AI and Software Moats: Why Pricing Models Are Breaking

Thomson Reuters and Intuit have both been hit hard in the market on AI disruption fears. Two established software and data businesses, priced as if AI is about to eat their lunch.

Everyone is talking about AI features.

That's the wrong conversation. Every company will have them within two years. The feature itself won't be the differentiator. The question that actually matters to an investor is simpler and far less exciting:

How will this company make money from AI?

For thirty years, software companies sold seats. One user, one login, one monthly fee. It was predictable, it scaled cleanly, and it made SaaS the best business model Wall Street had seen in a generation. AI is breaking that model, and the market has already started pricing in the fallout.

Whether it's pricing the right companies for the right reasons is a different question, and it's the one this piece is actually about.

The seat license worked because the cost of serving one more user was close to zero. Add a customer, collect the subscription, keep almost all of it as margin. That's why investors loved SaaS: recurring revenue, high retention, and expenses that didn't move much no matter how many people logged in.

AI removes that assumption.

Every AI feature runs on compute. Every query burns tokens. A company can sign the exact same customer, at the exact same subscription price, and have no idea whether that customer will cost them ten cents or ten dollars this month. Usage isn't a login anymore. It's a variable cost that moves with how hard the product gets used.

It’s like making a long distance phone call from a phone booth in the 80’s. You add quarters to add minutes.

Nobody has solved AI pricing yet.

Flat subscriptions expose companies to heavy users who quietly wreck the margin. Usage-based pricing sounds like the fix, until you try to implement it.

Set a company-wide average allowance and someone still has to decide what happens at the individual level. Cap each employee and the light users end up paying for capacity they never touch. Don't cap them and the heavy users quietly draw down the pool meant for everyone else, which puts the company right back to eating the exact margin loss usage-based pricing was supposed to prevent.

Nobody has settled on which version to run, let alone how to forecast revenue once they do. This isn't one company behind on pricing strategy. It's every software company, at the same time, with no existing model that forecasts variable AI cost the way the seat license forecast a subscription.

That's not because pricing is unsolvable. It's because nobody has the bandwidth to solve it. Every software company right now is racing to ship AI features before a competitor does, and that race leaves no room to sit down and fix the pricing model underneath it. Losing the feature race is the risk that gets managed first. Token cost and margin pressure get managed later, if at all.

I saw a version of this problem up close at work. Traditional software forecasting rests on cost categories with years of history behind them: headcount by location, rent, electricity, support hours. A finance team can build a defensible forecast because there's a track record to build it from. Tokens have no track record. There's no multi-year history of what normal consumption looks like for a given workflow, because most of these workflows didn't exist a few years ago. The old forecasting playbook doesn't just get harder under AI. It has nothing to reference. It’s all guess work.

Anthropic just proved the point about itself. It split access to its own newest model, Claude Fable 5, into two tiers: bundled at no per-token cost for its highest subscription plan, capped at half of normal usage limits, and a one-time $100 usage credit for standard Pro subscribers, after which they have to switch to metered billing at $10 per million input tokens and $50 per million output tokens, roughly double what its next most expensive model costs per token.

Anthropic's stated reason was that demand was "very high, and difficult to predict." That's the company selling the tokens, not a customer trying to guess someone else's usage, admitting it couldn't hold a flat price against its own model's demand. If the seller of the raw input can't forecast it, expecting every company building on top of that input to have already solved forecasting is the wrong expectation.

Anthropic said it would restore Fable 5 to a standard subscription feature once it has "sufficient capacity," without committing to a date. A company with more visibility into AI infrastructure costs than almost anyone else still can't put a number on when its own pricing problem gets solved.

Not every AI business faces this problem the same way, and the split comes down to something insurers have understood for a century.

An insurance company can't predict whether any single driver will crash this year. What it can predict, with remarkable precision, is how many drivers out of a million will crash. Individual usage is unpredictable. Average usage across a large enough pool becomes a stable number you can price against.

Consumer AI doesn't have that pool to estimate yet.

→ Usage is highly variable person to person
→ Companies respond with usage caps
→ Premium tiers appear to separate light users from heavy ones
→ Token bundles get sold like phone minutes used to be

A single consumer can 10x their usage overnight and there's no large enough base yet to smooth that out. That's why consumer AI pricing still looks experimental: caps, overage fees, tier confusion. Nobody has found the stable number, including Anthropic, which had to pull its own Pro subscribers off a flat plan and onto metered billing for exactly this reason.

Enterprise AI has the ingredient the insurance model needs. It doesn't have the pricing yet.

→ A large employee base is the kind of pool that should smooth out individual usage spikes
→ That's a structural advantage consumer AI doesn't have
→ Almost nobody is actually pricing against that average with real confidence today

The token space is too immature for that advantage to show up yet. Most companies, including large enterprises, can't tell you with any precision how many tokens a given workflow burns, let alone forecast it across a thousand employees.

That gap is exactly why a new category of tooling has shown up over the past year: platforms built specifically to track, cap, and compress the tokens a workflow or an AI agent burns before that usage ever reaches the AI provider's meter. Tools don't get built to solve problems that are already solved. Their existence is evidence that enterprise AI usage is still a black box internally, well before it becomes a stable number a vendor can price against.

Enterprise software may eventually adapt to AI pricing faster than consumer software, because the pool that makes averaging possible already exists inside a large customer base. It just hasn't gotten there yet. Consumer and enterprise AI are both pricing in the dark right now. Enterprise has a clearer path out of the dark. It hasn't walked it.

