Does any AI assistant you can text have a moat?

Seven kinds of durable advantage exist in the standard account of them. Across the 62 AI assistants catalogued here, five are unavailable, one is available and unoccupied, and one company is building the seventh.

This is a question about structure rather than quality, and the two get confused constantly. Every product in this directory could be excellent and none of them defensible.

What a moat is, and what it is not

A moat is not a head start, a feature, or a brand somebody likes. Hamilton Helmer's 7 Powers gives the tightest definition: Power is "the set of conditions creating the potential for persistent differential returns," and every Power has two halves. A benefit, meaning you charge more or spend less. And a barrier, meaning a specific reason a competent competitor does not simply copy you.

The barrier is the half people skip. "We have great memory" is a benefit with no barrier attached. The question is always what stops the next company from having great memory too, and the answer is usually nothing.

Helmer's seven, each with the barrier that makes it one:

• Scale economies. Share gains cost a competitor too much.
• Network economies. The value rises with every user, so the market tips.
• Counter-positioning. Copying you would damage the business they already have.
• Switching costs. They must compensate you to move.
• Branding. The reinforcement takes years nobody can skip.
• Cornered resource. Somebody else decides who gets it.
• Process power. The routines are opaque and slow to build.

Take them in order against what this directory records.

The three that are dead here

Scale economies. The benefit would be falling unit cost with volume. These products rent inference from the same two or three model providers on published per-token pricing, and the channel is a flat monthly line or a per-message fee. Nobody here buys anything at a volume that changes their price, and the one thing that does scale, a dedicated iMessage line, gets more expensive with growth: providers publish daily caps, and passing one means buying another line rather than a bigger plan. No.

Process power. Requires organisational routines a competitor cannot replicate quickly. The oldest product in this set is measured in months. There has not been time. No.

Branding. Requires years of reinforcement before anybody pays a premium for the name. In a market where 36 of 62 decline to publish a price at all, nobody is charging a premium for anything yet. Not yet, and possibly never, because the brand people already trust in this thread belongs to Apple or Meta.

The two that look real and are not

Cornered resource means preferential access to a coveted asset, held by somebody else's decision. iMessage access looks exactly like this. Apple publishes no general API, 50 of 62 listings reach the channel anyway, and getting there is genuinely hard.

It is still not a cornered resource, for a reason the data makes plain. Of the 9 listings that name who carries their messages, 8 name the same company. The access is rented, from a vendor that will rent it to the next product on the same terms, and a directory page lists the alternatives with their prices. A resource everybody can buy is a supplier.

The exception is real and singular. One listing of 62 reaches iMessage through Apple's own programme, which Apple decides and has closed to new partners. That product has a cornered resource, and it is inbound-only by design, which caps what the resource is worth.

Switching costs are the strongest-looking candidate at first glance. 33 of 62 state persistent memory, and an assistant that has learned your calendar, your family and your preferences is genuinely annoying to leave.

The barrier is thinner than it looks. Helmer's test is whether a competitor must compensate you to switch, and the compensation here is a few weeks of the new one learning you. Real, small, and shrinking every time context gets cheaper.

It would be a serious barrier if memory were both deep and unportable. Nobody has shown it is either. 15 of 62 state what happens to your data when you cancel, and not one publishes an export.

Counter-positioning, available and empty

Counter-positioning is the one Power a small company gets to use against a large one. It requires a business model the incumbent will not copy, because copying it would damage the business they already have.

There is an obvious one sitting here. An assistant that reads your mail and refuses to train on any of it is positioned against a company whose economics depend on the data, and refusing is cheap for a startup and expensive for an incumbent. Twenty of 62 products state they do not train on your conversations. Reading every tagline on the site, one of them says so in the sentence it uses to describe itself.

So the Power is available, 20 companies qualify for it, and nobody has claimed it. That is not a moat anybody has. It is a moat sitting unattended.

And the incumbent has noticed. The day after this was written, Meta's founder described giving every Muse user a cloud machine holding private data that Meta says it cannot see, and said he knows of no other agent product with anything close. Counter-positioning only works while the incumbent will not follow. If the company with two billion people in its app is willing to say that on the record, the position is being taken from the other end, by the one player it was supposed to work against.

Network economies, and the one company building one

This is the Power that matters, and it needs one distinction first.

The NFX taxonomy separates sixteen kinds and ranks direct network effects as by far the strongest. The relevant one here is personal utility: your identity is tied to the network, the utility is daily, and value compounds through overlapping small groups rather than through raw user count. WhatsApp and iMessage are the canonical examples. It is the strongest defensibility available to anybody.

Which is the problem. The personal utility network these products live inside is already built, and Apple and Meta own it. 50 of 62 listings reach you through Apple's network; 12 through Meta's. Every one of them is a guest.

On 9 September one company tried to build a second network inside the first. Instinct shipped a Trusted Person network, and the wording is precise about the shape: your Instinct can talk to other Instincts, and only to agents belonging to people you have named as close to you. Spouse, parents, older children, colleagues, small local businesses.

Read against the taxonomy, that is not a coordination feature. It is a direct network effect of the strongest available kind, deliberately constructed: identity-bound, daily, and organised into exactly the overlapping sub-groups that make personal utility networks compound. If your spouse has to be on Instinct for your agent to talk to theirs, every new user makes the product more valuable to four or five specific people, and those people have a reason to join that no competitor's marketing can supply.

It is one of 62 listings. A field was added to this directory the day it shipped, and it has one entry.

Two things would kill it, and both are visible. If the protocol opens, the effect belongs to whoever owns the protocol rather than to Instinct, which is the trade every messaging network has faced. And a bounded circle bounds the effect. A network of five people around each user is defensible and small. The 5 listings that work by being added to a group thread get a weaker version of it for nothing, because the second and third person are already in the room.

What this adds up to

Of seven Powers, three are structurally unavailable in this market, two look available and are rented or shallow, one is available and unclaimed, and one is being built by a single company three days old in its current form.

That is a market with no moats in it yet. It is the honest answer to the question, and not a prediction that there will never be any.

The thing worth noticing is that both live Powers are about other people. Counter-positioning is a promise to whoever is on the other end of your email. Network economies are whoever is on the other end of the plan. Nothing on the capability list is a moat at all, and the capability list is where nearly all the product effort in this directory is going.

What would change the answer

Three things, each checkable against a field on this site rather than against a narrative.

A second product shipping agent-to-agent. One is a bet, and the capability field will say when there is a second. If the second one interoperates with the first, the Power moves to the protocol and neither company has it.

Somebody claiming the empty position. The first product to put a training answer in its own headline takes a spot 20 others already qualify for and none has bothered to occupy.

Anybody publishing a memory export. That would tell us whether switching costs here are real, and it would tell us by making them smaller, which is why nobody will publish one and why the absence is the finding.