How does AI affect software company moats?

A moat is a durable advantage that protects a company from competitors copying it and taking its customers.

There are two main shifts at play right now:

The cost of writing software has dramatically fallen due to coding agents.
The users of software will be agents.

This essay is a thought experiment imagining an extreme hypothetical future case where the cost of writing software falls to zero and the users of software are 100% agents, and estimating how software company moats shift in that world.

Inspired by https://x.com/nicbstme/status/2023501562480644501, which was a post about the moats of vertical SaaS. But I wanted to both broaden the analysis and also add the axis of software users being AI agents, as opposed to humans.

Another change is that, as vibe coding becomes more common, non-software engineers can also become competitive at replacing SaaS products. So, your moat-less SaaS's competitors could be another SaaS company, or the services company that your SaaS is selling to, implementing the software itself with an AI agent.

The Software Moats I'm analyzing:

Network effects
Switching costs
Distribution
Branding
Economies of scale
Proprietary data

I did some thinking to analyze which of the moats can survive these two trends and which moats erode. The examples are theoretical, in that most of these companies are not yet experiencing the pressures of software being completely free to make and their users becoming 100% agents. Not yet, but soon!

Network effects: Each additional user adds utility for everyone else already on the network.

Bull case: Reputation of the network; with more volume comes better ability to resolve disputes, fraud, refunds, compliance, settlement, reviews. This is not really an argument that the network effects become stronger with AI but rather they remain the same strength. Example: Airbnb. Because the reviews are attached to the lister and that is handled by Airbnb, agents may still pay through them rather than lower-quality networks where reviews are not attached like Craigslist or Facebook.
Bear case: It becomes easier to list a product on a bunch of different networks. It also becomes easier for buyers to search for products on a bunch of different networks. Example: Etsy. It is a high amount of work to maintain a listing: you must handle supply, shipping, and customer queries. But people are willing to trust a variety of different vendors, so an agent would have a real advantage in maintaining listings across Amazon, Etsy, Shopify, etc. And then, when a buyer agent searches, it doesn't mind searching across networks.
Verdict: Network effect moats become weaker. An agent on either the listing side or the buying side does not mind using 30 networks vs 1 because the agent reduces the per-channel listing/catalog/CRM/support/shipping/supply chain overhead (presumes these are all APIs).

Switching costs: What is the overall cost of switching from one product to another in the same category?

So who owns the system of record?

Bull case: The SaaS owns the system of record; you can't switch off because now the SaaS also owns the AI implementation: you don't own the prompts, the agent feedback loop, the eval data. Example: Rippling. The compliance layer for any specific large company is stored in Rippling in a way that a company isn't able to process itself even if it had the data.
Bear case: The traditional main switching cost before was the retraining time cost for people to relearn how to use different software, but as software is commanded with plain English or by other agents, there's no relearning to do. Agents can quickly switch software because they store all of the information they need to use software in files that are easily reused in the new system. Example: Coding agents. The company using the coding agent owns the code, the skills, and the prompts. Prompting is in regular English, so there is no learning curve at all. Many people, including myself, have switched from Cursor to Claude Code to Codex, with each switch complete within a single hour. This is probably the highest PMF category within AI products so far, but it also has the least amount of moat, empirically.
Verdict: Switching cost moats become weaker. Vertical companies are currently rushing to embed their workflows into external AI vendors, but my belief in the end state is that business context will be embedded into wiki-style Markdown in file systems and databases that they own, allowing for rapid vendor switching.

Distribution: Distribution is whether buyers already know about you.

A person can only hold one or two names for a specific service in their mind, and it takes a long time for them to learn about more options. An AI can almost instantly grab half a million tokens of available options.

Bull case: Well, what is pre-installed on the agent and pre-baked into its training? Will a Gemini agent auto-route to Google Cloud services, and will Anthropic auto-route to Anthropic-managed sandboxes? Example: Google. If Google is [somehow] able to get its act together and make an agent that has adoption, the number of products it could point the agent at is very high.
Bear case: If a person does a single Google search, they may only read the first two results. But agents can do 20 Google searches and read 20 result pages per search. The AI can write down one file per vendor that you're looking up. An agent can document what it remembers about way more vendors. Example: DocuSign. Despite having the strongest brand name and being a verb, DocuSign is unmoated. Agents will discover that there are tons of DocuSign clones from vendors like Adobe Sign, PandaDoc, Box, etc., that all have identical feature sets.
Verdict: Distribution moats become weaker. I think developers of agents will fight the defaults; harnesses will be custom. The advantage of going shopping for vendors is too high. With the cost of vendor discovery and research dramatically falling with web research, who wouldn't try to find the perfect vendor for themselves even if it means their agent has to do deep web research?

