What should the decision actually turn on?

The controls and the shape, rather than the models. Whether it satisfies your security review, where material is held and how it expects work to be structured are the things you will still be living with in three years, and the model line-up will have turned over twice by then.

How hard is it to leave?

Harder than the data suggests, because the data is usually the easy part. What does not move is everything assembled inside the platform's own conventions, so the exit cost is measured in rebuilt work rather than in exported files, and it is worth estimating before signing rather than after.

Do we need one at all?

Not if the work is narrow and not if your requirements are ordinary. Assembling the pieces directly is entirely viable and frequently cheaper; a bundle earns its price when the controls are genuinely hard to build, which usually means a regulated setting or an organisation large enough that governance is the real problem.

Glossary

AI platform

An AI platform is a bundle of models, tools and controls sold as a base to build on, and what you are really buying is an opinion about how work should be assembled.

In plain terms

A vendor's bundle sold as something to build on rather than something to use. It gathers the pieces, adds the controls a company needs, and asks you to assemble your own things inside it. The pieces get replaced over time; the shape it expects your work to take is the part that stays.

Why it matters

Because the reversible part and the irreversible part are easy to confuse. Which models it offers changes every few months and matters far less than it appears to, while the way it expects work to be structured shapes everything built inside it, and that is what makes leaving expensive years later.

How it works

It bundles rather than builds, and the bundle is the product. Access to models, tooling for assembling things, somewhere to keep company material and the controls a security team asks for arrive together, which is the difference between this and a finished application you simply use.

The models inside are the least durable part despite being the headline. They are replaced, renamed and repriced constantly, and a platform choice made on which model it carries this quarter is a choice made on the most volatile thing in the bundle.

The lasting commitment is structural, and it is rarely discussed at purchase. How the platform expects work to be described, where material lives, what an assembled thing looks like: everything built inside inherits those decisions, and rebuilding them elsewhere is the real cost of leaving.

The controls are frequently the actual reason to buy, and are worth separating from the rest. Audit records, access rules, residency guarantees and the paperwork a security review demands are genuinely difficult to assemble independently, and for a regulated buyer they can justify the whole arrangement on their own.

What changes, what stays

There is a reason the comparisons cover the wrong things, and it is not carelessness. The left column is legible: model names, context sizes, prices per unit, all of it public and tabulated and updated weekly by people who do nothing else. The right column resists that treatment entirely, because how a platform expects work to be structured cannot be stated in a cell, and its consequences do not appear until somebody has built inside it for a year. So the available information is abundant, precise and about the part that will be obsolete by spring, while the information that matters is qualitative, hard to obtain and available mainly from people already committed. The practical response is to weight what the existing customers say about their second year over anything in the specification, and to treat the model line-up as a tiebreak rather than a criterion.

Seen in the wild

An enterprise platform with its own model family, retrieval against company knowledge and an agent layer operating under brand and compliance controls.
Writer
A no-code platform where autonomous workers are assembled from tools, triggers and instructions rather than written, then deployed against real work.
Relevance AI
One key and one balance in front of many models across many providers, which is the thin version of the same idea.
OpenRouter

Common misconceptions

People assume

The choice is mainly about which models it offers.

In fact

That is the fastest-changing part of the bundle and the easiest to replace. What persists is the shape it imposes on everything built inside it, which is why a decision made on this quarter's model line-up frequently looks arbitrary a year later.

People assume

One platform means one supplier and simpler procurement.

In fact

It usually means one contract over several suppliers underneath, since most bundles carry models they did not build. That is not a defect, and it does mean questions about where material goes and who processes it need asking about the layers rather than about the badge on the invoice.

Telling them apart

AI platform vs AI stack

AI platform

One vendor's bundle, with their opinion built in.

AI stack

The set of pieces you assembled yourself.

You buy the first and you accumulate the second, which is why one has a renewal date and the other has an owner.

Questions

What should the decision actually turn on?: The controls and the shape, rather than the models. Whether it satisfies your security review, where material is held and how it expects work to be structured are the things you will still be living with in three years, and the model line-up will have turned over twice by then.
How hard is it to leave?: Harder than the data suggests, because the data is usually the easy part. What does not move is everything assembled inside the platform's own conventions, so the exit cost is measured in rebuilt work rather than in exported files, and it is worth estimating before signing rather than after.
Do we need one at all?: Not if the work is narrow and not if your requirements are ordinary. Assembling the pieces directly is entirely viable and frequently cheaper; a bundle earns its price when the controls are genuinely hard to build, which usually means a regulated setting or an organisation large enough that governance is the real problem.