A flat rate per completed task is not guaranteed. Martech Futurist | August 18, 2026
Your budget treats a completed task as a flat-rate item. Gartner and Harvard Business Review both published on Monday, and between them they price that assumption out of the market.
Last week I argued that AI spend buys execution, and that explanation of outcomes had become the good buyers pay a premium for. Four items published between August 17 and 18 take that argument to the invoice. The price of execution is about to stop being flat, and the work of deciding what deserves the expensive treatment moves onto the CMO's desk.
Price each workflow by the mechanism it runs on
Gartner published a forecast on Monday that AI inference costs per agentic workflow will rise more than fivefold through 2028. Will Sommer, Sr. Director Analyst, names three trends driving it: foundational model economics improve rapidly, that improved efficiency unlocks deployment of more powerful and more expensive models, and sophisticated workflows consume far more tokens than a chatbot exchange. Gartner calls the result the Inference Paradox, defined as better unit economics escalating the overall cost of AI without a clear pathway to commensurate and predictable value. The operative number for planning: routing a task to an agentic reasoning model raises provider inference costs at least five times over a basic chatbot interaction, and often much more as complexity grows.
Read that as a design constraint. In the capability framework I use, mechanism sits on its own axis crossing every pillar: rules, then predictive, then generative, then agentic. Determinism, auditability, and cost-efficiency fall as you move rightward. Flexibility and failure surface rise. The discipline the Gartner data enforces is the one that axis was built for. Use the lightest mechanism that clears the task's accuracy bar, and pay for reaching rightward twice, in compute and in verification.
Here is what that looks like with numbers attached. Take a lifecycle team running 500,000 personalization decisions a month. Push all of them through one agentic reasoning pipeline at roughly a penny each and you carry $5,000 a month. Hold the design constant and apply Gartner's curve through 2028 and the same volume costs you $25,000. Now tier it. Segment eligibility, send-time, and template selection resolve on rules and a predictive model. Say 90% of the decisions clear that bar. Route the remaining 50,000 to the reasoning model at five cents each and you carry $2,500, plus a rounding error for the deterministic majority. Same output, one-tenth the run cost, and you have removed 450,000 probabilistic decisions from the surface a human might later have to check.
Sommer puts the failure mode plainly: defaulting to generic autonomous intelligence produces unbounded costs orders of magnitude higher than an optimized product ecosystem. Every marketing organization I have worked with has at least one workflow currently defaulting.
It reminds me of what Sharon Argov, Chief Marketing Officer at AI21 Labs, said when I interviewed her on The Agile Brand podcast: "When you think about enterprise, they want something very solid, responsible, controllable, predictable, maybe even boring. So we created a brand campaign that took the most boring tasks and turned them into the most boring agents; one tagline was 'dull in chat, never invent facts.'"
I guess you could say that boring is a cost strategy.
Model the end of the vendor subsidy before your 2027 budget locks
Stacia Garr of RedThread Research published in HBR on the same day, and her piece answers the question Gartner's forecast raises: who has been paying. Her argument is that major enterprise software vendors have absorbed the cost of GPUs, inference, and tokens as a customer-acquisition play, and that vendors are now closing out the unmetered and complimentary agent tiers as they shift to usage-based charges. Garr's framing is that leaders should stop treating AI as a software purchase and start treating it as an organizational design problem covering budgeting, workforce planning, and risk management.
Put the two findings side by side and the marketing implication is direct. Your unit costs rise, and the party who was quietly covering them stops. Most martech renewals signed in the last eighteen months priced AI features as included. Pull those contracts now and find the metering language, because your 2027 plan needs a variable line where a fixed one sits today.
This is the compute-economics Force in the framework acting on the operating model, and it lands hardest on marketing operations. In the Book 2 material I am developing, Operations runs horizontally across every team as a shared capability. Routing decisions belong there. Somebody has to own the question of which workflows earn an expensive mechanism, and the answer changes as prices move.
Set a materiality threshold for disclosure and hold to it
IAB released Version 2 of its AI Transparency and Disclosure Framework on Tuesday, and the logic mirrors the routing argument on a different axis. IAB calls for disclosure when AI materially affects authenticity, identity, or representation, and rejects universal labeling. Caroline Giegerich, VP, AI at IAB, gives the reason: labeling everything teaches consumers to ignore labels, which damages advertisers.
