Pricing is the part of an AI pivot nobody wants to talk about in public, mostly because there's no clean answer yet. Most teams are still testing, adjusting, and occasionally reversing decisions a few months in. That's not a failure of planning. It's the actual state of the market right now.
This article looks at how two SaaS CEOs priced their AI features in practice, what the wider market data says about where pricing is heading, and what that means if you're about to make this call yourself. The on-the-ground experience comes from Valentin Huang, CEO of Harvestr, and Charles Miglietti, CEO of Toucan, speaking during a live webinar on AI product pivots.
TL;DR
- Seat-based pricing breaks down for AI features because the cost behind them (tokens, compute, API calls) doesn't scale with headcount, it scales with usage.
- Harvestr moved from plan-based to usage-based pricing gradually, testing willingness to pay on its highest plans first before rolling changes out further.
- Toucan kept its core plan structure but layered usage-based pricing for AI on top, while migrating existing customers progressively rather than all at once.
- Industry-wide, 42% of SaaS companies were already monetizing AI through usage-based or hybrid pricing as of the most recent benchmark, and Gartner projects 70% of leading vendors will offer consumption-based pricing by 2027.
- Hybrid pricing (a base subscription plus a usage layer) is the dominant real-world pattern, not pure consumption billing. Pick the dose of usage pricing that matches how unpredictable your AI costs actually are.
- Billing infrastructure is the unglamorous blocker most teams underestimate. Usage-based pricing needs metering and tracking that flat or seat-based plans never required.
Why seat-based pricing doesn't fit AI features
For most of the last decade, SaaS pricing followed a predictable formula: pick a tier, pay per seat, add users as the team grows. That model worked because the cost of serving one more user was close to zero. A CRM doesn't get meaningfully more expensive to run because you added a fifth sales rep.
AI features break that assumption. Every AI-powered interaction has a real, variable cost behind it, tokens processed, model calls made, compute consumed, and that cost doesn't track headcount at all. A team of five people using an AI feature heavily can cost more to serve than a team of fifty using it occasionally. Charging everyone the same flat seat fee means either overcharging light users or quietly eating margin on heavy ones.
This is the exact tension that pushed both Toucan and Harvestr to rethink pricing mid-pivot. Neither team set out to redesign their pricing model. The AI features forced the question.
What Harvestr actually did, step by step
Valentin Huang described a sequence that was deliberately incremental rather than a single repricing event. "We started by moving AI features from just plan-based to usage-based pricing progressively, because of the cost it's triggering for us and also because of the value we deliver to customers."
The first move was narrow on purpose: AI features only went onto the highest existing plans. That wasn't a packaging decision so much as a research one. "We first started to add our AI features on the highest plans to really test the readiness to pay, and see how much real value customers were getting out of it." Only after watching usage and gathering market feedback did the team widen the rollout and adjust the structure.
The core plan architecture stayed mostly intact. Harvestr still prices per seat with feature-gated tiers. What changed is that a usage-based layer for AI now sits on top of that structure, rather than replacing it. That detail matters more than it sounds: a full switch to consumption-based billing is a much bigger lift, both for customers to understand and for the team to bill correctly.
One change was more structural. "We dropped some legacy feature gates, features that were only available in certain plans but weren't AI-based, and because customers were getting so much value out of AI features, it didn't make sense anymore to price those features separately." Harvestr also retired its premium plan entirely. "We had a premium plan, and with AI and the value we're bringing and the cost behind it, it didn't make sense anymore to have this premium plan."
What Toucan did differently, and why
Charles Miglietti's team made a similar bet but moved more cautiously on the customer-facing side. "The pricing of the new product is different because it's mostly usage-based on AI, and the tiers are different. For existing customers, we are approaching it progressively."
The progressive part wasn't only about pricing mechanics. It was tied to the broader migration story: Toucan's legacy product is still maintained, not sunset, while the AI-first product is positioned as where the company is investing going forward. That meant pricing had to support two product lines running in parallel rather than forcing every customer onto a new pricing model on a fixed date.
