AI Pricing Models: Value-Based Pricing Under Scrutiny

AI pricing: "Cost per outcome" isn't the silver bullet it seems There's a growing consensus that AI changes everything about software pricing. The answer everyone's converging on? Cost per outcome. The logic seems intuitive - it should correlate with customer value, as well as the provider’s cost structure, where AI model expenses are a key component. We tried to use this for our Variation Intelligence product, where we process subcontractor claims for large contractors in construction. In practice, it just didn't work, and there are a number of overlooked downsides. 1. Outcome pricing is hard to predict. With fixed deals or seat pricing, finance teams know what they're spending. Per-outcome pricing makes spend much more unpredictable for enterprise customers. In our case, our customer couldn't tell us how many variations they'd process in a year, and their finance team could not sign off on an unknown cost. 2. Not all outcomes are equal. Per-outcome pricing assumes each action carries roughly equivalent value. In practice, that's often not true. In our case, some variations were worth a few hundred pounds, others hundreds of thousands, yet per outcome pricing treated them identically. 3. AI inference costs are dropping fast. Per outcome pricing used to be justified by really high inference cost, but these are dropping. This might make it hard to justify per outcome pricing in the long term, unless you drive down your prices as well. Per-outcome pricing is a genuinely useful new tool in the pricing toolkit. But it's not the universal answer to AI pricing that the current narrative suggests. The companies that get pricing right will match the model to the customer's reality.

Clay just changed their pricing and wow, AI pricing is so complicated. They absolutely nailed the communication. Seriously, watch their announcement video. Highly transparent. Really well done. And overall, I’d expect the thing they were trying to prevent (existing users nailing them to the Reddit cross) was avoided. They raised some prices, decreased others and aren’t forcing old users on to new pricing. However, their actual pricing got more complex. They introduced something called "actions." Every time you do some things in the product, it costs you (on top of your subscription, on top of data credits). The behavioral science analogy for this is: if you're eating pizza and you're paying per bite, you're going to eat less pizza. This is the pain of paying. When people feel the cost at the moment of use, they use less. And when you're trying to get someone to adopt your product, this is the opposite of what you want. I saw this idea come alive in my own onboarding. They gave me the choice to enrich 10 vs 1000 emails. I chose 10 because I was nervous about using too many tokens. The onboarding is training me to be conservative.  And instead of thinking how useful the product is, I was doing math. You know a relationship is going south if you’re calculating all the things your partner does and doesn’t do around the house. I expect the same with products -- this pricing change for Clay will increase their support ticket volume from angry spouses who feel the relationship is unfair :) Overall, AI pricing is hard. And in order to simplify (hide the cost complexity from users) it takes some guesswork. The question is do you put the guesswork and math on the user or take it on yourself? Clay seems to be opting for heavy transparency (cool!), but that also translates to asking the user to do more math. The ideal: Pricing should increase usage vs decrease it. The mechanism I like is the ability to buy packages after you hit the plan limit -- the idea here is to hide  the complexity for 90% of users and for the heavy users, push them to buy more credits (instead of asking all users to be meticulous trackers w/ gas meters). I do believe AI pricing will get simpler, the same way AOL went from charging per minute of dial-up to a $20 flat fee. We're just not there yet -- instead, it’s getting more complex! Companies have to charge by the model they use, the effort the model provides and now actions and data. Oh my!