The regulatory context tightened since the January version. California's SB 942 metadata and labeling requirements took effect August 2. EU AI Act Article 50 transparency obligations became binding the same day. New York's synthetic performer law took effect in June, and South Korea imposed labeling mandates earlier this year. IAB's research found more than half of surveyed consumers want disclosure when an ad is fully AI-generated or uses AI imagery or video, and 73% of Gen Z and Millennial respondents said clear disclosure would increase or have no effect on their likelihood to purchase.
That last figure deserves a moment. The commercial risk of disclosing sits near zero for the cohort most likely to notice. The risk lives in disclosing indiscriminately until the label carries no information.
So you now run two threshold decisions on the same content pipeline. One sets which mechanism produces the asset. The other sets which output gets a consumer-facing label. Both reduce to the same question of stakes, and both belong in a written policy that outlives the judgment of whoever happens to be shipping that day. Agile Brand Principle 4 applies without modification here: stay true to your values, and respect customers and their data. A materiality threshold is that principle expressed as an operating rule.
Name one accountable owner for every agent permitted to act
MarketingTech reported Tuesday on research from AI governance platform Optro showing where these decisions currently sit, which is nowhere. One in three organizations already use AI in critical resilience workflows and 30% have never tested for agentic AI failure. Some 58% of business leaders believe their governance controls are evolving alongside adoption while 18% report dedicated AI risk safeguards in place. In the past year, 40% encountered misleading AI outputs, 27% identified AI-related data breaches, and 26% drew regulatory scrutiny tied to their AI use.
Guru Sethupathy, GM of AI Governance at Optro, frames the structural problem: governance models designed for static manual processes cannot keep pace with autonomous systems of action. That maps to the Identity and Permissions layer, which grows load-bearing exactly as orchestration moves toward the unsupervised end. What is this agent allowed to do, on whose authority, and what is it allowed to spend. The spending clause used to be theoretical. Gartner's forecast makes it a budget control.
A RACI matrix is unglamorous and it solves most of this. One accountable owner per agent, per decision class. Marketing, IT, legal, and security consulted, with the convening kept to a minimum.
It reminds me of what Richard Rutkowski of enGen said when I interviewed him on The Agile Brand podcast: "With Agentic AI, you can now take all the clinical data you have and compare it to the medical policy and render a decision, but it can be done in parallel to what the medical director's doing until you feel comfortable that they're aligned."
Parallel operation is the cheapest form of verification anyone has found. Run the agent alongside the human, compare, and move the work over when the evidence supports it. It also generates the calibration record you will want when someone asks how you decided a workflow was safe to automate.
What to do with this
Inventory by mechanism, not by tool. List your top 20 recurring marketing workflows and mark which mechanism each one runs on today. Most teams discover they bought agentic execution for work a decision table resolves. The Agentic CX pattern applies where interactions are high-volume, rules-aware, and multistep, which describes fewer workflows than vendors imply.
Pull the metering language. Every martech contract renewed since early 2025 with included AI features. Find the usage clause and the repricing trigger. Bring the list to your 2027 planning cycle.
Write the materiality threshold down. One page defining which AI involvement in customer-facing content requires a label, mapped against California, New York, and EU obligations. Review it quarterly.
Assign single-owner accountability per agent. Name the person accountable for each agent's actions and spending limit before the next deployment ships.
Run one workflow in parallel for 30 days. Pick the workflow with the highest stakes and the clearest ground truth. Compare agent output against the human decision and record the divergence rate.
I think there’s a lesson from my college days, where I majored in photography before digital cameras were common. Hours (and more than a few mistakes) in the darkroom taught the discipline this moment requires. You do not enlarge every negative. You contact-sheet the whole roll cheaply, read it, and commit expensive paper and chemistry to the frames that earn it. Inference is the enlarger. Most of your workflows belong on the contact sheet.
Your budget treats a completed task as a flat-rate item. Fix that this quarter, and start by finding the one workflow where you are paying reasoning prices for a lookup.