The “AI” Unit Cost and How to Account for it. Martech Futurist | July 28, 2026

Recent editions traced a chain of constraints on enterprise AI: proving what an agent did, bounding what it was permitted to do, the supply of human judgment available to supervise it, and the oversight design that lets scarce judgment cover more ground. This week a fourth constraint surfaced across the trade coverage, and it sits closer to the finance function than the technology one.

Agents that run continuously consume compute continuously. That consumption now appears inside marketing budgets as a variable cost with no settled price. Two pieces published July 27 show the industry improvising around the gap: holding companies are burying AI infrastructure charges inside principal media allocations, and one retailer disclosed per-interaction performance data for a customer-facing assistant. Alongside them, an OpenAI proposal for a new enterprise AI metric and a 500-respondent survey of creative operations leaders point the same direction. Marketing operations is acquiring a cost-accounting responsibility it did not carry twelve months ago.

My AI capability framework treats compute economics as a Force: an external pressure that bends the system from outside. Layers get designed in. Forces get contended with. This is the quarter marketers start contending.

Three Themes

Compute entered the marketing P&L before the industry built a way to price it

For roughly two years agencies absorbed AI costs and declined to bill for them, because clients expected more output for less money. Total costs have since risen far enough that absorption stopped working. The cost had to land somewhere, and it is landing in principal media, the inventory holding companies buy at wholesale and resell with a markup. Clients receive no separate token bill. They commit a fixed share of spend to principal inventory, and the markup funds the AI. Transparency in these arrangements gets negotiated case by case, which layers fresh opacity onto a practice buyers already scrutinized.

The honest reading is that nobody has a defensible price for tokens yet, and building a durable billing model takes longer than the market will wait. Compute costs are already hitting P&Ls. Capital decisions are already being made. Routing the expense through existing infrastructure is what happens when a real cost needs a home immediately.

The measurement unit is moving from time saved to cost per completed task ‍

OpenAI CFO Sarah Friar has proposed that enterprises evaluate AI on useful intelligence per dollar: whether the system completes meaningful work, what successful tasks cost in total, how accurate output remains after human review, and whether value compounds faster than spending as usage grows. The framework arrives as organizations adopt routing strategies, sending routine work to cheaper models and reserving frontier systems for genuinely hard problems.

That discipline mirrors the mechanism axis in my capability framework: use the lightest mechanism that clears the task's accuracy bar, and reach for a heavier one only when the task's openness demands it. Reaching right costs twice, once in compute and once in verification.

It reminds me of what Marcio Arnecke, Chief Marketing Officer at Apollo.io, said when I interviewed him on The Agile Brand podcast: "Time saving these days is table stakes. The real metrics you should be tracking are pipeline velocity, conversion rates by stage, ACV and expansion, win rates, and productivity per head."

Forrester's 2027 planning guidance reaches a compatible conclusion from the budget side. Larger AI budgets will not produce better outcomes where data quality, governance, and operating models remain weak, and token economics now belongs in the planning conversation.

Value concentrates where AI sits inside the workflow

Screendragon surveyed 500 marketing, creative, content, and operations leaders in the US and UK and found AI in near-universal use while only 24% of organizations have fully integrated it into everyday workflows. AI positioned alongside a workflow introduces one more system and one more handoff. The spend registers immediately. The saving never arrives.

The counterexample published the same week carries numbers. Michaels reports that shoppers using its Gemini-powered assistant convert at more than double the rate of traditional site search, with 27% of interactions producing a product click or an add to cart. Google Cloud credits the six-week build to infrastructure and catalog structure the retailer had already modernized. The economics worked because the foundation existed first.

This is Agile Brand Principle 2 applied to spend: knowing when evolution serves better than revolution. The organizations posting returns are improving workflows they already govern.

Featured Insights

Agencies Are Routing AI Costs Through Principal Medi

Digiday | How AI costs are quietly reshaping principal media deals | July 27, 2026

Seb Joseph reports on a holding company offer to absorb an entire AI infrastructure bill in exchange for committing 70% of a client's media budget to principal inventory. Multiple executives confirmed similar conversations. IAB Europe chief economist Daniel Knapp observes that agencies have long operated as futures markets through principal media and possess the risk-assessment and credit machinery to price an AI commitment, provided they can price the outcome. Outcome-based pricing remains concentrated in a handful of large advertisers with the budget and internal alignment to anchor spend to revenue.

