The Gist

Inference paradox. Gartner forecasts AI inference costs per agentic workflow will rise more than fivefold through 2028, even as token prices fall. Compounding multipliers. Reasoning agents can cost up to 150 times more than a simple chatbot to complete the same task. Service exposure. High contact volumes, multi-turn loops, failed handoffs and usage-based billing concentrate the cost risk in customer service. Tiered design. Routing each contact reason to the cheapest model that can resolve it, with a capped reasoning budget, keeps agent costs bounded.

Every AI business case I have reviewed in the last two years rests on the same unspoken assumption: that cost per interaction falls over time. Models get cheaper, tokens get cheaper, therefore the economics improve on their own. Sign now, and the maths gets better while you sleep.

Three days before I wrote this, a forecast landed that takes that assumption apart. Token prices are indeed falling, steeply. And the cost of running an agentic workflow is going up more than fivefold through 2028.

Both statements are true at the same time. That is the whole problem, and it has a name.

I am writing this for customer experience and service leaders specifically, because our function is where the exposure concentrates. We run the highest interaction volumes in the enterprise. We are the function most aggressively sold on agentic automation. And our business cases were built on a cost curve that is about to bend the wrong way.

What follows is what the research actually says, the arithmetic that should reset your numbers, and — more importantly — what can be done about it. This is a solvable problem, but not by waiting.

What Gartner’s Agentic Workflow Cost Forecast Says

Gartner’s forecast, published on Aug. 17, is that AI inference costs per agentic workflow will rise more than fivefold through 2028, as products move from assistive features to multistep execution. Senior Director Analyst Will Sommer puts the warning plainly: product leaders cannot rely on more efficient token economics to rationalize AI costs, because each successive generation of capability will demand more, and often more expensive, tokens.

How Gartner’s Inference Paradox Works

The mechanism is named the Inference Paradox — better unit economics escalating the overall cost of AI without a clear pathway to commensurate and predictable value.

Add CMSWire as a Preferred News Source

Three forces drive it, and they are worth separating because they get conflated constantly. Foundation model cost economics really are improving fast. Those efficiency gains unlock deployment of more powerful and more expensive models. And sophisticated workflows consume far more tokens than a simple exchange. Net result: tokens get cheaper, but not nearly fast enough to offset what we are now asking them to do. In the analysts’ phrasing, the rate of innovation is outpacing the cost curve.

Why an AI Agent Costs More Than a Chatbot

Sommer’s comparison is the clearest explanation I have seen of why an agent is not just a better chatbot. A chatbot reads a query and returns a probabilistically reasonable answer. An agent must constantly reason, negotiate and question itself. Every one of those steps is billable.

How Agent Cost Multipliers Reset the Customer Service Business Case

The headline number is fivefold. The detail underneath it is considerably more alarming, and this is where most readers stop paying attention exactly when they should start.

Reporting by Computer Weekly on the underlying research sets out the modelling. Training a medium-sized agentic model with advanced reasoning costs about 2.5 times more in hardware than a simple chatbot of the same size. Inference for that agentic model runs roughly five times greater. And the agentic model then needs five to 30 times more tokens on average to solve an equivalent task.

Those multipliers compound. The conclusion, and I want you to read this one twice: a task that would cost a simple chatbot one cent could cost an AI agent up to one dollar fifty.

That is a factor of 150 for the same job. Not five. 150.

What the Agent Cost Gap Means for a Service Operation

Now apply that to a service operation. If you handle two million contacts a year and route even 10% of them to a reasoning agent rather than a deterministic flow, you are not looking at a rounding difference. You are looking at a line item that did not exist in the approved business case.

The cost varies enormously by what you ask the system to do. Computerworld’s account of the Tokenomics Model the firm built — running twelve model types across scenarios covering architecture, hardware, data, infrastructure, energy and labour — puts basic workflows at around $0.05 per inference token, summarisation and knowledge retrieval at roughly $0.10, more complex workflows near $0.30, and planning and learning tasks at about $0.40. Provider cost per token for planning and learning therefore runs eight to ten times that of basic workflows.

Read that as a design instruction rather than a statistic. The cost of an interaction is set by the class of reasoning you invoke, not by the vendor’s price list. Which means it is a decision your team makes, not a market condition you accept.

Related Article: 5 Contact Center AI Trends Reshaping Customer Service

Why Agentic AI Costs Hit Customer Service Hardest

I want to be specific about why this lands harder on us than on other functions, because the generic coverage of this forecast has not made the connection.

4 Forces Concentrating Agentic AI Costs in Customer Service

Editor’s note: The following table highlights the most important lessons, actions and strategic considerations emerging from the forces that make customer service the most cost-exposed function in agentic AI deployments.

