Not every AI feature has to pay for itself
• 6 min read
Companies are still working out how to charge for AI, and the answers keep changing. This year GitHub moved Copilot’s agent work to token-based credits because the old model was “no longer sustainable,” Atlassian added credits and overage to Rovo, and Notion added credits for its agents. Those are all answers to “how do we charge for this?” PostHog, the product analytics company, asks a question that comes before it: should this AI feature make money at all? Its answer depends on what the AI does for the rest of the product. Its public handbook says features that are core to using the platform “should be free, or priced close to cost,” and for its AI, “we should charge close to cost, so we don’t put users off using it.”
The prices follow from that rule. The wizard that sets PostHog up in your codebase is free. The MCP server, which lets an outside agent like Claude query your analytics, is free apart from a few tools that run PostHog’s own models. Asking the AI about PostHog’s docs is free. PostHog AI, the agent built into the product, is metered at “a simple, consistent 20% markup over the underlying LLM provider’s cost,” after 500 free credits (about $5) a month. Its co-CEO, James Hawkins, told Stripe the question they started from was “Is this an N+1 feature, or is it a platform component?”
I’m the kind of customer the free part is for. PostHog’s own AI can do things the MCP can’t, like session summaries and deep research, but when I want something from PostHog I usually just ask Claude, since I’m already working there. The model cost lands on my Claude subscription, and the thing that gets used more is PostHog’s analytics, which is what they charge for.
I think Hawkins’s question is the right one for any AI feature, and the version I’d ask is whether it makes people use the core product more or whether it’s new functionality on its own. If it feeds the core, price it near cost or give it away, because the money is in what it unlocks. If it’s a product of its own, it should earn its own price, and a meter is fine. Shopify’s Sidekick shows the same split in practice, without saying why. The assistant is included with every Shopify plan, while generating apps with it is limited to the Grow, Advanced and Plus plans, with hourly and weekly limits. “Free” also usually means “in the plan price.” When Google and Microsoft folded AI into their plans in 2025, both raised the base price.
A free MCP also changes who pays for the AI. The customer brings their own model, so PostHog gives away the access, the customer pays for the inference through a subscription they already have, and the work still happens in PostHog. That isn’t a niche path. Hawkins wrote in May that “the majority of dashboards were created by agents via PostHog AI, MCP, our API, or our onboarding wizard,” and that MCP usage was “roughly doubling every month.” It isn’t only PostHog, either. Fullstory, which also sells session replay and analytics, makes its MCP available to all paying customers.
I feel the other side on tools where the AI is the product. On the free plans of v0 and Lovable, I catch myself saving prompts to stay under the limit, the same way I used to watch Cline’s running cost before letting it keep going.
None of this is easy to measure. Atlassian made Rovo included at no additional cost in its Premium and Enterprise plans in 2025, and later reported that adopters grew their annual recurring revenue “more than 2x faster than non-adopters” and completed 20% more Jira work items. That’s the closest public number I’ve found, but the customers who adopt a new feature first tend to be the healthy ones anyway. Jeff Bezos had the same problem with low prices and free shipping. In his 2005 letter he wrote that Amazon could estimate a price cut’s effect this quarter but “cannot numerically estimate” its long-term effect, and called free shipping and Prime “similar judgments.” What you can watch is whether accounts that use the AI end up using more of the core product, and what the AI costs you per account while you find out.
PostHog’s own history is a good test of the idea. The AI launched completely free in June 2025 “while we gather more information about usage and decide pricing.” It got real pricing in November with $20 of free credits a month, which dropped to $5 in June 2026. PostHog hasn’t said why. It could be cost, cheaper models or just calibration. From September, startup credits can’t be spent on its AI tools at all, because “token-based pricing makes the cost of these tools prohibitive and unpredictable.” That could look like the idea falling apart, but the wizard, the MCP and the docs help all stayed free, and the tightening landed on the agent doing the work. Copilot drew the same line. Completions “remain included in all plans,” and the long agent sessions moved to credits.
In Your pricing is the product I argued that the meter changes how people use what you built. Free shipping shows where this probably goes. It started as a bet Amazon couldn’t measure and turned into something customers expect, built into the price. MCPs are most of the way there already, treated like an API that comes with the plan, since the customer’s own model does the thinking. I think the AI that feeds the core product goes the same way, included and eventually expected, while the AI that’s a product of its own keeps its meter.