How to Price Your AI-First Product: The Death of SaaS Pricing and the Rise of Transactional Models with Defy Ventures’ Medha Agarwal

For years, SaaS pricing was comfort food for software companies: charge per seat, throw in an annual contract, add a discount big enough to make procurement feel heroic, and call it strategy. It worked because the product mostly helped people do work. More users meant more value, more seats, and more revenue. Nice, neat, and wonderfully spreadsheet-friendly.

AI-first products break that logic in half. In many cases, the software is no longer just helping a human click faster. It is writing, reviewing, resolving, classifying, drafting, summarizing, answering, routing, and sometimes finishing the job before a person even shows up. That changes the value story, the cost structure, the buyer conversation, and, most importantly, the pricing model.

That is the core of Medha Agarwal’s argument at Defy Ventures: when AI products complete work end-to-end, they can increasingly sell into labor budgets rather than software budgets. That is a huge shift. Labor budgets are bigger, stickier, and often tied more directly to business outcomes than classic software spend. In other words, the prize is larger, but the old per-seat playbook suddenly looks like bringing a butter knife to a machine shop.

Still, let’s not hold a funeral for SaaS pricing just yet. Seat-based pricing is not dead-dead. It still offers predictability, easier budgeting, and a cleaner cash flow profile. Public markets have long rewarded those qualities. What is dying is the idea that per-seat pricing should be the default for every software business, even when the software behaves more like a worker than a tool. The future belongs to founders who know when to charge for access, when to charge for usage, and when to charge for completed work.

The Old SaaS Formula Is Cracking

Traditional SaaS pricing made sense in a world where software amplified human effort. Think CRM, collaboration, project management, or analytics dashboards. The product’s value scaled with the number of people using it, so seats were a reasonable proxy for value. It was never perfect, but it was simple, predictable, and easy for finance teams to explain without breaking into a light sweat.

AI-first products do not always behave like that. If one AI agent can do the work of five people, your customer may get more value while adding fewer seats. That is a problem if your revenue still depends on seat count. The customer wins, your product works, and yet your pricing lags behind the value you created. Congratulations: you built a great product and an awkward monetization trap.

This is why transactional pricing is gaining momentum. Instead of charging for access to the software, you charge for the work the software performs. That work might be measured by tokens, documents, minutes, calls handled, tickets resolved, claims processed, leads qualified, or another unit of completed work. The charge metric moves closer to the actual value delivered.

That shift is also mirrored in the infrastructure stack underneath AI. Leading model providers and cloud platforms often bill based on usage, throughput, caching, batch processing, or reserved capacity. In plain English: your costs are already variable. If your revenue remains stubbornly fixed while usage explodes, your margins can start doing backflips for all the wrong reasons.

Why AI Economics Push Founders Toward Transactional Pricing

Variable Cost Structures Are No Longer a Side Note

In classic SaaS, cost of goods sold was usually predictable enough that founders could charge flat subscriptions without losing sleep. AI changes that math. Model usage, context length, multimodal inputs, latency requirements, and orchestration layers can all influence cost. OpenAI, Anthropic, AWS, Google Cloud, and Microsoft all support some version of metered or throughput-based pricing, which tells you something important: AI economics are not naturally seat-shaped.

That does not mean every AI company should expose raw token pricing to customers. Quite the opposite. Most customers do not want a finance puzzle disguised as a pricing page. But founders do need to respect the fact that AI cost curves can move with usage, performance tier, and reliability requirements. Ignore that, and your “simple” pricing may quietly turn into a very sophisticated way to lose money.

Customers Care About Finished Work, Not Model Calories

A buyer rarely wakes up excited to purchase 17 million tokens. They want invoices extracted, charts generated, conversations summarized, tickets resolved, safety reviews completed, or outbound campaigns launched. The closer your price maps to that outcome, the easier it is for the buyer to defend the spend internally.

This is where AI pricing starts looking less like traditional software pricing and more like operational pricing. A customer support leader can understand cost per resolved ticket. A revenue operations team can understand cost per qualified meeting. A healthcare operations team can understand cost per claim reviewed. Those are business units, not technical units, and business units tend to unlock bigger budgets.

AI Is Competing with Labor, Not Just Other Software

Medha Agarwal’s most useful lens is the budget lens. If your product replaces or materially reduces human work, you are not only competing with other software vendors. You are competing with headcount, outsourcing, back-office process costs, and the slow, expensive, deeply human miracle known as “manual work.” That means your pricing power may be based on labor savings, not software comparisons.

That also explains why some AI-first founders can justify charging far more than a normal SaaS benchmark would suggest. When the comparison is “software license versus software license,” budgets stay relatively tight. When the comparison is “this platform versus a team of people doing repetitive work,” the value conversation gets much bigger, much faster.