Key AreaWhat HappenedWhy It MattersRecommended ActionVolumeLegal reviews a few hundred contracts. Service handles millions of contacts.A per-interaction cost multiplier that is trivial in a low-volume function is existential in a high-volume one. We are the largest consumer of interactions in the business, so we are the largest consumer of tokens.Tier your intents before you tier your models. Most service volume is retrieval, and retrieval does not need a reasoning agent.LoopsCustomer conversations are not single-shot. They involve clarification, retrieval across systems, policy checking, escalation and follow-up. Every loop is another round of reasoning.The messier the customer’s problem, the more expensive it becomes — which means your worst-served customers are also your most costly to serve. That inverts the logic the entire deflection case was built on.Set a reasoning budget per contact, with a hard stop. Cap the number of reasoning cycles per interaction and force a human handover when the cap is hit.Failure is not freeWhen a deterministic script fails, it fails cheaply. When a reasoning agent fails, it has already burned through its reasoning budget before handing over to a human who then costs full price.You pay twice for the same unresolved contact.Keep deterministic paths alive. A well-built deterministic journey for a high-volume, low-variance task is faster, cheaper and more predictable.Billing models are moving against usThe Register notes that the shift by some providers from flat-rate subscriptions to usage-based billing makes token-heavy workflows capable of producing runaway costs.Flat-rate contracts capped your exposure. Consumption pricing transfers that risk to you, and service is the function generating the consumption.Renegotiate before renewal, not after the invoice. Secure volume caps, alerting thresholds and a contractual position on who absorbs cost when the agent loops.

This is why I keep arguing that trust and economics are the same conversation rather than two separate ones, as I set out in The Economics of Trust in AI-Driven CX. A system that reasons expensively and resolves poorly is not merely a bad customer experience. It is a bad customer experience with a metered bill attached.

Related Article: The Next CX Leaders Will Build Systems, Not Just AI

The Token-Deflation Illusion: Why Cheaper Tokens Won’t Cut Agent Costs

There is a specific belief circulating in procurement conversations that needs killing, and the researchers name it directly.

Will Sommer and Sabine Zimmerhansl describe the market as captured by a token-deflation illusion — buyers assuming that as providers improve token economics, those savings will show up in their own roadmaps. Their assessment of that assumption is blunt: they will not.

The supporting projection makes the point sharper. Token costs are expected to fall by roughly 95% by 2030, and agentic workflow costs still rise more than fivefold in the meantime. If a 95% price reduction does not rescue your economics, no vendor negotiation will either.

So the strategic conclusion is uncomfortable but clarifying. Agentic AI does not benefit from economies of scale in the way software historically has. In classic software, the second million users cost almost nothing to serve. Here, the second million interactions cost roughly what the first million did — and more if they are harder. Every assumption about operating leverage that our industry imported from SaaS needs re-examining.

How Inference Tiering Controls Agentic AI Costs in Customer Service

None of this argues against agentic AI. It argues against deploying it indiscriminately, which is precisely what most organisations are doing. The recommended remedy is inference tiering, routing and orchestration — calibrating complex tasks against more cost-efficient intelligence. Sommer’s warning is that defaulting to generic autonomous intelligence produces unbounded costs orders of magnitude above an optimized ecosystem.

Here is how I would translate that into service operations.

Translating Inference Tiering Into Service Operations

Editor’s note: The following table highlights the most important lessons, actions and strategic considerations emerging from applying inference tiering, routing and orchestration to customer service operations.

Key AreaWhat HappenedWhy It MattersRecommended ActionIntent tieringMost service volume is retrieval.Retrieval does not need a reasoning agent, and paying agentic prices for a lookup is the single most common waste I expect to see in 2027 budgets.Tier your intents before you tier your models. Classify your contact reasons by the reasoning class each genuinely requires: retrieval, summarisation, multistep resolution, or open-ended planning.Task-based routingVendors sell one system that handles everything, because that is easier to sell.That architecture guarantees you pay planning-class prices for password resets.Route by task, not by product. Insist on the ability to route different intents to different models, and treat that capability as a commercial requirement rather than a technical nicety.Reasoning budgetsAgents that question themselves indefinitely will do exactly that.A hard stop protects the customer and the budget simultaneously, which is rare enough to be worth doing on its own.Set a reasoning budget per contact, with a hard stop. Cap the number of reasoning cycles per interaction and force a human handover when the cap is hit.MeasurementCost per contact is now meaningless without knowing which class of intelligence handled it.If you cannot produce that number today, that gap is your first project, and it takes weeks rather than quarters.Change the unit you measure. Move to cost per resolved contact, split by reasoning tier.Contract renewalBuyers are moving from subscription to consumption pricing.Your provider’s answer on failed and retried workflows will tell you a great deal about whether they have thought about this at all.Renegotiate before renewal, not after the invoice. Secure volume caps, alerting thresholds, and a contractual position on who absorbs cost when the agent loops. Ask your provider directly what happens to your bill when a workflow fails and retries.Deterministic pathsThe instinct to replace every scripted flow with an agent is expensive and often worse for the customer.A well-built deterministic journey for a high-volume, low-variance task is faster, cheaper and more predictable.Keep deterministic paths alive. Reserve reasoning for the situations that actually require judgement. The Reasoning Tier Question for Your Next Service Operating Review

The organizations that will do well here are not the ones spending least on AI. They are the ones who know which class of intelligence is answering which customer, and what each of those decisions costs.

That is a design discipline, and it belongs to service leadership rather than to procurement or engineering. Nobody else in the building knows which contact reasons genuinely need reasoning and which have been over-engineered because a demo looked impressive.

So one question for your next operating review, and I would put it in exactly these words. For our top 20 contact reasons, which reasoning tier is handling each one, what does it cost per resolved contact, and who approved that allocation?

If nobody can answer, you do not have an AI cost problem yet. You have about 18 months before you do.

fa-solid fa-hand-paper Learn how you can join our contributor community.