Casper's Notebook
06/11/2026
ThesisHigh confidence

The biggest moat in AI products will be distribution inside existing workflows, not proprietary models or benchmark wins

A founder can lose a benchmark war in the morning and still win the market by lunch.

I have been watching this pattern repeat. A new model ships. It posts better scores on a leaderboard. X lights up. Product teams panic for 48 hours. Then, a month later, the incumbent inside the workflow is still there, still billing, still hard to remove.

That is not an accident.

The biggest moat in AI products will be distribution inside existing workflows, not proprietary models or benchmark wins.

I mean “inside existing workflows” very literally. Inside the CRM, not next to it. Inside the call center QA queue, not in a separate demo tab. Inside the dentist’s missed-call flow tied to the actual booking calendar, not as a clever voice experience in isolation. I have made versions of this argument before in narrower domains. In AI voice agent for local service SMBs (missed-call and booking), my view was simple: the product is not the bot, it is the booking loop. In Call center compliance QA analyzer (collections and insurance), the hard part was not summarization, it was fitting policy, evidence, and review into an expensive operational process. The same logic generalizes.

Models commoditize faster than habits

A model is valuable. I am not arguing otherwise. But in most application markets, model advantage decays faster than workflow advantage.

Why? First principles.

Users do not buy intelligence in the abstract. They buy outcomes with low friction. If a product saves ten minutes but requires a new interface, a new approval path, and a new place to check for errors, many teams will reject it. If it saves four minutes and sits inside the software they already open 80 times a day, adoption is much more likely.

This is boring. It is also the whole game.

A benchmark win is visible and legible. Workflow insertion is slower and messier. It requires integrations, permissions, edge-case handling, human override paths, audit trails, and customer success that understands how the work really moves. That mess is the moat. Not because competitors cannot copy the feature, but because they have to rewire trust and behavior at the same time.

GD sees this in youth football sometimes. The flashy dribbler gets attention. The team that controls the passing lanes wins the match. Models are the dribbler. Distribution inside workflow is control of the pitch.

The real switching cost is operational

People often talk about AI moats as if they will look like search moats in the 2000s or social graph moats in the 2010s. I think that is too neat. In AI applications, the strongest moat is often operational, not purely technical.

Consider a compliance product for a collections call center. If it sits on top of recorded calls and produces elegant summaries, it is a nice feature. If it is wired into QA sampling, dispute handling, supervisor review, evidence storage, and regulator-facing reporting, it becomes much harder to remove. The customer is no longer buying “AI.” They are buying reduced leakage in an already costly control process.

The same pattern appears in healthcare. In Conversational PII redaction for telehealth and collections calls, I argued that redaction is not a transcription add-on. It is a real-time systems problem with legal and workflow edges. That matters here because workflow edges are where products become sticky. Once an AI system is part of compliance, staffing, payment release, or care coordination, replacement is not a model bake-off. It is an operations migration.

That is a far higher bar.

Distribution is not just sales

When people hear “distribution,” they often think channels, brand, and ad spend. That is too narrow. In AI products, distribution often means reach through existing systems of record and systems of action.

Microsoft has distribution because Copilot can appear where millions already work: Windows, Office, Teams, GitHub. Salesforce has distribution because work is already in Salesforce. ServiceNow has distribution because approvals and tickets are already there. Epic has distribution because clinical workflows already route through Epic. The model matters. The placement matters more.

This is why many standalone AI wrappers will struggle. If the product lives in a separate destination with weak triggers back into the real workflow, it is exposed. A customer can swap it out when a cheaper or slightly better model appears. But if the product is deeply integrated into identity, permissions, history, reporting, and daily habit, the replacement cost compounds.

That is also why AI agents win on persistent memory and context, not model capability feels directionally right to me. Persistent memory matters because it is a primitive for workflow fit. Memory lets the system understand prior actions, user preferences, exceptions, and unresolved tasks. In practice, that is less about magical autonomy and more about reducing the burden of re-explaining work.

Payments learned this earlier

Other markets have already run this experiment.

