Stop Overpaying AI: Match the Mechanism to the Job

TL;DR: Enterprise AI costs are compounding fast because operators keep deploying agentic systems for jobs a simple rule could handle. EY’s analysis puts a single orchestrated agent interaction at roughly $1.20 versus four cents for a basic chat — a 30x premium for the same output. The fix is not cheaper models; it is choosing the right mechanism before you sign the contract.

Four Mechanisms, Four Cost Profiles

Every AI feature on your roadmap runs on one of four architectures, ranked from cheapest to most expensive: rule-based, predictive, generative, and agentic. Understanding the gap between them is the entire ballgame for operators managing six-figure monthly ad budgets.

Rule-based logic is if-this-then-that code someone wrote by hand. It costs almost nothing to run and produces identical output every time. Throw a case the author never anticipated at it, and it fails loudly — you find out immediately. Predictive models score inputs against historical patterns. Lead scoring, churn probability, propensity models — most marketing automation platforms already include these. Cheap per call, consistent across calls, useless outside the training data.

Generative models — large language models (LLMs) — produce new text or images from open-ended input. You pay more per call, and you take on a failure mode the lighter options cannot produce: wrong output written fluently, in complete sentences, with no visible seam. Agentic setups wrap a generative model in a loop with access to external tools. The loop plans a step, calls the tool, checks the result, and repeats. Every iteration resends the accumulated context window, so every small error compounds into the next step. This is the forklift. It solves the messiest long-tail problems, but it costs accordingly.

The EY Number That Should Concern Every Budget Owner

EY priced a single customer-service interaction first as a simple chat, then as an orchestrated agent workflow with tools, subagents, and retries. Result: four cents in 2023, $1.20 in 2026. Same job, 30 times the cost. The per-token price has actually fallen since early 2023 — published trackers confirm that. Enterprise AI spending still increased over the same period because consumption exploded as tasks got more complex.

Gartner estimates an agentic task uses 5 to 30 times as many tokens as a standard generative chatbot interaction. By step 20 of a multi-step job, you have paid for the original context 20 times over. Marketing compounds this faster than almost any other business function because it repeats small tasks at massive volume. A tenth of a cent per product description is a rounding error until you are refreshing 200,000 SKUs every quarter. Ninety cents a support ticket sounds cheap until you multiply it across 400,000 tickets a year and carry that number into a budget meeting.

Operators running paid media campaigns at scale feel this acutely — the same compounding math applies to AI-assisted copy generation, audience segmentation, and bid optimization when each task triggers a full agentic loop.

Why Teams Consistently Pick the Wrong Mechanism

Nobody deliberately chooses the most expensive option. It happens for four predictable reasons.

Pilots hide unit economics. A 500-interaction proof of concept produces a token bill so small it goes unnoticed. The real cost shows up in production volume, usually a quarter after the contract is signed. If your pilot did not run at 10% of projected production load, its economics are fiction.

Vendor vocabulary distorts buying decisions. “Fully agentic” describes an architecture. Buyers hear “premium tier.” These are not the same thing. Plenty of serious multi-system coordination runs on rules with no model involved at all. A vendor calling something agentic is describing how it is built, not guaranteeing it outperforms a simpler approach for your specific use case.

The mechanism question falls between teams. Procurement negotiates the per-seat price. Legal reviews data handling. Marketing judges output quality. Nobody owns the question of whether a $1.20-per-interaction agentic system is doing a job that a four-cent rule could handle. A thorough marketing technology audit that cuts across all three departments is the only way to close that gap.

Verification headcount is invisible in the AI line item. A broken rule fails loudly on the first pass. A generative model citing a refund policy that does not exist looks identical to correct output. Someone has to read it. That reading shows up as a headcount expense, not as an AI cost. Teams model computation, skip verification, and then discover the real running cost of the feature is a full-time employee reviewing its output.

What This Means for High-CAC Verticals

Operators in forex, iGaming, crypto, and legal run some of the highest cost-per-acquisition numbers in digital marketing. The mechanism question is not abstract for these businesses — it is a direct multiplier on margins that are already thin relative to acquisition spend.

A forex broker running automated lead qualification does not need an agentic loop to route a demo-request form submission. A predictive score against CRM history handles that at a fraction of the cost. Where a genuine agentic setup earns its premium is in handling the messy long tail: a prospective client who submits partial documentation, follows up across three channels, and needs context carried between interactions. That is the $1.20 job. Routing a clean form fill is the four-cent job. Operators who cannot distinguish between the two end up paying forklift rates to carry coffee cups.

The same logic applies to iGaming acquisition funnels, where AI-assisted onboarding flows can balloon in cost when agentic orchestration is applied to steps a simple decision tree already handles. For law firm intake operations, the verification problem is especially acute — a generative model confidently citing the wrong statute in a client-facing message is a liability issue, not just a cost issue.

Operators using AI agents for lead qualification should be running a mechanism audit on every workflow: which steps genuinely require generative reasoning, which can be handled by a predictive score, and which are simple enough for a rule. The answer will almost always reveal that the agentic layer is being applied two or three steps earlier in the funnel than it needs to be.

Running the Mechanism Audit Before the Contract Renews

The practical step is a mechanism-by-mechanism review of every AI feature currently in production or under evaluation. For each one, answer three questions: What is the actual failure mode if this produces wrong output? Does the task require open-ended language understanding, or does it work within a defined set of inputs? What is the production volume, and what does that cost at current per-call rates?

Most operators will find that 60 to 70 percent of their current AI workload could run on a lighter mechanism without any meaningful loss in output quality. That is not a guess — it is consistent with what happens when teams actually map tasks to the four-tier architecture rather than buying at the feature level.

For crypto operators managing token-launch acquisition at volume, the math is particularly stark. Crypto lead generation campaigns already carry high CPAs — stacking agentic overhead on top of paid media spend without justification is burning margin twice. Operators scaling forex client acquisition face the same dynamic every time an agentic workflow gets applied to a task that a predictive model already handles in their CRM.

The mechanism question is a procurement discipline, not a technical one. The teams who will control AI costs over the next 18 months are the ones who build a forcing function into every buying decision: what is the lightest mechanism that produces acceptable output at production volume? Until that question is answered with a number, not a demo, the bill will keep climbing regardless of what per-token prices do.

Operators who want a concrete starting point should map current AI workloads against the four-tier framework, price each at production volume, and run the comparison against the precision targeting and automation decisions already embedded in their paid media stack. The savings are not theoretical — they are sitting in your current contracts.

Originally reported by MarTech, August 2026.