Published April 22, 2024 · By CampaignsLive · Industry
The default assumption through 2022 and 2023 was that brand-side AI adoption would happen primarily through third-party platforms. Brands would buy access to commercial tools (Midjourney, DALL·E, Firefly, the emerging specialized platforms) and integrate the output into their existing production stacks. The model was familiar — it was how brands had adopted most prior creative tooling categories, from Photoshop forward.
Through 2024, the largest brands began doing something different. Rather than only consuming third-party AI tools, they began building their own AI infrastructure — proprietary models, in-house fine-tuning pipelines, brand-asset-anchored generation surfaces, internal team-facing tooling that wrapped around third-party APIs. The wave was less publicized than the platform-side announcements but was probably more consequential for the long-term shape of how brand creative production will be organized.
This is a working description of who built what, why, and what the pattern implies for the next few years.
What “building their own AI” actually meant
Three distinct types of in-house investment emerged through 2024.
Custom fine-tuned models on brand archives. The largest fashion, automotive, and FMCG brands began training internal models on their own historical campaign archives. The work used open-source foundations (Stable Diffusion through mid-2024, FLUX after August) plus per-brand fine-tuning. The output models were not products in any commercial sense; they were internal infrastructure, available only to the brand’s own creative team and approved external partners.
Wrapper interfaces on top of third-party APIs. A more common pattern was building brand-specific internal interfaces on top of the commercial AI providers’ APIs. The interface enforced brand guidelines, integrated with internal asset management, enforced approval workflows, and presented a brand-specific surface that hid the underlying API’s defaults. The brand’s creative team interacted with the wrapper; the wrapper called out to Midjourney, OpenAI, or whoever was underneath.
Full internal AI tooling teams. A handful of the very largest brands — predominantly in tech, automotive, and consumer goods — began standing up internal AI tooling teams that combined ML engineering, creative direction, and product management. The teams’ work spanned model selection, fine-tuning, infrastructure operations, internal interface development, and the broader coordination of how AI fit into the brand’s production stack.
The pattern was visible at varying depth across many of the brands that have been the most public about their AI work — Coca-Cola, Estée Lauder, Mattel, BMW, Microsoft, and others — though the specifics of what each brand actually built were rarely fully disclosed.
Why brands built rather than only bought
Three reasons recurred in the working conversations about this.
Brand-specific output character. The third-party tools, even at their best, produced output trained on broad corpora. Brand-specific visual register — the particular composition, palette, talent direction, atmospheric quality that defines a brand’s creative identity — was reachable through prompting but not reliably reproducible. The brands that ran the highest-volume creative production found that the cost of working around the third-party defaults exceeded the cost of building infrastructure that started from the brand’s defaults.
Rights and compliance ownership. Larger brands, particularly in regulated industries, found that the rights and compliance posture they needed for AI-derived assets was easier to maintain when the infrastructure was theirs. Provenance documentation, talent rights tracking, model-version control, and the audit trail their legal teams required were operationally cleaner when the brand controlled the surface.
Vendor leverage. A brand running significant creative volume through a single third-party AI tool was, by definition, exposed to that vendor’s pricing, terms, capability changes, and strategic priorities. Brands with significant ongoing volume found that maintaining their own infrastructure — even if it cost more in absolute terms — gave them leverage in the commercial conversations and continuity if the vendor pivoted.
Where the third-party platforms still fit
The in-house tooling wave did not eliminate the role of third-party platforms. The pattern that emerged was more nuanced.
Foundation models came from third parties. Almost no brand was building image-generation foundation models from scratch. The compute requirements and ML expertise needed are at a scale that few brands have or want. The brand-side work was almost always on top of third-party foundation models, either as fine-tuning or as wrapper interfaces on commercial APIs.
Specialized brand-creative platforms competed with in-house tooling. Tools whose value proposition was brand-creative-specific training — CampaignsLive among others — found themselves competing not just against general-purpose tools but against brands’ own internal infrastructure. The competitive case was about whether the specialization, the operational maturity, and the production-stack support of a dedicated platform justified its cost compared to the brand running its own.
Long-tail tools remained third-party. The specific operations that did not justify in-house investment — design finishing, image upscaling, background removal, video stem separation — continued to be served by best-of-breed third-party tools, integrated into the in-house production stacks.
What this implied for the market
The brand-side build-vs-buy dynamic created a specific competitive shape for AI creative platforms entering 2025.
Tools whose differentiation was raw model capability were vulnerable. A brand with sufficient internal capacity could replicate the capability in-house. Differentiation on model capability alone was not durable.
Tools whose differentiation was the surrounding production stack were durable. The work that brands did not want to build themselves — the operational maturity, the print and color workflow, the rights documentation, the brand-fine-tuning automation, the audit trail — was the differentiator that survived.
The market consolidated around tools that solved the second category. Tools that focused on capability alone increasingly found themselves displaced by either better capability or in-house alternatives. Tools that focused on the production-stack work increasingly captured the brand-side volume that did not move in-house.
What this means for brand-side teams in 2026
Three implications.
The build-or-buy decision is not binary. Most brands that have done well have built specific pieces of their AI infrastructure in-house (typically the brand-specific fine-tuning and the internal team interface) while continuing to use third-party platforms for foundation capability, print and color workflow, and specialized operations. The decision is which pieces to build, not whether to build.
The internal capability bar is real. Building in-house AI infrastructure requires ML engineering, creative direction with AI fluency, infrastructure operations, and the cross-functional coordination to keep all of it aligned with the brand’s creative direction. Brands that have not built that internal capability cannot replicate the in-house pattern, and end up at a structural disadvantage to brands that have.
The third-party tool selection criteria have shifted. The right third-party tool, by 2026, is not the tool with the most attractive demo. It is the tool whose specific capability fills the gap the brand’s own infrastructure does not cover — typically the production-stack work that the in-house build did not include.
For the related discussion of where the AI tooling market sat by late 2025, see From Generation to Production-Ready: The Quiet 2025 Shift. For the broader pattern of how brand-side production reorganized, see How Agencies Restructured Around AI Creative in 2024.