Don't Use a Forklift to Move a Coffee Cup: Your AI Bill May Have Gone from 4 Cents to $1.20

A while back, a friend of mine who runs e-commerce vented to me. His company deployed AI customer service last year. During the demo, the whole room applauded. When renewal time came around this year, finance turned pale. He asked me: weren't AI prices supposed to be falling? Why does ours keep climbing the more we use it?

I said, before you look at the invoice, answer one question for me: do you know what's actually running underneath the "AI" you bought?

He froze.

He's not alone. Among the enterprise teams I've worked with, almost nobody deliberately picks the most expensive option — yet overpaying happens every single day. Today, let's untangle this muddled account.

Four Machines, One List

Peel back the AI features on your list, and underneath there are only four kinds of machines.

What's a rules engine? It's "if... then..." someone wrote down in advance. Under $50 and within 30 days? Auto-refund. That simple. It costs almost nothing to run, and it produces the exact same result every time. The flaw is equally obvious: when a situation the rule-writer never imagined shows up, it stalls on the spot. But stalling is a good thing — you know immediately that something broke.

The second is the predictive model. It takes patterns learned from historical data and scores each new input. Lead scoring, churn probability — your marketing automation platform is already full of them. Cheap per call, stable output. But the moment it meets something absent from its training data, it silently hands you a useless answer.

The third is generative AI. Large language models understand however the customer writes — odd words, slang, sentence patterns they've never seen; nothing fazes them. The cost: pricier per call, plus a brand-new failure mode the first two machines can't produce: it will cite, in fluent, complete, seemingly airtight sentences, a refund policy you don't actually have.

The fourth is the agent. Wrap a generative model in a loop that can call tools: look up the order, check the policy, issue the refund, send confirmation through the customer's preferred channel — all in one flow. This is the forklift among the four machines.

Of course a forklift can move a coffee cup.

But every step of the loop resends all the context accumulated so far, and every small error at one step carries into the next. Flexibility and coverage go up — and so do the bill and the risk.

A single customer refund request can be handled by all four machines. The difference is that the price can be 30x apart.

4 Cents, and $1.20

EY ran the numbers on this, and when I saw them, my heart sank.

The same customer service interaction cost roughly 4 cents in 2023 with a simple chatbot. By 2026, swap in an agentic pipeline with tools, sub-agents, and retries: roughly $1.20.

Same task. Thirtyfold.

EY added a stinging reminder: the invoice from the model vendor is only part of what you actually spend. Where the rest hides — we'll get to that.

Some of you may object: no, token prices have been in free fall since early 2023 — the public price trackers all show it. Costs are clearly dropping.

Right, the unit price is dropping.

But the total bill is rising.

Gartner has estimated that an agentic task consumes 5 to 30 times the tokens of an ordinary generative conversation. Why? Because by the time the machine reaches step 20, it has to recheck the original request and pay for the opening context all over again. Over 20 steps, you bought the same context 20 times.

The unit price drops to 30% of what it was, while usage grows tenfold. You tell me: is the total rising or falling?

Why Marketing Takes the Hit

There's also an amplifier, and it lands squarely on us marketers.

A ten-cent rounding error per product description — sounds trivial, right? Except you refresh 200,000 SKUs every quarter. Ninety cents per ticket versus twelve dollars of manual handling — practically free, right? Except you have 400,000 tickets a year. Take those numbers into the budget meeting. Go ahead, try.

Nowhere else in the company does any function repeat the same small task this many times.

Repetition is the interest you pay on choosing the wrong tool. Another team chooses wrong and pays once; marketing chooses wrong and pays every quarter.

Four Pitfalls

So why do people still jump into a bill this clear? From what I've observed, there are four utterly ordinary reasons.

First, pilots can't measure production costs. A 500-interaction PoC produces a token bill too small for anyone to notice. The real economics only surface after the contract is signed and a quarter or two of production volume has run. By then, you have no leverage.

Second, mistaking the construction method for the product tier. "Generative" and "agentic" describe how a system is built, but the market has slowly turned them into price tiers. When a vendor says "fully agentic," they mean architecture; the buyer hears "upgraded edition." Those are two different things. Plenty of ordinary assistive work runs on nothing more than a generative model underneath, and plenty of genuine cross-system coordination runs on plain rules — no model at all.

Third, the people who own the mechanism aren't in the room. Procurement negotiates unit prices, legal reviews data, marketing looks at results. As for whether the system is a sledgehammer cracking a nut? That's a question for the engineering and data teams. And they're usually not at the procurement table.

The fourth pitfall is the deepest: verification costs headcount. When rules break, you can see it on the first pass. Generated content that cites a policy you don't have looks exactly like the correct answer. So someone has to read it. That "reading" doesn't appear in the AI budget line — it appears in headcount. Many teams budgeted for compute and forgot verification, only to discover that the true running cost of the feature is one full-time person reading its output.

What to Do

After all that, the prescription is really one sentence:

Always start with the lightest machine; only move up to a heavier one when the light one runs out.

Concretely? Three things.

Add one question to every vendor conversation: what mechanism actually powers this feature underneath? At our real volumes, including human review, what does it cost per year? Paying generative prices to do a predictive model's job is a bleed that never triggers an alarm — and the more you use it, the faster it flows.

Audit your inventory across the four layers. Go find that combination of "the simplest task running on the most complex semi-automated pipeline." It's most likely the biggest muddled line in your AI budget — and often the easiest to cut, because your tech stack almost certainly already has a simpler tool that gets roughly the same result.

Price at production volume before the pilot. Take the vendor's per-interaction price, multiply by your real annual volume, then add review and rework time. Finish this math while you still have negotiating leverage.

Of course, some needs genuinely merit the heavy artillery of agents; some jobs are cheapest with a mix — rules as the base, agents catching the long tail. And for some jobs, going back to that honest old rules machine is actually faster and steadier.

Before you call in the forklift, look at what you actually need to move.

And here's hoping your next AI bill is no longer a shock.