Embedding AI features that customers actually pay for

# EmbeddingEmbeddingAn embedding is a numerical vector that represents data (text, images, or items) in a way that captures meaning, so similar items sit close together in space.View full definition → AI features that customers actually pay for

In 2023, Notion launched Notion AI as a $10-per-member add-on, separate from its base subscription. Skeptics called it a bolt-on gimmick. By 2024, Notion had folded much of that AI into higher-priced core plans, making "AI-included" a reason to upgrade rather than a side purchase. The lesson: the way you package AI often matters more than the AI itself.

This lesson dissects how two companies (Notion and Intercom) turned AI from a free experiment into revenue, and gives you a framework to decide which AI features to charge for versus which to give away.

Why most AI features fail to earn money

Adding AI is easy. Most SaaS products can wrap a large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → (LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition →, a text-generating AI like the ones behind ChatGPT) around existing data in a weekend.

That is exactly the problem. If your competitors can do the same thing in a weekend, the feature is a commodity: something so common it carries no premium price. A "summarize this" button is now table stakes, not a differentiator.

AI features earn money in only a few situations:

• They do work the customer would otherwise pay a human to do.
• They are hard to replicate because they depend on proprietary data or workflow.
• They create measurable outcomes (tickets resolved, hours saved) you can attach a number to.

Everything else is a feature you ship to stay competitive, not to grow revenue.

Case 1: Intercom Fin and outcome-based pricing

Intercom sells customer support software. In 2023 it launched Fin, an AI agent that answers customer questions automatically by reading a company's help documentation.

Here is the crucial design decision: Intercom priced Fin per resolution, publicly stated as roughly $0.99 per resolved conversation (check Intercom's current pricing for the latest figure). You pay only when Fin actually closes a customer issue without a human.

Why this works:

It maps to a real cost. A human support agent costs a company far more than a dollar per ticket. Fin's price sits comfortably below the alternative, so the value is obvious.

It aligns incentives. Customers do not pay for AI that fails. That removes the biggest objection to AI: "what if it hallucinates and I still pay?"

It scales with value delivered. A company handling more volume pays more, but only because it received more resolved tickets. This is outcome-based pricing (charging for results, not for access or usage).

The AI here is not magic. Many vendors built similar support bots. What Intercom nailed was tying price to a countable business outcome.

Case 2: Notion AI and the packaging pivot

Notion took a different path. Its AI writing and Q&A tools were initially a separate add-on. The friction was obvious: every team had to decide, seat by seat, whether AI was worth another $10.

Notion's move was to stop selling AI as a standalone product and bake it into its Business and Enterprise plans. AI became a reason to move up a tier, not a checkbox to argue about.

Why the pivot made sense:

Writing assistance is becoming a commodity. Every document tool now has it. Charging separately for a commodity invites churn and comparison shopping.

Bundling reduces decision fatigue. One price, AI included, removes a recurring internal negotiation.

It defends the core product. By putting AI inside higher tiers, Notion made those tiers stickier without risking that a cheaper competitor's free AI undercuts a paid add-on.

The takeaway is the mirror image of Intercom. When AI is differentiating and measurable, charge for the outcome. When AI is becoming standard, bundle it to protect and upgrade your core.

A framework: charge, bundle, or free

Sort every AI feature into one of three buckets.

1. Charge separately (or per outcome)

Use this when the feature:

• Replaces expensive human labor (support agents, analysts, copywriters).
• Produces a countable result you can price against.
• Depends on data or workflow rivals cannot easily copy.

Example: an AI that drafts and files compliance reports inside a fintech product. It saves specialized labor and touches proprietary data.

2. Bundle into a higher tier

Use this when the feature:

• Is genuinely useful but increasingly common.
• Makes the whole product stickier.
• Would cause friction if sold as a nickel-and-dime add-on.

Example: AI meeting notes inside a project tool. Nice, not unique, better as an upgrade lure.

3. Give it away free

Use this when the feature:

• Is now expected by every buyer.
• Losing it would make you look dated.
• Charging would annoy customers more than it earns.

Example: autocomplete or basic grammar suggestions. Charging for these in 2026 signals you are behind.

🎬 [VIDEO: "How to Price AI Features" - youtube.com - a practical walkthrough of usage, outcome, and bundled pricing models for software products]

The commoditization trap

The biggest strategic error is charging a premium for something about to become free everywhere.

When you build an AI feature, ask: will this be a default in every competing product within 18 months? If yes, do not build a pricing strategy around it. It buys you table stakes, not margin.

The features that resist commoditization share one trait: they are anchored to your unique data or workflow. A generic summarizer is copyable. A summarizer trained on a customer's own five years of internal tickets, sitting inside the workflow where those tickets live, is not.

This is why the moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.View full definition → is rarely the model. Anyone can call the same LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition →. The moatmoatA lasting edge over competitors: a resource, capability or position they cannot easily replicate, letting a firm earn above-average returns over time.View full definition → is the proprietary context you feed it and the workflow you embed it in.

A quick technical reality check

Non-technical readers often assume AI cost is fixed. It is not. Each AI call has a variable cost (usually priced per tokentokenA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition →, a chunk of text roughly three-quarters of a word). This matters for your margins.

Rough mental model:

Cost per AI feature use = 
   (input tokens + output tokens) x model price per token

Gross margin on feature = 
   price charged to customer - cost per use - support overhead

If you bundle unlimited AI into a flat monthly plan, a few heavy users can quietly erode your margin. This is why many vendors add usage caps or credits even inside "included" tiers. Know your per-use cost before you promise "unlimited."

Bringing it together: a decision sequence

When evaluating any proposed AI feature, run it through these questions in order:

1. Does it replace paid human work? If yes, lean toward charging per outcome (the Intercom Fin path).

2. Is it becoming standard? If yes, bundle it into a higher tier to defend the core (the Notion path).

3. Is it already expected everywhere? If yes, give it away and compete elsewhere.

4. What is my per-use cost? Model this before committing to flat pricing.

5. What is my moat? If the answer is only "we call an LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition →," expect it to commoditize fast.

Notice that the same AI capability can land in different buckets for different companies. AI support answers are a paid outcome for Intercom because support is their core business. For a project management tool, the same capability might be a free convenience. Context decides pricing, not the technology.

Key takeaways

• Price AI against the alternative it replaces. Intercom charges per resolution because a resolved ticket has a clear human-labor comparison. Outcome-based pricing removes the "what if it fails" objection.
• Bundle commoditizing AI to defend your core. Notion folded AI into higher tiers rather than selling a fragile standalone add-on, turning AI into an upgrade reason.
• Your moat is proprietary data and workflow, not the model. Anyone can call the same LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition →. Context and embeddingembeddingAn embedding is a numerical vector that represents data (text, images, or items) in a way that captures meaning, so similar items sit close together in space.View full definition → are what resist copying.
• Know your per-use token cost before promising "unlimited." Variable AI costs can silently destroy margins on flat-rate plans.
• Ask if a feature will be a default within 18 months. If yes, treat it as table stakes, not a revenue driver.