June 25, 2026

The Bill You Can’t Budget: How AI Is Rewriting the Rules of CX Procurement

The seat-based SaaS model was, whatever its limitations, legible. A contact centre bought licences, counted users, and planned around headcount. Procurement teams knew the unit. Finance could model it. Vendors and buyers, however unequal their footing, at least shared a common language.

Agentic AI has deftly disrupted that arrangement. Today, the commercial unit is shifting from people to consumption. That means tokens, credits, actions, automated resolutions, and conversations. The invoice now reflects how much the software was used, not merely how many people had access to it. That may be a more honest accounting of how AI actually works. However, it also introduces something many CX leaders have spent years trying to eliminate: cost opacity.

“Charging for AI based on how much you use it isn’t automatically fairer,” says Tim Banting, Founder and Principal Analyst at So What? Now What!

“Really, it just moves the financial risk from the vendor straight onto the customer. The software companies know exactly how much computing power their models chew through, but the buyers are completely in the dark.”

What Buyers Are Actually Facing: The New Landscape of AI Pricing

The market has not settled on a single model, and that instability is itself part of the problem. Across the major CX and contact centre platforms, buyers now encounter a proliferating set of commercial structures that did not exist three years ago.

Salesforce’s Agentforce is perhaps the clearest illustration of how rapidly this space is evolving. The company has shipped three distinct pricing models for Agentforce in roughly 18 months. First, there was $2 per conversation at launch, then Flex Credits at $0.10 per action in May 2025. Most recently, per-user licences started at $125 per user per month. All three were running simultaneously on the same product. Rather than being a stable commercial framework, it’s a live experiment in which the customer bears the uncertainty.

Elsewhere, the picture is no less complex. Zendesk was among the first platforms in the CX space to offer outcome-based pricing, with committed automations priced at $1.50 per resolution. Amazon Connect runs pure consumption pricing with no minimum monthly fees or upfront licence charges. Hybrid pricing, blending traditional seat-based licences with variable credit pools, is an increasingly common structure. This is driven partly by the sheer difficulty buyers face in forecasting variable AI spend against fixed budget cycles.

The result for leaders and buying committees is a procurement environment of remarkable fragmentation. Comparing vendors on cost alone has become close to impossible when the underlying metre differs across every contract.

Why Vendors Are Moving This Way — And What It Costs Buyers

The commercial logic behind consumption-based pricing is not difficult to follow. AI incurs authentic variable costs at the infrastructure level: compute, inference, and tokens processed. Unlike a software licence, which can be sold at near-zero marginal cost, large language model deployments scale with usage. Vendors are, in a narrow sense, passing through real costs.

But the asymmetry of information in that arrangement is stark. Vendors hold detailed visibility into their own compute economics. Buyers hold virtually none. As Banting puts it:

“This completely messes up the way companies usually buy customer service software. For years, managers could plan their budgets around predictable things like the number of desks, software licences, or users logged in at the same time. AI throws all that predictability out the window.”

There is also a structural conflict at play that deserves more scrutiny than it typically receives. Legacy CX vendors built their revenue models on seat-based pricing. The more effective their AI becomes at deflecting contact volume and reducing agent dependency, the fewer seats their clients need. Naturally, this directly undermines the vendor’s own commercial position.

Outcome-based pricing resolves this tension by aligning vendor incentives with customer outcomes. However, it requires vendors to accept tangible performance risk, and many are reluctant to do so without significant contractual safeguards.

For organisations investing in AI to reduce costs, the paradox is tangible. The tool deployed to drive efficiency can simultaneously make platform spend harder to predict and control. A recent Salesforce study of CIOs found that 90% report managing AI costs is already limiting their ability to drive value. This is truly a striking figure given how early most organisations remain in agentic deployment.

Incentive Distortion, Governance Gaps, and the Benchmark Problem

The pricing structure an organisation chooses does not merely determine what it pays. It shapes how the entire operation behaves. This is why the choice deserves far more strategic attention than it currently receives.

Banting is direct on this point.

“The way you pay for AI actually changes how your call centre staff and systems behave. If a vendor charges you for every single task the bot does, your team will try to cut down on those tasks. If the bill is based on tokens, they will get obsessed with making prompts as short as possible, opening the door to potential hallucinations. But if you pay only when a problem actually gets fixed, the focus stays on doing a good job.”

This incentive distortion runs deeper than operational behaviour. It corrupts the benchmarking process. Tokens, the most granular consumption unit, are a measure of computational activity, not customer value. Two vendors resolving an identical customer query may consume wildly different token volumes. Comparing them on cost per token produces a number that is commercially misleading.

The governance challenge is equally pressing. Consumption-based pricing demands a level of real-time spend monitoring that traditional procurement processes were not designed to support. Purchasing departments accustomed to annual budget cycles now need to track usage against shifting denominators, set granular spending limits, and attribute costs to specific workflows and business units. None of that infrastructure is in place at most organisations.

The stakes are rising. Forrester’s 2025 CX Index fell to a new all-time low of 68.3 out of 100 in the United States, a fourth consecutive annual decline, with disappointing AI implementations cited as a contributing factor. CX leaders cannot afford to add cost unpredictability to an already deteriorating performance picture.

What CX Leaders Should Demand Before Signing

The pricing shift does not reduce to a simple directive to seek outcome-based contracts and avoid token-based ones. The reality is more nuanced. What constitutes a sensible commercial structure will vary by deployment type, use case maturity, and organisational risk appetite.

What does not vary is the need for greater contractual rigour. Banting outlines the baseline:

“You want to see the actual costs behind the tech, set strict limits on your spending, and have the option to pay only when a problem is solved. You also need to be able to track every single step the bot takes and make sure the vendor shares some of the risk if the system breaks down or does a terrible job.”

Several of these are, in principle, achievable today. Yet many organisations enter contracts without negotiating the usage dashboard access, ramp-period exclusions for test traffic, or true-down rights that would give them meaningful cost control in practice. The first 90 days of an agentic deployment routinely see consumption run well above forecast. Test traffic, internal QA sessions, and retry loops all register on the metre, but with buyers, not vendors, absorbing the difference.

Arguably, the most strategically urgent question is what the pricing model implies about the vendor’s own confidence in its product. A vendor willing to charge only on verified resolution is a vendor that believes in its automation quality. A vendor that insists on token or action-based billing regardless of outcome is, potentially, hedging its own performance risk at the buyer’s expense. That is a reasonable commercial position for the vendor to take. It is also important information for the buyer.

Banting puts the broader principle plainly:

“At the end of the day, if AI is going to take over jobs usually done by people, we have to track and manage its costs just like we do with regular staff budgets.”

Pricing as a Strategic Lens

The shift to consumption-based AI pricing will likely not reverse. Variable compute costs are structural, and aligning price to value delivered is a more defensible commercial model than the legacy seat licence, which charged for access regardless of use, rewarded shelfware. It also created no incentive for vendors to maximise the value their software delivered.

But that logic cuts both ways. The same move that makes pricing more rational for vendors also makes it more demanding for buyers. It requires procurement competence, governance infrastructure, and commercial literacy that most CX teams are still building. The crux of the issue is what the pricing structure reveals about vendor incentives, what governance infrastructure is required to manage it, and whether the commercial framework in place actually connects what the vendor earns to what the customer experiences.

Pricing, in short, has become a due diligence lens on AI. It is not an afterthought once the tech decision has been made. CX leaders who treat it as the latter will find themselves reading bills they cannot explain, for outcomes they cannot verify, on systems whose cost behaviour they cannot predict.

Customer Experience Today's pick