Every big firm’s AI needs weeks of data plumbing before it does anything.

Back in April 2024, asking a client to spend twelve weeks modernizing its data before deploying an AI agent was reasonable. The models were new, the data pipelines were immature, and pointing an autonomous agent at messy source material was genuinely risky. The so-called digital twin, a clean digital model of the business built before any agent runs, was a legitimate precondition. That argument aged out in the eighteen months between then and the April 28 announcement that OpenAI now runs on AWS Bedrock alongside Anthropic, Meta Nova, and Google Gemini. Access to the model is no longer the variable. Setup time is.

The May 9 round of consulting results confirmed the problem the Big Five do not discuss in public. McKinsey’s Kate Smaje printed a 20 percent profit-margin uplift claim from AI adoption. PwC’s Joe Atkinson put a 7.2x performance multiplier on AI-fit companies. BCG’s Sylvain Duranton coined tokenmaxxing as a boardroom metric. Each number is real, and each rests on an assumption the firms do not surface: that the client has already finished the data integration work needed to run the framework that produces the number. The 7.2x multiplier holds if the client’s data is unified, normalized, and wired into the layer that measures the output. The 20 percent uplift is achievable if the meaning layer is pre-built and the agent can reach the right business model. The claims sit downstream of infrastructure most clients do not have, and that the firms bill separately to build.

A technical review of the five dominant frameworks found one consistent thing: every one of them anchors truth in an outside system that has to exist before the agent runs. McKinsey’s Agentic Mesh needs a meaning layer and business models that define how the company works across scattered systems. Deloitte’s Zora AI needs a single source of truth and a digital twin of the business before it can run finance operations. Accenture’s AI Refinery needs its Intelligent Digital Brain integration to unify data across sales and resource-planning systems. Wipro needs data modernization through Foundry IQ. PwC needs a connector server set up before its Agent OS can coordinate across providers. None of these are unreasonable as architecture. They are all unreasonable as a sales pitch when the competitor can run from a raw transcript in under a minute.

Three signals from the May 9 round directly contradict the idea that the digital twin model is holding. First, Wipro’s Bandaru launched a standalone AI-Native Business and Platforms unit with its own profit-and-loss and an invest-build-partner mandate, which is a structural admission that the five-layer data prerequisite is not converting to revenue fast enough. Second, Cognizant’s Kumar paired a $600 million acquisition of Astreya with a Project Leap restructuring the firm calls a broader and shorter pyramid, buying the managed service rather than building the integration layer and compressing the very delivery pyramid the old model required. Third, EY’s Truncale launched a client-zero agentic audit across 130,000 auditors, running on existing workflows and existing engagement data rather than waiting for a digital twin to be built. All three are exits from the setup-tax model, not extensions of it.

The architectural counter is not a better integration framework. It is a different idea of where truth lives. Source-native extraction means the ground truth is the structure already inside the input document, not an outside system that has to be built and kept in sync. The agent runs against the transcript, the earnings call, the contract, or the procedure directly. No business model to build, no data modernization, no digital twin. The practical consequence is measurable: a workflow that needs twelve weeks of data plumbing before it can produce an auditable recommendation does not compete with one that produces the same recommendation from the raw document in the time it takes to upload it. That is a cost-to-first-insight argument, and the CFO, who the May 9 round confirmed is now the primary AI buyer, closes it in one conversation.

Expect the firms to fight back. The leading indicator named it directly: within thirty to forty-five days, consulting firms will start echoing multi-vendor foundation model strategy as table stakes, and they will attach a new reason for the old data engagement, namely that juggling many models across many clouds requires a unified data foundation. The digital twin returns wearing a multi-vendor wrapper. The counter-question to put in front of any buyer reading that briefing is simple: can the firm run a strategic audit from your last board presentation today, with no data infrastructure, and show the path that produced each finding? If the answer requires a prerequisite conversation about data maturity, the setup tax is still the product, and the AI is just the vehicle that delivers the invoice for it. The Wipro restructure, the Cognizant acquisition, and the EY rollout all point the same way: the digital twin is being shortened, not deepened. The setup tax is becoming optional. The extraction advantage is not.