The Four Pricing Models That Matter in AI-First Software

1. Fixed Subscription Pricing

This is the classic SaaS model: per seat, per team, per location, or platform fee. It still works well when usage is frequent, habitual, and hard for users to mentally meter. If a product is used dozens of times per day, charging per action can create friction. Nobody wants to wonder whether sending one more message, generating one more note, or asking one more question will trigger a miniature billing crisis.

Fixed pricing is strongest when the end user is a human, adoption matters more than exact value capture, and predictability is a major part of the sale. It is clean, familiar, and budget-friendly. It is also the pricing equivalent of sweatpants: comfortable, popular, and not always appropriate for a board meeting.

2. Input-Based Transactional Pricing

This model charges for the resources consumed. Think per token, per API call, per document, per minute, per file, or per workflow run. It works especially well for developer tools, infrastructure products, and AI services where the buyer understands usage as part of the product itself.

Input-based pricing is often the easiest model to launch because it maps neatly to your internal cost structure. The downside is that it can feel too technical for nontechnical buyers. If your customer is an operations leader, billing by tokens may be accurate but emotionally useless. It is like billing someone for the number of electric sparks in their toaster.

3. Output- or Outcome-Based Transactional Pricing

This is where AI pricing gets interesting. Instead of charging for what goes in, you charge for what gets done. A support AI could charge per resolved ticket. A legal AI could charge per contract reviewed. A recruiting AI could charge per qualified candidate delivered. A finance AI could charge per invoice reconciled.

Outcome-based pricing is powerful because it aligns price with customer value. It also requires confidence. You need strong measurement, clear definitions, auditability, and enough product maturity to stand behind the result. Founders love outcome pricing in theory. Finance teams love it after the instrumentation is finally built and everyone has stopped arguing about what “resolved” actually means.

4. Hybrid Pricing

Hybrid pricing is rapidly becoming the practical winner. Bain has described hybrid models as the dominant interim strategy in AI software, and that makes perfect sense. A hybrid structure usually combines a base subscription or platform fee with included usage, minimum commitments, volume tiers, overages, or success-based components.

Hybrid pricing gives customers predictability and gives vendors a way to capture upside. It is the compromise model, but in a good way. Think of it as the grown-up answer to a messy market: enough stability for procurement, enough flexibility for finance, and enough upside for founders who do not want their best customers capped by seat count.

Medha Agarwal’s Four-Part Test for Choosing the Right Model

One of the most actionable frameworks from Medha Agarwal’s pricing discussion is the idea that founders should choose a model based on four factors, not vibes. Pricing by intuition alone is how you end up with a dashboard full of “strong engagement” and a revenue plan held together by optimism.

Frequency of Usage

If the product is used constantly throughout the day, fixed pricing usually reduces friction. High-frequency usage and pure pay-per-use pricing can be a bad combination because customers start mentally calculating the cost of every action. That is poison for adoption. If usage is occasional, episodic, or tied to variable workloads, transactional pricing becomes much more attractive.

Magnitude of Cost Savings

If your AI creates dramatic cost savings compared with a human workflow, you have room to price more aggressively on a per-transaction basis. If the savings are modest or diffuse, customers may prefer a fixed fee because the ROI is harder to attribute per use.

Workflow Integration Point

Where your product sits in the customer’s workflow matters more than many founders realize. If all inbound volume runs through your system first, usage can be predictable and easy to model. If your product is only used sometimes, at the edge of the workflow, a pure transactional model can feel random and harder to budget.

Budget Type and Buyer Persona

This is the big one. Are you selling to IT out of a software budget, or to operations and finance out of a labor or process budget? If you are replacing labor, outcome and transaction pricing can feel natural. If you are simply adding a new tool to the stack, subscription pricing may still be the better fit. The buyer’s mental model should guide the pricing model. If they think in licenses, show licenses. If they think in units of work, show units of work.

How to Build a Smarter AI Pricing Strategy

Hide the Complexity, Not the Logic

Customers do not need to see every moving part of your cost stack. They do need to understand why the pricing is fair. The best AI-first pricing pages explain the value metric clearly, keep the number of variables under control, and avoid turning checkout into advanced calculus.

Separate Your Cost Metric from Your Customer Metric

Your model vendor may charge you for tokens, caching, or throughput. That does not mean your customer should be billed the same way. Internal cost metrics are for finance and product. Customer-facing metrics should reflect business value. A buyer wants to understand what they bought, not reverse-engineer your infrastructure bill.

Use Guardrails for Predictability

If you choose usage-based or outcome pricing, add guardrails. Minimum commitments, included usage bands, spend caps, annual true-ups, and volume discounts help customers feel safe. This is also why reserved or provisioned capacity models exist across the AI infrastructure ecosystem: predictability matters, especially at scale.