Branding: The company or person preferred to do a job, once buyers know they exist.

This one will be kind of a tug of war between internal-facing agents and external-facing agents: what gets cheaper faster, the cost to make software or the cost to evaluate other people's software?

Bull case: Despite the fact that it becomes nearly free to create software, the end-to-end verification loop of services remains expensive, and so branding still plays a role because instead of doing an evaluation of every single possible software vendor, you'll just stick to the ones with good brand reputation. Example: CrowdStrike. CrowdStrike is enterprise security software. Enterprises can't really afford to run an end-to-end evaluation loop to see if they get hacked and lose all of their data.
Bear case: The verification loop also becomes cheap alongside creating software. Example: vector databases. My agent can now immediately test the QPS claims of different vendors, and so it's possible to evaluate many vendors directly, which is plausibly a higher signal than brand. I can have my coding agent test out the gazillion QPS claimed by Turbopuffer and then compare it to Pinecone, even though I have seen ads for Pinecone and I know other people who've used it, but haven't heard of Turbopuffer.
Verdict: Branding moats are weaker for easily verifiable work, and still just as strong for hard-to-verify work. What fraction of vendor work becomes quickly and easily verifiable? I think the fraction increases, but not to the point that most vendors can be assessed without some layer of trust.

Economies of scale: There is a fixed cost of setting up a type of service. If you have more people use it, you have a lower amortized fixed cost.

Bull case: If the costs of providing the service were mostly non-software, then those advantages remain, as AI has not made it cheaper to build electricity, data centers, RAM, etc. Example: Amazon. The expenditure required to own airports, trucks, and warehouses is very high. Similarly, the cost to own GPU hardware for data centers is very high.
Bear case: If the amortized cost was the cost to build and maintain easy-to-validate software, then if the cost of software goes to zero, you lose significant advantage. Example: Jira / Atlassian. Ticketing software was designed for people to split up work and divide tasks among themselves. If agents are doing the work, then agents are actually able to create a new system to assess the quality of the splitting up of work, at very low cost.
Verdict: Economies of scale become weaker for software, but stay the same for capital, physical infrastructure, compliance.

Proprietary data: And it's not only the data but rather the full cycle loop of being able to see a customer react to your specific website and then understand their product journey step by step.

Bull case: Companies can now auto-improve their products based on their data in an agentic loop faster than ever before. Customers interact with companies via chat, which contains the information about why they are dissatisfied with the product. They can then auto-improve their software from those interactions. Example: X. If xAI got it together on making its models better, people reacting to @grok tweet replies live is gold-standard data and could be used to quickly improve the next model; to my knowledge, there aren't really other popular contexts where multiple people can interface with an AI; other sources only have 1:1 chat data.
Bear case: However, non-proprietary public data now is even more of a utility, because any LLM can process 1,000 pages of text from a Form 8949 to categorize trades to help file your taxes. Example: Intuit TurboTax. TurboTax gets its information from publicly available documents that describe the law. An AI is able to understand these public documents and process them for you into a tax return.
Verdict: Proprietary data moats become stronger. This is one that I think we haven't really seen in action yet at all. The idea of the self-improving codebase that observes customer interaction and improves itself on the fly.

Conclusion

Theme of the weaker moats: Human memory / UI learning curve, showing up first in vendor research, public information processing, code volume

Theme of the stronger moats: Trust, payment, capital infrastructure, proprietary data

What's still missing:

A common theme of these is that I'm presuming that agents have authority to purchase and can sign up for software without human involvement. This currently is not true, but I'm presuming we are trending towards it.

Right now I would say it's not quite easy to list a product on a bunch of different networks because agentic sign-up is not yet ready. This is actually too much cognitive overhead for a person: signing up for many websites to list products on a bunch of different websites. It seems like not a big deal, but it makes a ton of difference for being able to evaluate 3 vendors vs. 300.

The area that I think is missing right now is an agentic trust layer; now that A2A payments are approximately built out, how can an agent trust a vendor? Existing online reviews don't take you through the trust process and lack context. (I'm building something here, more to come soon)

On models vs. apps & SaaS vs. services:

I think services have a better angle into having more of these moats. Today, June 26, 2026, OpenAI + Anthropic + Gemini + open-weight models' run rate is maybe ~$100B. All AI apps combined might be ~$20B. This has skewed directly in the model layer's favor in the last couple of months. This is a trend that states that, as the models get smarter, easier to use, and cheaper, many companies are rolling their own software without SaaS. This might be intuitive since vertical companies own more of these moats than SaaS companies. As the trend continues, vertical companies will continue to write a higher fraction of their own software rather than buy vertical SaaS.