The detail both founders flagged that most teams miss
Toward the end of the pricing discussion, Valentin raised something that gets skipped in most pricing strategy conversations: the billing infrastructure itself. "The last thing I wanted to add is also to anticipate the impact it has on your billing infrastructure and tooling, because usage-based pricing is not something that's always easy to implement and to track. That's also something you need to consider when you start pricing usage."
This is the unglamorous part of an AI pricing decision. Flat and seat-based billing is simple to meter: you count users, you invoice monthly. Usage-based billing requires tracking consumption in near real time, reconciling it against a plan structure, handling overages, and surfacing usage data back to customers so they understand their own bill. None of that exists by default in most billing stacks built for traditional SaaS. Budgeting for that build, or for the vendor integration that handles it, needs to happen at the same time as the pricing decision, not after the pricing page is already live.
There's a second observation that's easy to miss but worth sitting with. Alim, who moderated the conversation, summarized it well: "pricing is always in iteration." Neither founder described their current pricing as final. Both described a continuous process of adjusting tiers, watching how customers react, and changing course based on real usage data rather than a one-time launch decision.
What the broader market data says
Toucan and Harvestr's experience isn't an outlier. It tracks closely with what's happening across SaaS more broadly as AI features become standard rather than novel.
The most concrete data point comes from the 2025 SaaS Benchmarks Report from High Alpha, cited in recent SaaS pricing research, which found that 42% of companies were already monetizing AI features through usage-based or hybrid models. The same research points to Gartner projecting that 70% of leading SaaS vendors will offer consumption-based pricing across at least part of their portfolio by 2027. That's not a niche shift. It's a majority of the market moving in the same direction within a few years.
What's notable is that pure usage-based billing isn't actually winning out over hybrid models. Research from OpenView Partners, referenced in a 2026 SaaS pricing analysis, found hybrid pricing structures, a base subscription combined with a usage-based component, represent 46% of SaaS companies, well ahead of pure pay-as-you-go models at 15%. That distinction maps almost exactly onto what both Toucan and Harvestr actually built: keep the base plan structure customers already understand, add a usage layer specifically for AI on top of it.
Choosing how much usage pricing you actually need
Not every AI feature needs the same amount of consumption-based pricing layered on top. The right amount depends on how unpredictable your underlying AI costs actually are, and how that unpredictability scales with customer usage.
Low cost variability: if your AI feature has a roughly fixed cost per user regardless of how heavily they use it (a single daily summary, for instance), a flat add-on fee per seat can work fine. You're not exposed to runaway costs from a handful of power users.
Moderate variability: this is where Harvestr and Toucan both landed. A base plan structure stays in place, with a usage-based layer specifically for the AI features whose cost scales with how much customers actually use them. This protects margin on heavy users without forcing every customer to learn an entirely new billing model.
High variability: if your AI feature's cost can swing wildly between customers (an agent that might run a handful of tasks or thousands, for example), a closer-to-pure consumption model, credits or metered usage, becomes necessary. The tradeoff is that this is harder for customers to predict and budget for, and it's the version that most strains existing billing infrastructure.
Most teams don't need to pick a single archetype forever. Both founders described their pricing as something they're still actively adjusting based on what they're seeing in usage data, not a decision made once and left alone.
A practical sequence if you're about to do this
Based on what worked for both teams, a reasonable order looks like this.
Start narrow. Add AI pricing to your highest tier or most engaged customer segment first, the way Harvestr did, rather than repricing your entire base at once. This gives you real willingness-to-pay data before you commit to a structure broadly.
Decide what stays and what gets dropped. If AI value makes certain legacy feature gates irrelevant, as Harvestr found with features that no longer justified being locked behind a separate tier, removing them simplifies your pricing page and your sales conversations at the same time.
Build the billing infrastructure before you need it, not after. Usage tracking, overage handling, and customer-facing usage dashboards take real engineering time. Underestimating this is the single most common practical mistake based on what both founders described.