My takeaway: ask your agency where AI cost currently sits in your commercial arrangement and what audit rights you hold over it. A cost you cannot see is a cost you cannot manage, and this one compounds with usage.

OpenAI Proposes Useful Intelligence per Dollar as the Enterprise AI Metric

MarketingProfs | AI Update, July 24, 2026 | July 24, 2026

The roundup covers Sarah Friar's proposed framework for evaluating AI investments on completed work, total cost of successful tasks, post-review accuracy, and whether value scales faster than spend. The same issue reports Forrester's 2027 planning guidance that stronger foundations matter more than larger budgets, and separate coverage showing enterprises routing routine workloads to inexpensive open-weight models while reserving premium systems for demanding tasks.

My takeaway: adopt cost per successful task as a reported metric this quarter. Add the verification hours to the numerator, since a cheap output a human has to rebuild is an expensive output.

Marketing AI Progress Stalls at the Integration Layer

Screendragon | The State of AI in Content and Creative Operations 2026 | July 21, 2026

The survey of 500 leaders across the US and UK finds near-universal AI use against 24% full workflow integration. Screendragon CMO Anne Cogan frames the gap as an operational connection problem, with AI running beside the places where work gets requested, created, governed, approved, and measured. The report positions orchestration as the layer between AI capability and business impact.

My takeaway: map every AI tool in your stack against the workflow step it serves. Any tool requiring a person to export, paste, or re-enter work is adding cost on both sides of the handoff.

Michaels Publishes Per-Interaction Numbers for Its AI Assistant

Digiday / Modern Retail | Michaels claims Google-powered AI assistant doubles conversion rate of traditional search | July 27, 2026

Sara Jerde reports that Michaels logged 75,000 conversations with its Ask Mike assistant between a quiet May launch and the formal July 21 announcement. President and chief customer officer Heather Bennett says engaged shoppers convert at more than twice the rate of traditional search users, and the company reports 27% of interactions ending in a product click or add to cart. Google Cloud attributes the six-week build to existing cloud infrastructure and an AI-ready product catalog.

My takeaway: instrument assisted sessions separately from unassisted ones before you scale. Michaels can defend this spend because it measured at the interaction level from launch.

Key Takeaways

Give the compute bill an owner. Someone in marketing operations should hold the AI cost line, forecast it against usage, and report it monthly. Absent an owner it stays distributed across tool budgets and agency fees where nobody sees the total.

Price the task. Per-seat and per-tool accounting hides the variable that actually moves, which is how much work the system attempts and how often it succeeds. Cost per completed task exposes both.

Route deliberately. Cheaper models handle a large share of marketing work at acceptable accuracy. Reserve expensive reasoning for tasks where a wrong answer carries real cost, and document which tier each workflow uses.

Fix the handoff before you add the tool. The Screendragon finding and the Michaels result point at the same mechanism from opposite ends. Integration determines whether AI spend converts to margin.

Negotiate visibility into agency AI economics. Where a partner absorbs your AI cost, that cost is funded somewhere in your commercial terms. Ask where.

It also reminds me of what Megan Lukitsch, Vice President of Global Sales, CX at CSG, said when I interviewed her on The Agile Brand podcast: "You're going to start with your baseline economics. What's your current cost to serve per interaction?"

Some Closing Thoughts

I have run two companies and sat through the diligence on five transactions, three as an acquirer and two as a seller. In every one of them the questions that decided the outcome were about unit economics. What does it cost to serve one customer, and what does that customer return. Marketing has historically been permitted a softer standard than that, partly because attribution genuinely is hard and partly because the costs were fixed enough to budget once a year and forget.

Agent-driven marketing removes that shelter. A system that acts continuously bills continuously, and the bill scales with ambition. My second book argues that Intelligence tells you what is and what is likely, while deciding what an organization should do stays a human responsibility. Deciding what a completed task is worth paying for belongs squarely in that human column. No dashboard resolves it.

Twelve years of piano lessons taught me that the practice which compounds is the practice you measure. The same holds here. Start counting cost per completed task this quarter, while the numbers are still small enough to correct.