If AI makes software easier to build, and AI itself becomes available to every competitor, the natural next question follows immediately.

Why will customers keep paying?

This is where most investors stop thinking about pricing and start thinking about moats, which is the right instinct. But the way most people frame a moat question is already outdated.

The old question was: does this company have a moat?

That question no longer separates winners from losers, because almost every software company can point to something that looks like an advantage. The better question is this:

If every competitor had access to the same AI tomorrow, what couldn't they recreate?

That single reframe does more work than an entire afternoon of reading a 10-K. It forces you to separate what a company has built from what a company actually owns.

These create real friction for customers. None of them are durable moats on their own.

→ Software
→ Workflows
→ AI features
→ Integrations
→ Customer-owned data
→ Switching costs

Each of these can be copied, and increasingly, AI is the tool that makes copying them faster. A workflow that took a competitor eighteen months to build two years ago might take three months now.

Switching costs aren't the same everywhere, and the depth matters. A restaurant reservation system holds almost nothing hostage: the data is thin, the workflow is shallow, and a competitor can replicate the experience in a weekend.

A deep enterprise platform like ServiceNow sits at the other end: years of workflow logic, custom fields, and integrations wired into a company's other systems, where ripping it out is a multi-year migration with real operational risk.

That depth buys real time. It's just not the same thing as ownership. The ServiceNow-style integration slows a competitor down, worth paying attention to as an investor. But it's still the customer's cost to leave, not something the vendor owns outright, and a well-funded competitor with a genuinely better product, or an AI tool built specifically to automate that migration, will eventually make even the deep version cheaper to absorb. Depth changes how long the friction holds. It doesn't change what it is.

Proprietary data. Not data a company collects from its customers and could lose access to, but data nobody else can get at any price. Thomson Reuters, Bloomberg, and Moody's don't sell software. They sell access to information nobody else has assembled at that scale, and AI can't manufacture a data set that took decades to build. Thomson Reuters is worth naming specifically because the market didn't spare it. Its stock got hit alongside the rest of software on AI disruption fears, even while its underlying data moat is exactly the kind of thing this section argues AI can't touch. The selloff priced the business model story. It didn't price the moat.

Network infrastructure. Visa and Mastercard aren't payment apps. They're physical and regulatory networks: credit cards used in tens of millions of locations, merchant acceptance built over decades, and PCI compliance infrastructure that took years to certify. An AI model can write better fraud-detection code tomorrow. It cannot build a trusted network overnight.

Regulatory assets. Stock exchanges, credit rating agencies, and licensed financial institutions operate inside a permission structure that AI has no way to shortcut. You can't out-code a license.

Brand. Discuss this one carefully. Brand alone is rarely enough. A strong brand earns a customer's first look, not their permanent loyalty, and the same slow erosion that quietly turns a "forever" dividend stock into a weaker holding applies to brand-based moats too. Brand supports a moat. It rarely is one by itself.

A single Bloomberg Terminal seat runs roughly $32,000 a year, and Bloomberg still has an estimated 355,000 subscribers paying it. That price has nothing to do with software features. It's the cost of accessing data nobody else has assembled, which is exactly why no AI feature can undercut it.

I've used Intuit for years. I don't like the annual price increases, and I complain about them every time they land. I stay anyway, because migrating years of financial history to a new platform costs me more time than the price increase costs me in dollars. My switching cost sits somewhere between a restaurant app and a ServiceNow integration. Not a multi-year migration, but not a weekend project either.

Here's the question that changed how I think about this: if another company could migrate my history with AI tomorrow, in an afternoon instead of a weekend, would I stay?

Probably not.

That's customer friction. Not a moat. AI is precisely the technology that closes gaps like mine faster than either of us expects, and Intuit's stock has already taken a real hit on that exact fear. Not that AI makes Intuit's product worse. That a competitor uses AI to make switching frictionless, and the pricing power built on customers like me not bothering to leave quietly disappears.

When you look at any software company through an AI lens, ask five questions in this order:

1. How does it make money today?

2. How will AI change its pricing model?

3. What does the customer actually pay for?

4. If switching became effortless, would customers leave?

5. What can't competitors recreate?

The fifth question is the one that separates a real thesis from a story you want to believe. The patterns behind the holdings that superinvestors keep for decades almost always trace back to an answer to that fifth question, not to a feature list.

This piece doesn't end with a pricing model that works or a list of stocks that are safe. Nobody has solved AI pricing yet, and nobody gets to skip the company-by-company work of figuring out what a business actually owns.

What changes is the question worth asking.

Not does this company have a moat, but what couldn't a competitor recreate with the same AI: proprietary data, payment infrastructure, regulatory authority, a genuine network effect that gets stronger with every additional user. Turning that question into a habit rather than a one-time exercise is what keeps you from re-litigating the same mistake with every new AI headline.

That's the position Thomson Reuters and Intuit are actually in right now. Some of that selloff is very likely an overreaction, the market punishing a real moat for a story it doesn't own. Some of it might be the market correctly pricing in a weaker moat than the stock used to command. The five questions above are how you tell the difference, company by company, not by reading the stock chart.

If you can't answer those questions with any real confidence for a given company, that's useful information too. It means you don't have an edge on that name right now, and staying out of it, or sticking with an index fund and companies you actually understand, is a better decision than guessing. Understanding what's actually happening, in the pricing and in the moat, is the edge. There isn't a shortcut past it.

Inner Circle members see every holding, allocation change, and trade I make. Evrey month.