AI agents are already making purchase decisions on behalf of customers. But most commerce platforms weren't designed for this. They were built around human behavior. A search query here. A few clicks there. Maybe a product comparison. AI agents operate completely differently. They execute rapid, multi-step queries. They pull structured data at scale. They expect real-time responses across every interaction. The load this creates isn't incremental. It's a different infrastructure challenge entirely. And here's the issue most companies are missing: The businesses that fall behind won't be the ones ignoring AI. They'll be the ones assuming their current systems can handle the volume when it arrives. That assumption is where the gap opens. This blog from commercetools looks at what this shift actually means for commerce infrastructure. It breaks down the six things enterprises need to have in place before the pressure builds. Worth reading if your team is starting to have these conversations internally. https://okt.to/cZCSKF

I tried to estimate how much of the world's code is being written by AI. There's no single source that measures this cleanly, but the trajectory probably looks something like this. Even the skeptics would agree a version of this is where we're heading. What's more interesting to me is that words are probably following the same curve, just a year or two behind code. The exact numbers are debatable, but I think most people in this space would agree the direction is right. But the two lines tell very different stories. When AI writes more code, developers don't slow down. They become builders who can ship more things. Because the appetite for what they produce is still far greater than what they can actually deliver. Every company I visit has software on their wishlist they've never had the budget or people to build. That backlog doesn't shrink when development gets faster. It expands. Demand for 'code' is essentially infinite. 'Words' are a different thing entirely. Last month I was at a wholesaler outside Rotterdam. Their sales team spends half their day writing quotes. Product specs, pricing, terms, follow up emails. We automated most of it. They didn't suddenly have five times more customers requesting quotes. They had the same pipeline, handled faster, with fewer people writing. The quotes are what they are. Same pattern everywhere. When a lawyer reviews contracts 5x faster, there aren't 5x more contracts waiting. When a support team handles tickets in a fraction of the time, the ticket volume stays the same. Code creates demand. Words absorb compression. That's the split nobody is framing clearly enough. The question isn't which tasks AI can do. It's whether the demand for the output has a ceiling or not. For software, there's no ceiling. Companies have decades of unbuilt tools sitting in someone's head or on a whiteboard somewhere. AI just makes the list longer. For most professional text, there is one. The same work gets done with fewer people. And that's a harder conversation than most AI optimists want to have.

AI may be the most expensive “free feature” in software history. Everyone is adding AI. Few are recalculating their unit economics. In traditional SaaS: More usage = more value = more profit. In AI: More usage = more compute = more cost. If you’re not measuring contribution margin at the customer level, you may be celebrating the exact cohort compressing your gross margin. Blended gross margin hides the truth. Customer-level margin exposes it. The real AI advantage won’t come from better prompts. It will come from: • Cost-to-serve visibility • Contract enforcement • Overage discipline • Margin intelligence built into billing AI monetization isn’t a pricing tweak. It’s an infrastructure decision. #AIMonetization #CFO #RevenueIntelligence #UsageBasedBilling https://lnkd.in/e4BWb7Rk

1% of people will have self-recursive AI agents in 12 months. Meaning it can call itself, evaluate its own output, correct its own mistakes, and loop back through its own process without human intervention. This will reduce the need for SaaS (but not kill it) Markets will respond. Pricing will shift. And the people with these agents will operate at a completely different speed. I won't be in that 1%. Let me be honest about that. But I will be building toward it. Slowly. Deliberately. Learning what these systems can do, where they break, and how to make them useful for real work. You should be doing the same. There's going to be a gap. People who have AI agents handling their workflows, researching, executing, and iterating on their own. And people who are still clicking through 15 SaaS dashboards to do the same job. The gap won't show up overnight. But in 12 months it will be visible. In 24 months, it will be obvious. You don't need to be in the 1%. You need to be building in the right direction. Start small. Pick one workflow. Automate it. Learn how agents pass information. Learn where they fail. Because they will fail. A lot. But the people who started building now will be ready when the tools catch up. And the people who waited will be scrambling.

