How to stop overpaying for AI complexity

You wouldn’t purchase a forklift to carry a coffee cup. Yet that’s effectively what happens when enterprises apply complex AI systems to tasks that don’t need it. As AI adoption increases, it is essential to understand the costs of running these systems and their impact on existing people and processes.

The initial promises of what an AI system can do often don’t account for long-term costs, the human oversight required, and the total cost of ownership. Those, along with potential failure points when a highly automated system breaks down or misses a step, are often glossed over in the excitement of adding a system to the stack. In most marketing organizations, that decision is made by a different person or team than the one managing procurement.

Definitions matter when there are four potential approaches

Nearly every AI feature on your list of priorities uses one of four mechanisms, which, ordered from simplest to most elaborate, are: rule-based, predictive, generative, and agentic. Briefly exploring each will quickly identify where they diverge.

Rule-based is if-this-then-that logic someone wrote by hand. It costs almost nothing to run and produces the same output every time. Hand it a case the author never anticipated, and it fails in the open, so you find out on the spot.

A predictive model scores an input against patterns in historical data. Lead scoring, churn probability, and propensity models are already sitting in your marketing automation platform. Cheap per call, consistent across calls. Outside the training data, you get nothing useful back.

A generative model produces new text or images using a large language model (LLM) from open-ended input, so you can send it phrasing that no rule anticipated. You pay more per call. You also take on a failure mode that the two lighter options can’t produce: wrong output written fluently, in complete sentences, with no visible seam.

An agentic setup wraps a generative model in a loop with access to tools. It’s the biggest forklift of the four scenarios. The loop plans a step, calls the tool, checks the result, and runs again. That’s how you get coverage of the messy long tail, and it’s also where the cost comes from, because every step resends the accumulated context and every small error carries into the next one.

As complexity increases, flexibility and reach go up, predictability and auditability fall away, and risk increases for everything from small errors to mistakes that could cost you and your customers.

Platform buying conversations often occur at the feature level, with the mechanism sitting a level below, where the real costs (and risks) lie.

Example: 1 refund request with 4 different cost implications

To illustrate the vast differences among these four approaches, let’s use a customer refund request as an example and see how each method handles it.

Rule

The AI follows a clearly labeled policy table stating that if the refund request is under $50 and within 30 days, it can be automatically approved by the system. This method is free to run once it’s set up, yet it encounters immediate challenges when a request doesn’t fit the predefined table. Even with complex rules, any requests not accounted for by those rules will grind the process to a halt.

Predictive

Using this approach, a machine learning model scores each request as one of several potential outcomes, such as approve, deny, or escalate. This approach is fast and consistent, but, like the rule-based approach, it runs into issues if a request doesn’t resemble its training data.

Generative

An LLM drafts a resolution based on what the customer actually wrote, regardless of previous requests. Odd phrasing or “out of the norm” requests stop being an issue, so that’s the upside. On the other hand, the cost per call increases, and there’s a risk that, every so often, the draft cites a refund policy you don’t have.

Agentic

Finally, an agentic model runs in a loop, taking the customer’s request, retrieving the order record, applying the policy, issuing the refund, and sending the confirmation via the customer’s preferred channel, with minimal human interaction. This approach lets you get coverage of the cases the other three miss. That said, due to information processing and option consideration, this has the highest price by a wide margin.

Those costs are a primary concern as enterprises move out of their AI experimentation phases and into a new phase of accountability. What was four cents in 2023 is roughly $1.20 in 2026. That’s EY’s accounting of one customer-service interaction, priced first as a simple chat and then as an orchestrated agent workflow with tools, subagents, and retries. As we just saw, this is essentially the same job with about 30 times the cost. Their broader point is that the invoice from your model provider captures only part of what you’re actually spending.

Recommended Articles

The next AI opportunity is deciding where to spend

Aug 17, 2026

Here’s the first martech category replaced by AI

Aug 13, 2026

B2B growth doesn’t end at closed-won

Aug 11, 2026

Despite growing industry maturity, the bill keeps climbing

As costs go up, it’s tempting to push aside the tokenomics doomsaying by citing past precedents in which the cost of delivering technology solutions has fallen as demand has increased. That’s already happening, as per-token prices have fallen sharply since early 2023, according to published trackers. Yet enterprise AI spending increased over the same period.

