When you switch browsers, the first thing the new one offers is to import your profile: bookmarks, extensions, saved passwords, and cookies. Chrome offers to import from Firefox, and Firefox offers to import from Chrome. The reason is simple. Switching costs are high, and browser makers know it.
Fifteen years ago, the browser wars were a fight for usage share, and the only way to win was to make switching cheap enough that people would actually do it. Even then, the inertia was enormous. Bookmarks and saved logins aren't just data. They represent accumulated effort, and people don't want to rebuild that effort from scratch.
How moats used to work
Before AI, companies built moats through data gravity and API friction. Facebook made it easy to upload your photos and hard to export them. LinkedIn made it easy to add connections, and it's still surprisingly difficult to export your network. These were subtle choices. The product looked open, but the back door was locked.
The result was a kind of escape velocity. Once enough people and data were inside, the network became the product, and the switching cost became social. You couldn't leave Facebook because your friends were there, and you couldn't leave LinkedIn because your professional network was there. The moat was invisible until you tried to cross it.
There's an element of boiling the frog in how these moats work. At the beginning, the experience is genuinely appealing. The water feels warm and comfortable. You upload a few photos, add a few connections, and enjoy the payoff right away.
But the lock-in builds gradually. Every photo, every connection, and every message adds a little more to what you'd have to leave behind. No single day feels like a commitment, so the switching cost grows without anyone noticing. Facebook and LinkedIn are perfect examples of that slow, incremental entrapment. Then one day you consider leaving and realize you can't easily get out.
The AI shift
AI changes this equation because it's remarkably good at processing data. Much of today's code and documentation lives in Markdown files, plain text that both people and agents can read. Products are easier to build than ever, and the effort required to launch something new has dropped dramatically. That raises an obvious question: where does the moat end up?
A growing number of companies are trying to answer it. Meta has Muse. There's Instinct, and there's Grok Bot. New personal assistant platforms seem to launch every quarter. The branding differs, but the pitch is the same: let us be the one assistant that runs your life.
If you want a sense of how crowded the field has become, Assistant Benchmark is a useful resource. It tracks 139 assistants and has tested 32 of them on the same 15 tasks, scoring each one after real use and noting what it does best. A few examples from its leaderboard:
- Muse, "Meta’s personal AI agent for everyday tasks," is rated best at purchasing and email.
- Instinct, "A personal assistant that takes action across your apps," is rated best at travel and routines.
- Town, "The assistant that learns how you work, then gets to work," is rated best at connected apps and email.
- Pally, "The AI personal assistant you text," is rated best at routines and restraint.
- Poke, "The proactive assistant on Messages, WhatsApp and Telegram," is rated best at personality and picks.
- Wajo / Fo, "Action agents that get things done in the physical world," is rated best at phone calls and purchasing.
They're all competing for the same position, and the moat taking shape has two pillars. Neither of them is the model.
Pillar one: trust
An assistant that books your flights, pays your bills, and reads your email isn't just a tool. It's a relationship.
You don't give that kind of access to just anyone. You give it to the assistant that has earned it: the one that remembers your preferences, gets the dinner reservation right, and checks with you before spending your money. Every good interaction builds a little more confidence. After enough of them, starting over with a new assistant feels like a step backward.
The platforms understand this. Trust isn't a byproduct of a good product. It's the strategy. The thoughtful onboarding, the personality, and the memory of what you mentioned three weeks ago are all designed to make leaving feel like a real loss.
Pillar two: cookies and sessions
Trust gets you in the door. Access is what makes an assistant useful. To be valuable, an assistant needs to connect to the services you already use: Amazon, Netflix, your bank, your email, your cloud storage, and your recurring subscriptions. These are the accounts you've built up over years of life online.
The most common way to make that work is an always-on computer, either in the cloud or on your own device, that the assistant operates through. The assistant reaches each service the same way you would, by being logged in. Being logged in means cookies, sessions, and stored credentials.
The effort you put into connecting those services is the new switching cost. It's the trade-off you accept so the assistant can access your information and act on your behalf.
The key point is that those cookies and sessions don't transfer easily to another assistant.
The export problem
Say you use Grok Bot today and connect your email, calendar, bank, and cloud drives. If you decide tomorrow to switch to Claude, you can't export those connections. You start over. You log in again, re-authorize each service, and rebuild your session state from scratch.
There's no standard format for moving authenticated sessions between assistants, and there's no "import my profile" button in the agent world. Each platform keeps its logged-in state to itself, and whether by design or not, that state only flows in one direction.
That isn't an accident. It's the point. The always-on computer where your sessions live is the new moat. That's where your effort goes, and that's what's hard to leave behind.
The legal moat
There's a third layer, and it lives in the fine print.
Some of these platforms include terms of service that go beyond prohibiting you from using the service to build a competing product. They also prohibit attempting to reverse engineer the platform at all. That means not just copying it or launching a clone, but simply trying to understand how it works.
This isn't a hypothetical. Wajo, the company behind the Fo assistant, lists both restrictions in Section 5.5 of its Terms of Service. Here is how the section opens, followed by two of the restrictions it lists:
You will not, and will not permit any person or entity (including any Authorized User (defined below)) to, directly or indirectly:
reverse engineer, decompile, decode, or disassemble, or otherwise attempt to derive or gain improper access to the Services or any software component of the Services, in whole or in part;
use the Services, Output, documentation, or any Wajo confidential information for benchmarking or competitive analysis, or to develop, commercialize, license, or sell any product, service, or technology that could, directly or indirectly, compete with the Services, or to train or improve any AI model; or
Note the phrase "could, directly or indirectly, compete." The restriction isn't limited to building a clone. It reaches anything that might compete, and it rules out benchmarking along the way.