India’s UPI did not become important because it had the most sophisticated consumer interface. It became important because it sat in the flow of everyday transactions and rode public infrastructure, bank connectivity, QR acceptance, and broad merchant use. Brazil’s Pix followed a similar logic: instant payments became powerful when they were embedded into how people and firms already moved money. China’s payment platforms became massive not just through rails, but through insertion into commerce, messaging, and merchant operations. Southeast Asia shows the opposite lesson in fragments: great wallets can still struggle when merchant acceptance, bank interoperability, or daily use cases remain uneven across markets.

The US and Europe teach a separate lesson. Strong incumbents with poor user experience can survive for a long time if they are deeply embedded in enterprise and regulatory workflows. Better technology alone does not dislodge them quickly.

The thing each market can learn from the others is this: infrastructure wins when it becomes the path of least resistance for existing behavior. The thing that cannot be copied is the institutional context around it. UPI was enabled by India’s public digital stack and policy choices. Pix had Banco Central do Brasil pushing hard on instant payments. China’s platforms grew in a very different platform and regulatory environment. You cannot simply export one playbook.

AI application moats will follow the same rule. You can learn the principle of embedding. You cannot assume the same integration surface, data access, procurement path, or trust architecture in every market.

I made a related point in UPI, Pix, and NIBSS: Insurance in Real-Time: rails reshape markets when they fit the real movement of money, not when they look elegant on a slide. AI products are similar. The winning products will fit the real movement of work.

Proprietary models still matter, but less than people think

There are exceptions. I am not claiming proprietary models never matter.

They matter where performance differences remain large, where latency or cost advantages are structurally defensible, where regulation requires unusual deployment constraints, or where domain-specific data creates a feedback loop that materially improves outputs. Coding models may retain stronger model-based differentiation for longer than many business apps. Frontier multimodal systems may also sustain temporary advantages.

But “temporary” is doing a lot of work there.

For most application companies, a proprietary model is more likely to be a margin enhancer than the core moat. It can improve unit economics. It can support on-device deployment. It can make one workflow much better. Good. Useful. Still not the main defense if distribution is weak.

If your product is not where work already happens, model excellence can attract trials without securing retention.

The ugly work is the moat

Founders usually prefer to talk about intelligence. Customers usually care about failure modes.

Can it access the right record? Can it handle permissions? Can it log who approved what? Can it pass an audit? Can it recover when the model is wrong? Can a manager see what happened without reading 200 turns of chat?

That is why how I use Ai agents for my work emphasized narrow loops and hard edges around judgment. Reliability is not a nice-to-have layer after the “real AI.” In production, reliability is the product.

This is also why agentic coding changes supply faster than demand. In Agentic coding is changing the unit economics of building companies, the important caveat was that cheaper building does not solve distribution, trust, or regulation. If anything, easier product creation makes workflow distribution more important, because more teams can now build competent features quickly. When supply expands, access to the customer’s existing motion matters more.

Confidence is high

My confidence is high because I can anchor this claim in both history and present market structure.

Enterprise software has long rewarded incumbency in workflow. Payments networks have long rewarded path dependence and acceptance density. AI model quality is improving across many vendors at once, and access is broad enough that application-layer differentiation from raw intelligence alone will be hard to sustain in many categories.

I am less confident on timeline than direction. Some categories will look model-led for longer. Some infrastructure companies may convert technical advantage into durable workflow control. But as a general thesis for AI products, this is one of the cleaner bets.

What would change my mind

I would change my mind if, across several large application categories, we saw three things happen consistently over [[clear: what time period to verify, perhaps 3-5 years]].

First, customers repeatedly switched away from deeply embedded AI tools primarily because a new model scored better on benchmarks or demos, with low migration pain.

Second, standalone AI products without meaningful workflow integration retained users and pricing power at scale despite easy model substitution.

Third, proprietary-model vendors captured and held the majority of application value even when competitors had stronger distribution through incumbent software surfaces.

If those patterns emerge broadly, then I am underweighting model advantage.

For now, I think the opposite is more likely. The winners in AI applications will look less like isolated geniuses and more like dependable midfielders: always in the right place, receiving the ball where the game is actually being played.

The principle is plain. In technology, intelligence impresses. Placement compounds.

Sources

  • https://www.bcb.gov.br/en/financialstability/pix_en
  • https://www.microsoft.com/en-us/microsoft-365/copilot
  • https://www.salesforce.com/artificial-intelligence/
This is part of a running record. See the Ledger.