Walled Gardens: A Short History, and the Case for a Private AI Environment for Franchises

A private AI environment for franchises is a walled garden by another name: a bounded space where the brand's material, the network's conversations and the answers built from both stay inside, and nothing crosses the wall by accident. The idea is older than the web. Franchising is the case it fits best.

This industry already runs several thousand of these boundaries without calling them that: FRANdata's forecasting model tracks approximately 4,000+ US franchise brands, and each one is a documented way of operating that somebody pays a royalty to be inside.

This post is the idea and its history. How a particular product builds one is next month's argument, and mixing the two is how a concept ends up sounding like a feature list.

Where the phrase came from, and what the first gardens sold

The term arrived from cable television, where it described a system operator who decided which channels reached the set. It moved to computing when the online services of the late eighties and early nineties — CompuServe, Prodigy, AOL, France's Minitel — sold subscribers a catalogue rather than a network.

The bargain those services offered is worth restating, because it is the same bargain every walled garden has offered since:

• A known catalogue. Everything inside had been chosen by somebody, so a subscriber did not have to evaluate it themselves.
• Billing that worked. One relationship, one invoice, one place to complain.
• Identity. Inside the garden you were an account, not an anonymous connection, which made moderation and support possible.
• Predictable behaviour. The service worked the same way on Tuesday as it had on Monday.

None of that was fraudulent, and the caricature of these services as merely restrictive is unfair. They were solving problems the open network genuinely had. What they were also doing, of course, was keeping the subscriber's attention and the subscriber's money inside a boundary they controlled.

The open network won that round. The walls came back anyway

Then the web arrived, and the gardens lost decisively on the one dimension they could not compete on: selection. No curated catalogue could match a network anybody could publish to, and by the end of the nineties the walled garden was a term of abuse.

It has aged into something more interesting than that. Because the walls did not disappear — they moved. App stores, social platforms, messaging networks, the enterprise SaaS tenancy: each is a bounded space with a gatekeeper, entered voluntarily because the boundary provides something the open version does not.

The pattern worth taking from this is that gardens relocate to wherever the valuable thing sits. When distribution was scarce, the wall went around distribution. When attention became scarce, it went around attention. Communities did the same at smaller scale, which is why a network that wants a real conversation among its operators reaches for a private community platform franchise members can be identified inside, rather than a public forum.

The valuable thing now is context — the accumulated material and conversation that makes an answer specific to one organisation rather than generically competent. That is what the current round of walls is being built around, and it is worth being unsentimental about the fact that this round has commercial motives too.

What "private" is currently made to mean

The enterprise AI market has been rerunning the whole history in about three years, and the word doing the most work is the one with the least agreed meaning.

At present "private" is used for at least four different things: a contractual promise that your inputs are not used to train a general model; a tenancy where your data sits in a partition of somebody's cloud; a deployment inside your own network perimeter; and, most loosely, a settings toggle in a consumer product.

Those are not equivalent, and the first two are good and widely available. The major providers' business tiers handle them competently, and a brand that adopts one has genuinely improved on a network of personal accounts.

The gap they leave sits somewhere else entirely. A boundary drawn around a model tells you where your text is kept, and says nothing about which people on which side of a commercial relationship may see what the text produced. That second question is the one distributed businesses actually have, and it is a question about the shape of the organisation rather than about infrastructure.

Why franchising is the case a private AI environment for franchises was made for

Here is the argument, and I think it is the strongest version available to anyone in this industry.

Franchise GDP came in at $549.9 billion for 2025, on the IFA and FRANdata outlook, and almost all of that output is produced by people executing a playbook somebody else wrote.

The playbook is the product. A franchisee is not paying a royalty for a logo and a supply agreement; they are paying for the inside of a boundary — for the accumulated, expensive, specific knowledge of how this business is run well, which they could not assemble alone and which stops being worth paying for the moment it is generally available.

Which makes franchising unusual in a way that cuts against the brand. Most companies protect their operating knowledge by keeping it in few hands. A franchisor is contractually obliged to hand it out, to every operator, permanently, as a condition of the model working at all. The wall cannot be built by restricting who holds the manual, because distribution is the point.

So the boundary has to live somewhere else: around where the material may be taken, and what may be done with it once it is there. It is the only place left to put it, which makes this a structural question long before it is a technology one.

What a good wall restricts, and what it must not

The failure mode of every garden in this history was overreach — restricting things the boundary did not need to restrict, until the people inside started climbing out.

A defensible wall restricts four things: where the brand's material may travel; what any outside party may retain or learn from it; which layer of the relationship may see which facts, in both directions, since a franchisee's own numbers belong to them; and what the whole thing costs, which is a boundary in its own right and the subject of franchise AI cost control.

It must not restrict the questions people are allowed to ask, the quality of what answers them, or the exit. A network that cannot take its accumulated record out is not inside a garden; it is inside somebody else's business.

The gardens that failed sold safety and delivered a smaller world. The ones that lasted sold a boundary the members would have drawn themselves. Franchising is the rare case where the members already did — they signed it, and they pay for it every week.

In practice, here is what the boundary gets drawn around: AI franchise management software.