Where platforms differ: the layers below the demo
Every platform can show a striking image; the differences sit lower down. At the orchestration layer, ask how a brief becomes an instruction to a model, what happens when a model fails or is withdrawn, and whether users must choose models themselves. At the brand layer, ask whether rules are stored as data the system applies, or as a PDF the user is meant to read. At the workflow layer, check that approvals bind to a specific version, so an edit after sign-off reopens review. Analytics is the most neglected layer: without usage and review time by team, market and medium, nobody can say what the platform costs per asset or where work queues.
Build, buy or assemble
Building on model APIs appeals to companies with strong engineering teams, and the first prototype is quick. The long tail is not: contracts and security reviews with each model provider, an evaluation set to re-test every new model, brand rules turned into code, an interface non-technical staff will use, approvals, records, and someone on call when a provider changes its terms. Assembling point tools avoids the engineering but moves the work into hand-offs between them. Buying moves the engineering to a vendor at the cost of depending on its roadmap. Building tends to pay off only where generation is part of the product the company itself sells, not a service to its marketing team.
What drives total cost: seats, tenants and usage
Pricing units shape behaviour, so read them against how the organisation actually works. Seat pricing punishes broad access: if occasional reviewers, legal approvers and market managers each need a paid seat, people start sharing logins and the approval record stops meaning anything. Tenant pricing matters to groups with several brands or markets and to agencies with many clients, so ask what a tenant includes and whether shared rules, such as a group legal line, can be maintained once rather than copied into each. For usage, ask whether it is pooled across the organisation or allocated per team, and what happens when one market runs out in the middle of a campaign. Then add the internal costs no vendor quotes: writing brand rules down, connecting systems, training and administration.
Test it on a campaign you have already made
Demos and fresh test briefs flatter every platform, because nobody knows what good would have looked like. Replay a campaign you shipped last quarter instead: give the platform the original brief, products and guidelines, and compare the output with what actually ran, how long it took and how many rounds of changes it needed. Include one awkward requirement on purpose, such as a product with fine print, a right-to-left language or a legal line that must survive every crop. Let the people who reviewed the real campaign review the replay. Then run a short pilot on live work in one market, with the security review done before any live material goes in, so the pilot tests the platform rather than the paperwork.
Updated 25 September 2026