This increase stems from greater consumption as tasks become more complex. Gartner estimates that an agentic task uses 5 to 30 times as many tokens as a standard genAI chatbot interaction. When models have to continually reorient themselves to the original request and use generative AI to interpret it, by 20 steps into a job, you’ve paid for the original context 20 times over.

In addition to consumption, sheer volume accounts for the rest, and marketing is a prime driver of this volume. A tenth of a cent per product description is a rounding error until you’re refreshing 200,000 SKUs every quarter. Ninety cents a ticket reads as a bargain compared to a $12 human interaction. Then you multiply it out across 400,000 tickets a year and carry that number into a budget meeting.

Few functions in the business repeat the same small task as often as marketing does, which is why a carelessly chosen mechanism here compounds faster than it would elsewhere.

What distracts from choosing the best approach

In my work advising enterprises on martech and AI adoption, I can’t think of a single team that deliberately chose the most expensive way to do something, yet it happens all the time, for four ordinary reasons.

Pilots aren’t a true indication of unit cost at scale

The consensus is that a smaller-scale proof of concept is the best way to understand and pressure test how a larger system will function. Yet a pilot may run a mere 500 interactions and produce a token and infrastructure bill so small that it goes unnoticed.

The real economics are determined by production volume, usually a quarter or two after the contract is signed.

The vocabulary describes construction and gets confused with quality

Generative and agentic are ways to describe how a system is built. In the market, they’ve come to function as product tiers, so a vendor saying “fully agentic” describes an architecture, while the buyer hears it as a better version of the same product. These aren’t the same thing as product tiers. A generative model sits under a great deal of ordinary assistance work, and plenty of real multi-system coordination runs on rules with no model anywhere in it.

The system’s users don’t own the mechanism question

Procurement negotiates the per-seat price, legal reviews data handling, and marketing judges the output. This means engineering and data teams review whether the most complex agentic approaches are used in ways that are overkill, or whether the system will use the right tool for the job (e.g., a forklift for everything). There may be a disconnect in how all of this gets communicated and discussed.

Verification of system outputs requires headcount and a separate budget

The most complex agentic mechanisms fail differently than more straightforward rules-based ones. A broken rule is visible on the first pass. Generative output citing a policy you don’t have looks exactly like correct output, so someone has to read it, and that reading turns up in a headcount line instead of in the AI line item.

While all four of these need careful consideration, the last one poses the biggest challenge and takes the most time and money to solve. Teams model the computation, skip the verification, and then find out the running cost of the feature is a person reading its output.

Using the lightest mechanism that matches desired outcomes

When do you need a forklift to carry your work, and when is the job too small to justify the cost or complexity? The details differ by use case, but as a rule of thumb, aim to adopt the lightest approach that still meets your needs. 

Even when an innovative, fully agentic approach seems like the right fit, use approaches and technologies selectively. Avoid scaling up costs when added complexity doesn’t produce better results or solve a problem simpler approaches can’t handle.

Also, the end product isn’t always just one method in a silo. That means use agentic approaches for truly complex needs, and keep rule-based mechanisms for areas that either don’t need greater complexity or interpretation, or where those may require human oversight, introduce potential hallucinations, or other artifacts that slow down a process.

Consider these three things as you head into 2027 planning and potentially heated discussions about 2026 ROI and outcomes.

Add a question to every vendor conversation

Which mechanism is this under, and what does it cost at our volume, including review? Paying generative prices for a predictive problem is a quiet margin leak that widens as adoption grows.

Evaluate your current usage of the four different AI approaches

Find the most complex, semi-autonomous workflow sitting on the simplest task. That’s usually the largest unexamined line in an AI budget, and often the easiest to fix, because there’s likely a way to produce similar results more simply using another set of tools in your stack.

Price your pilot projects at production volume before you start

Take the vendor’s per-interaction figure, multiply by your real annual count, and then add review and/or revision time on top. Do the math during your negotiations while you still have leverage and time to reconsider initial assumptions.

Some needs can more than justify a more complex approach; some may do best by combining methods (scaling complexity only when/if needed); and others may actually perform better with simpler, tried-and-true methods. Figuring out whether you need that forklift means taking a closer look at what you need to carry.