Instinct takes a similar approach. Its Terms of Service include these restrictions:
You may not do any of the following in connection with your use of the Services, unless applicable laws or regulations prohibit these restrictions or you have our written permission to do so:
duplicate, decompile, reverse engineer, disassemble or decode the Services (including any underlying idea or algorithm), or attempt to do any of the same;
use the Services for benchmarking purposes, to create or develop any competing products or services or to improve or train other AI models, tools or technologies;
The reverse engineering ban covers "any underlying idea or algorithm," and the attempt alone is enough to break the rule.
Town's Terms of Service extend the same idea to the parts of an assistant that make it feel personal:
You will not, and will not permit any user or other party to: […] (c) reverse engineer, decompile, disassemble, or otherwise derive or determine or attempt to derive or determine the source code (or the underlying ideas, algorithms, structure or organization) of the Town Technology, or attempt to extract models, prompts, memories, or system logic, except as permitted by applicable law; […] (i) access or use the Town Technology to build a similar or competitive product or service or otherwise engage in competitive analysis or benchmarking;
Pally's Terms are shorter but make the same point:
You may not:
Reverse engineer, or attempt to derive the source code, models, or underlying components of the Service
Use the Service to build or train a competing product or service
Poke's Terms of Service go furthest in scope:
2.2 Restrictions and Acceptable Use. Except to the extent a restriction is prohibited by Applicable Law, you will not do, and will not enable any third party to do, any of the following:
2.2.1 disassemble, reverse engineer, decode, or decompile any part of the Service or license or resell or modify any part of the Service;
2.2.6 use the Service in competition with us, to develop competing products or services, for benchmarking or competitive analysis of the Service, or otherwise to our detriment or disadvantage;
Not every platform goes this far. Ollie's Terms of Service prohibit you from trying to "Reverse engineer or scrape the Services," but they don't include a separate ban on building a competing product. Tomo's Terms ask users not to misuse the service but say nothing about competing products. The pattern isn't universal, but it is common among the platforms competing hardest for your accounts.
This is a newer style of legal clause, and it reveals something important. The technology is no longer the hard part. An always-on computer, a browser, a model, and a set of connectors can be assembled by a small team in a matter of weeks. The platform itself is relatively easy to replicate.
What's hard is getting people to hand over their logins, their history, and their confidence. That's the real asset. When the asset is that valuable and the technology is that easy to copy, companies turn to legal protection. These clauses aren't really protecting the code. They're protecting the switching cost.
Capital and the arms race
If the technology is easy to replicate, why are so many of these companies raising so much money?
Because building the moat is expensive, and most of them don't have the resources to do it on their own. Trust and connected accounts don't appear overnight. They have to be earned one user at a time, through onboarding, support, reliability, and enough marketing to get noticed in a crowded field. That takes capital.
It's effectively an arms race. The goal is to build a cohort of users who have committed themselves through the logins and connections they've set up across the web. Once a user has connected enough accounts, there's a point where they simply won't switch. Every platform is racing to get users past that point before a competitor does.
That tells you something about where this market is heading. Personal assistants are becoming commoditized. The platform itself is increasingly interchangeable, since most of them are built from similar parts. What isn't interchangeable is the set of connections a user has established through a particular platform. The platform is the wrapper. The login graph is the asset.
The data center IP problem
There's also a less visible cost to building this moat, and every platform shares it.
To work, each of these assistants needs an always-on computer for every user, a sandbox where the assistant can log in and act on that user's behalf. Those computers almost always run in data centers. And data center IP addresses are easy for websites to identify, because they come from well-known ranges owned by cloud providers.
That creates real friction. Today, your assistant can reach your bank or your favorite retailer without a problem. Tomorrow, that same site notices the traffic is coming from a data center rather than a home or phone connection, decides it isn't a real person, and blocks or flags it. The login that worked yesterday stops working, and the assistant can no longer do the job you asked it to do.
For the platforms, this is a constant maintenance burden. The connections that make up the moat are only valuable if they keep working, and keeping them working means fighting a steady stream of blocks, challenges, and re-authentications. The moat isn't just expensive to build. It's expensive to keep from eroding.
Why the push is happening now
This explains why every assistant platform is working hard to get users connected early. The sooner you connect your accounts, the more invested you become. Eventually you'll ask yourself whether it's worth switching, and the answer will probably be no. You've already done the work, you don't want to do it again, and you trust the assistant you have.
The moat in the age of AI isn't a better model, more training data, or even a better interface. It's the trust a user builds with one assistant, combined with the accumulated login state of their digital life, all tied to that assistant's computer. So far, no one has built an export button.
The services that seem the most open and friendly often have the deepest moats. They talk about consumer choice and convenience, but the real competition is happening underneath, in the trust, cookies, and sessions that nobody sees but everyone depends on, reinforced by terms of service most people never read. Moats have always been about making switching feel harder than staying. In the agent era, that feeling is built on sessions you can't pack up and take with you.
No comments yet