Price at Par and Win on Value

Agarwal is blunt on this point, and she is right. Undercutting competitors can attract the wrong customers and create a false sense of product-market fit. Price-sensitive customers are often temporary customers. They love your discounts right up until somebody else shows up with shinier discounts. Instead of racing to the bottom, compete on accuracy, workflow fit, trust, auditability, service, and measurable ROI.

Common Pricing Mistakes AI Founders Keep Making

• Charging by token to buyers who do not care about tokens.
• Keeping a seat-based model long after the product is doing autonomous work.
• Offering usage pricing with no forecasting tools, making finance teams nervous.
• Hiding overages in the fine print and then acting surprised when renewal calls get weird.
• Launching outcome pricing without clear instrumentation, definitions, or success criteria.
• Assuming cheaper pricing is a substitute for sharper positioning.

What Good AI Pricing Looks Like in Practice

A solid AI-first pricing model usually has three qualities. First, the customer can explain it in one sentence to their CFO. Second, the vendor can protect margin as usage scales. Third, the price grows with value without making customers feel punished for success.

That often leads to structures such as a platform fee plus usage tiers, a workflow subscription with overages above a threshold, or a per-outcome fee with monthly minimums. In other words, the future is not simply “usage-based pricing.” The future is pricing that balances fairness, clarity, predictability, and upside. Founders who understand all four can turn pricing from a billing function into a strategic weapon.

Experience from the Field: What Founders Learn After Their First AI Pricing Reset

Once founders move from theory to real-world pricing, a few patterns show up again and again. The first surprise is that customers are often more open to transactional pricing than internal teams expect. Founders worry buyers will reject variable pricing on principle, but many buyers are perfectly comfortable with it when the value metric is concrete. They do not mind paying per claim reviewed, per conversation handled, or per report generated. What they hate is ambiguity. If they cannot estimate cost or explain what triggers spend, confidence drops fast.

The second surprise is that the biggest pricing challenge is usually operational, not philosophical. It sounds exciting to say, “We charge for outcomes.” Then someone asks what counts as an outcome, when it is measured, how disputes are handled, what happens when a human intervenes halfway through, and whether the system gets credit for a draft that a manager edits before sending. Suddenly the pricing meeting turns into a taxonomy workshop with light emotional damage. This is why strong telemetry matters. Outcome pricing is not just a packaging decision; it is a product and data decision.

Another common lesson is that hybrid pricing calms the room. Sales teams like it because they can anchor deals with a base commitment. Finance teams like it because revenue becomes easier to forecast. Customers like it because they get enough predictability to budget, plus room to scale if the product proves itself. Founders like it because they stop having the same philosophical debate every week and can get back to building. In many companies, hybrid pricing is not a compromise because the team lacks courage. It is a compromise because reality exists.

Founders also learn that different buyer personas hear the same price in very different ways. A technical buyer may be comfortable with usage, rate cards, and capacity language. An operations buyer usually wants to know how much work gets done and what savings or speed gains come with it. A CFO wants predictability, risk controls, and a believable ROI story. The product may be identical, but the pricing narrative has to translate across audiences. Great pricing is partly arithmetic and partly interpretation.

Then there is the emotional lesson nobody puts on the pricing slide. Raising prices, changing metrics, or introducing overages feels scary. Founders worry they will kill momentum. Sometimes they do it too late because a weak pricing model can hide behind strong growth for a while. But once the company starts seeing heavy usage, margin pressure, or mismatched expansion, the cost of waiting becomes obvious. In practice, the healthiest teams treat pricing as a product surface that evolves with customer behavior. They test, learn, and refine it. They do not carve it into stone after one sales call and hope destiny handles the rest.

The final lesson is the simplest one: customers will pay more when the product clearly does more. Not when the homepage says “AI-powered.” Not when the demo is flashy. Not when the founder uses the phrase “redefining workflows” with a heroic facial expression. They pay more when the software saves headcount hours, reduces errors, shortens cycle times, increases throughput, or improves conversion in a way they can feel. That is the real center of gravity in AI-first pricing. The model matters. The metrics matter. But value clarity is still king.

Conclusion

The death of SaaS pricing is not the death of subscriptions. It is the death of lazy default pricing. AI-first companies live in a new world where software can perform work, not just support it. That means founders need pricing models that reflect units of work, customer outcomes, and variable economics under the hood.

Medha Agarwal’s framework is a smart place to start: look at usage frequency, cost savings, workflow position, and budget type. Then choose the model that best matches how your product creates value. In some cases, that will still be seats. In many others, it will be transactional or outcome-based pricing. And for a growing number of companies, the real answer will be hybrid pricing that gives buyers predictability while letting vendors capture upside as value scales.

The winners in AI-first software will not just build better models or better agents. They will build better business models. In this market, pricing is not the boring slide at the end of the deck. It is part of the product. Treat it that way.