Brandtech Partner: AI Fails at Scale Due to Legacy Operating Models

Full analysis

Rebecca Sykes, a partner at The Brandtech Group, says the thing holding AI back inside companies isn't bad data. It's the old plumbing. Her example: a client cut one step of a task by 70% with AI, and the task still took four days, because the data still crawled through five disconnected systems with a human copy-pasting between them. For anyone building AI into a business, the lesson is that your shiny new model fixes the part that was never the bottleneck.

This is easy to undo as a framing. Nobody signs a contract off a podcast. What's actually being decided, for a reader running an AI-forward team, is where next quarter's budget goes: into better models and more tokens, or into rewiring the workflow around the models you already have. There's no deadline here, just the slow bleed of pilots that impress in the demo and move no deadline in production.

The Skeptic. This is the ERP consulting pitch with a new logo. Swap "AI" for "SAP" and this is a 2003 McKinsey deck: the software is fine, your org is broken, hire us. Sykes isn't wrong, but she isn't new either. And the "messy brand data stops the sea of sameness" line is convenient. It tells the client they need exactly the proprietary-data work that Brandtech sells. The real question she skips: can a company actually rewire a workflow end-to-end before competitive pressure forces it to just bolt on another point solution? History says no. Companies ship the 70% step and declare victory because the step is visible and the handoff is somebody else's problem.

The Builder. The 70% step that doesn't move the four-day cycle is the thing every internal AI team walks into, and nobody talks about it enough. You ship a gorgeous LLM step, Slack fills with applause, the SLA doesn't budge, because the bottleneck was always the handoff between systems. Instrument the whole workflow first, measure where the hours actually sit (usually the gaps between tools rather than the work inside any one of them), then decide where the model goes. And the "clean data first" pushback is operationally correct. Waiting for perfect data is how a pilot rots in the backlog for eighteen months. Ship on messy data, measure, then clean the parts that hurt.

The Compute Pragmatist. There's no compute story here, and that absence is the signal. If the bottleneck is workflow, not model capability, then inference cost and model quality have stopped being the limit for these jobs. Brand marketing tasks, copy, asset variants, routing, have hit a plateau where a better model doesn't move the business number. The constraint has moved entirely to integration and the lag between systems. That reprices the next spend cycle. It isn't GPUs or API tokens. It's systems integrators and middleware. The model providers and hyperscalers don't catch this wave. The SIs and the glue-code shops do. If you're a buyer, your 2027 AI line item looks more like a plumbing invoice than a compute bill.

The Safety Lens. Buried in the org-chart talk is a real admission. Sykes wants new roles for "AI governance, systems ownership, and workflow maintenance," which means nobody today owns what happens when the deployed model drifts off-brand or hallucinates at scale. That's a gap, not a hire. And the "sea of sameness" point hides a bigger one: if every brand routes through the same two or three foundation models with no proprietary grounding, a single model update or bias shift breaks everyone at once. That's correlated failure across clients on the same day. Nobody in this conversation names it as a supply-chain risk. It is one, and messy proprietary brand data is a crude hedge against it.

The Enterprise Buyer. Sykes is handing a CMO the excuse they wanted: the AI works, your processes don't, so the transformation budget is bigger and slower than the vendor demo implied. That's true, and it's also the sales motion. What a buyer should ask before signing: show me one end-to-end workflow you rebuilt, the cost of the rebuild, and the cycle-time it bought, versus the point solution. If the answer is a metaphor about frogs and a 70% step, you're buying change management, not AI. The procurement risk here isn't the model. It's committing to a multi-year re-platforming on the promise that end-to-end beats point-fixes, when the vendor gets paid either way.

Where they split. The Builder and the Compute Pragmatist agree the constraint has moved to integration, and treat that as a clean engineering problem: instrument the workflow, fix the handoffs. The Skeptic and the Enterprise Buyer see the same fact and smell a longer, pricier consulting engagement that clients rarely finish. Both can be right. The constraint really has moved to plumbing, and companies really are bad at rewiring plumbing. The second disagreement cuts deeper: the Researcher wants proof that messy brand data improves output, where the Safety Lens treats that same messy data as a necessary hedge against everyone's AI breaking on the same model update. One wants an eval; the other says the eval misses the systemic risk.

What it hinges on. One belief does the work: have brand-AI tasks actually plateaued on model capability, so that spend should shift from compute to integration? If yes, the Compute Pragmatist's reallocation is correct and the SIs win the next cycle. If a near-term model jump (better agents, native tool-use, longer context that swallows the copy-paste handoffs) collapses the integration problem, then the plumbing spend was premature and the model providers win after all. Before committing a budget to end-to-end rewiring, a buyer should run the cheap test the Builder named: measure cycle-time on one real workflow, find whether the hours sit inside the steps or between them. If they sit between them, Sykes is right. If a tool-using agent can already bridge the handoffs, she's selling you yesterday's problem.

Prediction: Between now and Q2 2027, at least one major model provider (OpenAI, Anthropic, or Google) will ship a general agent capability that handles multi-system handoffs (reading from one app and writing to another without a human copy-pasting) as a headline feature, directly attacking the integration bottleneck Sykes describes.

Confidence: Medium. The capability is already in preview; productizing it is the obvious next move.

Why: Sykes locates the whole problem in the gaps between systems, the manual copy-paste that kept a 70%-automated task at four days. That gap is exactly what agent and computer-use features (an AI that clicks through apps the way a person would) are being built to close, and OpenAI, Anthropic, and Google have all already shown early versions. The commercial pull is enormous, because whoever owns the handoff owns the workflow, which is why the model providers won't leave that ground to the systems integrators. The opposite outcome, that these features stay in demo and the plumbing stays manual, is the less likely bet precisely because the integration layer is where the money now sits, and three labs are racing for it.

Revisit by 2027-06-30: We're right if OpenAI, Anthropic, or Google has launched a generally available agent feature marketed for cross-application workflow automation by then. We're wrong if all three keep such capability in limited preview or research demo with no general availability.