By Samir Sharma ~ with research contributions from Pulse Labs
A few weeks ago, I was deep in a debugging session with an AI coding tool.
Not a quick question. A real session. Iterative, stateful, building on itself. The kind where you are 30 to 45 minutes in, the model has context, you have momentum, and you can feel the solution taking shape. Then it stopped. I had run out of credits.
It did not feel like hitting a technical limit. It felt like my call dropping mid-sentence because I ran out of prepaid minutes. That is when the mismatch clicked: we are building AI tools for postpaid workflows, but pricing them like prepaid services.
The analogy is more accurate than it sounds.
Prepaid pricing makes sense when usage is predictable and discrete. You buy a block, consume it, top up when needed. That works for utilities. But AI-assisted work is not a utility interaction. When I am working inside a tool like this, I am not issuing isolated requests. I am working through a problem: refining logic, testing edge cases, building context across a chain of back-and-forth. It feels less like querying an API and more like thinking out loud with a sharp colleague. In that mode, the cost per prompt is invisible. I am not thinking about tokens. I am thinking about solving the problem.
When credits run out in the middle of that, it is not just a billing event. It is a cognitive interruption, and in a coding workflow, that interruption is especially costly. Context has been built across multiple interactions, momentum is high, and the solution is often just within reach. A hard cutoff does not pause progress. It resets it.
This is a product design problem, not just a pricing one.
| Dimension | Current Model | Actual Usage |
| Usage pattern | Linear and predictable | Bursty and iterative |
| User awareness | Monitoring credits | Focused on the task |
| Interruption | Hard cutoff | Should be seamless |
| Unit of value | Tokens consumed | Problems solved |
That last row is the real issue. We are pricing inputs while users experience value through outcomes. Nobody cares how many tokens it took to fix a bug. They care that the bug is fixed.
We have seen pricing evolve before. SaaS was priced per seat. Cloud was priced per compute and storage. AI is something different. We are no longer just delivering software functionality. We are augmenting how people think, decide, and build. The value is not in tokens consumed. It is in time saved, decisions improved, and problems that actually get solved. That shift changes what good pricing design should look like.
A few directions seem promising.
Hybrid subscription and usage. A base plan provides continuity, with usage tiers layered on top. This is how major gas and electric utilities structure residential pricing: your baseline allocation covers everyday usage at a standard rate, and heavier consumption moves into higher tiers. You are never cut off, just priced accordingly. The base keeps the relationship intact; the tiers handle the variable work.
Soft limits instead of hard cutoffs. This is already standard practice for most B2B SaaS platforms across Sales, Marketing Automation, Service, Billing, and ERP when managing API or usage limits. If limits are necessary, let the session complete. Notify the user, reduce performance if needed, but do not break the flow mid-thought. The trust damage from a hard stop is disproportionate to the marginal cost of finishing.
Session-based pricing. Charge for meaningful units of work: a debugging session, a code review, a feature spec. Not the underlying token math that users never think about. This is similar to how leading cloud serverless platforms work. You pay per function invocation and how long it runs, not for the raw compute cycles underneath. The bill reflects the task, not the infrastructure. That is the same shift AI pricing needs to make.
Outcome-oriented pricing over time. Eventually, pricing should reflect value delivered: time saved, complexity handled, business impact. This is harder to implement, but it is the model that most accurately matches how people experience value.
Some AI support platforms are already doing this, charging roughly $0.99 per resolved issue rather than per token or per conversation attempt. For context, a human agent typically costs $5 to $10 per query. The point is not who handles the issue. It is that when pricing is tied to resolution, the incentive finally aligns with what the customer actually wants.
That said, businesses need to stay flexible. As customer frustration with AI interactions grows and churn becomes a real concern in some segments, the ability to dial the model up or down based on what actually works for your customers is just as important as the pricing structure itself. That includes knowing when to put a human in the loop sooner rather than later. The best customer service is the kind that resolves the issue and keeps the customer, regardless of who or what handles it.
Why does this matter if you are building AI products?
All of this means that a flexible pricing model is not an afterthought. It needs to be architected into the product from the start, one that can accommodate AI led resolution, human handoff, and everything in between. That is the direction AI pricing needs to move toward across the board.
I saw this pattern play out first-hand at the enterprise technology companies I have worked for. The companies that won in SaaS were not always the ones with the best features. They were the ones whose pricing matched how customers actually worked. Most teams think about pricing as a revenue decision. But pricing also determines your sales motion, how sales reps are compensated, how marketing tells the story, and how a buyer navigates internal procurement to get the deal approved. A token-based model is a hard sell to a CFO who wants predictability. An outcome-based model, pay per resolution or pay per session completed, is a much easier conversation.
AI will follow the same arc. Right now, token-based pricing is putting handcuffs on how people use these tools. But the right model will do more than remove friction. When pricing reflects outcomes rather than inputs, the entire buying cycle unlocks. Marketing can tell a clear ROI story instead of explaining token math. Sales can close on business value instead of defending usage caps. Procurement and Finance can approve a model that looks like every other outcome-based contract they have signed. And end users can actually work the way they need to work, without watching a meter. The companies that figure this out first will not just win on product. They will win on how easy they are to buy.
About the Author
Samir Sharma is a product leader at Adobe building AI-powered systems for enterprise decision-making. With over 15 years of experience across SaaS software, cloud and data center platforms, and hardware chip design, he brings a full-stack perspective to product development. He has led 0-to-1 products at Salesforce and Cisco, with deep experience across sales, marketing, and service platforms. He holds an MBA from the University of North Carolina Kenan-Flagler Business School.
Research Perspective — Pulse Labs
It is the interruption. Product leaders describe the same dynamic across their user bases: their most engaged users are not running discrete queries — they are deep in extended, iterative sessions where the tool has effectively become part of their thinking process. When a hard limit breaks that session, the damage is disproportionate. It does not read as a billing event. It reads as the tool failing them at the worst possible moment.
The trust implication comes up repeatedly. Product and research leaders we speak with are increasingly clear that reliability — not just features — is what converts a tool from an experiment into infrastructure. This is not unique to AI. They have seen it play out in SaaS, in developer tooling, in data platforms. A product that performs brilliantly ninety percent of the time but cuts off unpredictably in the other ten, is not one their users will build core workflows around. The ceiling on adoption is set by the worst experience, not the average one.
Research leaders raise a related point about how this maps to long-term retention. The moment any kind of interruption pulls a user out of deep work, the relationship with the tool changes. It shifts from invisible partner to visible obstacle — and that shift, in their data and in ours, is one of the most reliable predictors of churn. Rebuilding that trust after it breaks is significantly harder than preserving it in the first place.
The clearest takeaway from these conversations: pricing is not a back-office decision. It is a product decision. And the leaders getting this right are the ones treating frictionless continuity not as a growth strategy, but as a fundamental design requirement.
Member discussion