Treat the first version as a draft, not a launch. Plan to revisit pricing on a real cadence, informed by usage data, rather than treating the initial structure as something you'll only touch again if something breaks.
Where Toucan fits into this
If you're an ISV embedding AI analytics into your own product, pricing has an extra layer of complexity Toucan and Harvestr didn't have to deal with directly: you're not just pricing AI for your own customers, you're potentially helping your customers figure out how to price AI for theirs.
Toucan's usage-based AI pricing for embedded use cases is built around the same logic both founders described: a predictable base structure, with a usage layer specifically for the AI-driven queries and conversational analytics that actually consume compute. That keeps the model simple enough for an ISV to explain to its own end customers without needing to build custom metering from scratch.
If you're trying to figure out how AI pricing should work for your own embedded analytics, book a demo and we'll walk through how the usage model actually works in practice. Want to look at it yourself first? start a free trial.
Go deeper
- How to Generate New Revenue Streams with Data in a SaaS Product How analytics and AI features translate into monetization, beyond just the pricing model itself.
- How to Calculate ROI for Embedded Analytics A framework for justifying the investment in AI and analytics features before you decide how to price them.
- Average Revenue per User (ARPU): How to Calculate and Optimize It Useful context for tracking whether your new AI pricing model is actually moving the metrics that matter.
- Analytics Solution: Should You Build or Buy? The build vs. buy decision, which has its own cost and pricing implications for ISVs.
FAQ
Should I price AI features per seat or based on usage?
For most SaaS products, neither extreme works well on its own. Pure seat-based pricing ignores the real, variable cost behind AI features and risks losing money on heavy users while overcharging light ones. Pure usage-based pricing is harder for customers to predict and puts more strain on your billing infrastructure than most teams are ready for. The pattern both Toucan and Harvestr converged on, and what market data shows as the dominant real-world structure, is hybrid: keep your existing seat or tier-based plan as the base, and add a usage-based layer specifically for the AI features whose cost actually scales with consumption.
When should I start charging for AI features instead of including them for free?
Both Harvestr and Toucan introduced AI pricing progressively rather than all at once. Harvestr's approach was to add AI features to its highest plans first, specifically to test willingness to pay and observe real usage before deciding on a broader pricing structure. That sequencing matters: charging too early, before you understand how customers actually use the feature, risks setting a price that doesn't reflect real value. Waiting too long to monetize AI features that are clearly driving usage and retention leaves revenue on the table and makes a later pricing change feel more disruptive to customers who got used to it for free.
How do I know if my AI feature needs usage-based pricing or a simple flat fee?
The deciding factor is cost variability, how much the underlying cost of serving the feature changes depending on how heavily a specific customer uses it. If the cost per user is roughly fixed regardless of usage intensity, a flat add-on fee is simpler for everyone and avoids unnecessary billing complexity. If usage can vary dramatically between customers, the way an AI agent's compute cost might vary based on how many tasks it runs, a usage-based or hybrid model protects your margin and more fairly reflects the value different customers are actually getting.
What's the biggest mistake teams make when pricing AI features?
Underestimating the billing infrastructure required to support usage-based pricing. Flat and seat-based pricing is simple to bill: count users, send an invoice. Usage-based pricing requires real-time or near-real-time tracking of consumption, a system for handling overages, and a way to show customers their own usage so the bill doesn't feel like a surprise. This work needs to start at the same time as the pricing decision itself, not after the new pricing page has already gone live. Teams that treat billing infrastructure as an afterthought tend to find out the hard way, usually through support tickets about confusing invoices.
Is usage-based pricing actually replacing seat-based pricing in SaaS?
Not entirely, and that's an important nuance. Industry research consistently shows hybrid models, a base subscription plus a usage-based component, as the dominant pattern, well ahead of pure consumption-based billing. Most SaaS companies aren't abandoning seats outright. They're layering usage-based pricing for AI-specific features on top of a pricing structure customers already understand. That mirrors exactly what both Toucan and Harvestr did: the core plan architecture stayed largely intact, with AI consumption priced as an addition rather than a wholesale replacement.
