Faster response cycles
• Improved employee experience

But unless someone can explain where the recovered capacity went, it does not justify enterprise-scale workflow costs on its own.

Speed without deployment is a cost, not a return.

The organizations successfully extracting augmentation ROI all do the same thing:

They measure output at the system level, not the task level.

A team producing:

• 18 reports per week after AI adoption instead of 12

That is measurable ROI.

Producing the same 12 reports faster requires a different business case entirely. ^[3]

A Note on Model Selection

Most per-workflow ROI models ignore another major variable:

Whether the right model is doing the work.

Engineering teams frequently prototype using frontier models and never optimize cost before production rollout.

The result:

A simple routing or classification task ends up billed at 10–40× the necessary compute cost.

The economics approved during pilot phase no longer resemble production economics six months later.

That pilot-to-production gap — and how organizations should calculate true per-run economics — is the focus of Part 2.

The Four Questions

The budget conversation that matters is not:

What is our AI spend?

It is:

1. What does this workflow cost per run at production volume?
2. Is this replacing work or augmenting it?
3. If replacing: what happens to the freed capacity?
4. If augmenting: did throughput actually increase?

None of this requires advanced infrastructure.

It requires treating AI as a variable operating cost tied directly to measurable performance outcomes.

The organizations extracting durable AI ROI are not the ones with the largest budgets.

They are the ones treating every workflow like its own P&L line.

Measuring output instead of activity.

Tracking where capacity actually goes.

The bill for every AI workflow is already accumulating.

The question is whether the return is.

Endnotes

1. “Your Claude API bill is higher than your revenue: Why simple Python tasks are blowing up AI costs” — CIO, May 21, 2026.
   https://www.cio.com/article/4175244/your-claude-api-bill-is-higher-than-your-revenue-why-simple-python-tasks-are-blowing-up-ai-costs.html
2. “Every AI Subscription Is a Ticking Time Bomb for Enterprise” — State of Brand, May 2026.
3. “CIOs should beware the AI confidence trap” — CIO.com, May 21, 2026.
   https://www.cio.com/article/4175346/cios-should-beware-the-ai-confidence-trap.html
4. “2026: The Year AI ROI Gets Real and Forces a Strategic Fork in the Road” — Wndyr, 2026.
   https://www.wndyr.com/blog/2026-the-year-ai-roi-gets-real-and-forces-a-strategic-fork-in-the-road
5. “From Promise to Impact” — McKinsey & Company, April 2026.
6. “A Need for Nuance: The Economist's Andrew Palmer” — Me, Myself, and AI / MIT Sloan Management Review, May 19, 2026.
7. “Measuring AI ROI: Why you're doing it wrong and the 7 Steps to fix it” — Everyday AI / Jordan Wilson, March 3, 2026.
8. The Goal — Eliyahu Goldratt, North River Press, 1984.
9. “Avnet CIO Max Chan details AI-driven supply chain transformation strategy” — CIO Leadership Live, January 7, 2026.