Your AI analytics vendor just announced a 40% price increase. Your first thought: "We'll switch to a competitor." Your second thought, after talking to IT: "Actually, we can't." Welcome to the subscription trap. Here's what lock-in actually looks like: Your historical data lives in their proprietary format. Extracting it costs $15,000-$30,000 in consulting fees... assuming they even allow full exports. Your AI-generated insights are built on the vendor's specific prompting framework. New platform means rebuilding every query from scratch. Your integrations connect through their APIs. Switching means re-engineering every connection point. Your executives finally learned how to ask questions in the old system. New vendor means starting adoption from zero. As one director put it: "We're not customers. We're hostages with a monthly payment plan." So let's do the maths on that 40% increase. 5-year subscription model: $197,790 5-year ownership model: $47,000 That's four times more expensive. And that assumes only 10% annual increases going forward — not the 40% jumps we're seeing today. The alternative? Stop renting AI at runtime. Buy AI at setup time. AI generates the templates and logic once. You pay for tokens during the build. Then deterministic templates run forever with zero ongoing AI costs. No monthly fees. No vendor lock-in. No hostage situation. Your 40% price increase just became your wake-up call. What's your current AI analytics subscription costing you?

AI agents are already shopping for your customers, but many commerce platforms just weren't made for it. They were built around how humans usually browse—click here, search there. AI agents work in a whole different way. They do quick, multi-layered queries, gather structured data at scale, and want instant responses. This puts a whole new strain on backend systems! The companies that will fall behind aren't those ignoring AI; it's those thinking their current setups can handle the incoming demand. This blog dives into what this transition means for commerce systems and outlines six key things businesses should have prepared before the demand peaks. Definitely worth a read if your team is starting to discuss this topic: https://okt.to/EkNiPm

Everyone's ranking which SaaS tools are "dead" because of AI. Most of these lists are wrong. I've shipped 50+ production agents. Here's what they miss. 𝗧𝗵𝗲 𝗯𝗿𝘂𝘁𝗮𝗹 𝘁𝗿𝘂𝘁𝗵 𝗮𝗯𝗼𝘂𝘁 𝘀𝘂𝗿𝘃𝗶𝘃𝗮𝗹: The tools that die aren't the ones AI can replace. They're the ones that don't become the AI. Cursor is on the "probably dead" list. I think that's backwards. Cursor isn't competing with AI. Cursor is becoming the interface for AI. 𝗪𝗵𝗮𝘁 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗸𝗶𝗹𝗹𝘀 𝗮 𝘁𝗼𝗼𝗹: → Single-feature products with no workflow depth → Tools that sit between you and the AI instead of embedding it → Platforms that treat AI as an add-on instead of the core Cursor's bet: become the IDE where agents live. Not just autocomplete. Full workflow automation. Background agents that refactor, test, deploy. That's not a tool getting replaced. That's a tool becoming the replacement. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝘁𝗶𝗲𝗿 𝗹𝗶𝘀𝘁: → Dead: Single-feature tools AI can replicate in one prompt → Surviving: Platforms embedding AI as a feature → Winning: Tools becoming the orchestration layer for agents Cursor has a path to that third tier. Most tools on these lists don't. Which tools do you think are being underestimated?

Everyone is building AI tools. Few are solving problems they actually have. Three questions should be mandatory: Does this already exist? The AI SaaS graveyard is filling up fast. Point solutions with a thin AI wrapper on top of capabilities that established platforms already ship natively. If you work in the tool daily, you already know what’s missing. You also know what’s already there. Shipping something new doesn’t matter if the gap you’re filling closed six months ago. What problem is this actually solving? Not a hypothetical user persona. Your problem. The one costing you three hours a week, the one your team works around with a spreadsheet and a prayer. “AI makes it faster” is a feature. Cutting a manual reporting process from four hours to twenty minutes is a value proposition. What’s the business case? Start with your own workflow. If a tool saves you five hours a week across a team of ten, that’s a quantifiable ROI before you’ve written a single line of code. That’s your business case. That’s also your first case study. The market research phase that used to take months is now an afternoon. There’s no excuse for skipping it. Build for the problem in front of you. Validate it exists for others. Then scale it. The best tools get built by the people who were most frustrated by the gap. Are you building something to solve your own inefficiencies, or someone else’s assumptions?