We are turning this engagement into a practitioner playbook: the orchestration architecture, the enterprise-context design, the disagreement and governance patterns, and the economics scaffolding that turns a market signal into an investment-grade decision.
Coming soonA Fortune 500 manufacturer wanted to know whether agentic AI could compete with its traditional new-product innovation process.
The company's normal process took approximately 18 months. A serious new-product opportunity could require consumer insight, category and competitive research, product and industrial design, sourcing, manufacturing planning, pricing, channel strategy, marketing, ecommerce, unit economics, and ultimately a complete P&L.
Much of that expertise already existed inside the organization. The problem was getting it to operate as one system.
Traditionally, work was sequential: a question from one function generated another round of analysis somewhere else. Information could fall through the cracks between teams. Different functions sometimes worked from different assumptions. Decisions stayed largely siloed, leaving the Innovation Office to reconcile inputs across 15 to 20 teams.
The Innovation Office might develop only two or three concepts deep enough to recommend one, while the broader enterprise might make a major investment behind a new concept roughly once every two to three years.
The constraint was not the ability to generate another idea. It was the organizational capacity to research, challenge, economically underwrite, and align around an idea well enough to commit scarce people, manufacturing capacity, and capital.
BCG's 2026 research on large consumer brands identifies this as a systemic problem across the industry. Incumbents frequently possess substantial data, technology, and insight but struggle to prioritize opportunities, create shared conviction, assign end-to-end ownership, and convert signals into scalable innovation.1
The VP of Innovation, working within the COO organization, was not looking for another AI demonstration. Leadership wanted to understand what AI could actually change about the innovation system.
The mandate centered on specific questions. Could AI continuously identify attractive new market opportunities? Could it develop credible product concepts grounded in actual consumer and market signals? Could those concepts be evaluated against the company's proprietary product, manufacturing, marketing, and financial information? Could AI materially accelerate the work required before leadership committed the broader organization?
The benchmark was not another AI system. It was the company's existing way of working.
A conventional approach would have introduced AI function by function. Design gets an AI tool. Marketing gets another. Consumer insights gets a copilot. Supply chain experiments separately. Finance builds its own models. Everyone gets somewhat faster while the same organizational architecture stays intact.
The recommendation here was different.
Centralize orchestration. Distribute expertise.
Functional teams would keep their domain expertise and decision authority. The COO Innovation Office would hold the orchestration layer. Specialized agents would perform increasing amounts of the research, analysis, synthesis, scenario development, and coordination between functions.
The resulting architecture sat between two common extremes. It was not a centralized AI organization attempting to replace business functions. And it was not dozens of employees independently prompting copilots. It was an agentic innovation system designed to connect specialized expertise around a common set of business decisions.
A seven-person multidisciplinary team — spanning AI architecture, product innovation, consumer intelligence, creative, growth, customer experience, and product leadership — designed the operating architecture. The resulting system used more than 40 specialized agents across the major stages of innovation.
Their capabilities covered market and trend intelligence; consumer segmentation and validation; competitive analysis; opportunity identification; product concept and design; sourcing and manufacturing planning; material analysis; bills of materials and product specifications; compliance inputs; direct-to-consumer and channel strategy; media and influencer planning; pricing; demand forecasting; customer-acquisition economics; conversion scenarios; margin optimization; three-year P&L scenarios; and portfolio investment recommendations.
The number of agents was not the important part. The architecture connecting them was.
BCG's 2026 research on scaling enterprise agents reaches a similar conclusion: step-change value requires redesigning the end-to-end process rather than inserting agents into isolated steps. The research distinguishes domain agents from the common orchestration, governance, memory, evaluation, and infrastructure needed to make multiple agents work as a coherent system.2
We designed the agentic system to work with information that normally sits scattered throughout a large company.
Customer segmentation, purchasing behavior, reviews, focus-group findings, social listening, search signals, and cross-brand customer behavior.
Historical catalogs, product performance, category trends, competitive pricing, whitespace analysis, seasonal plans, and previous product decisions.
Materials, supplier information, hardware specifications, factory capabilities, capacity, quality metrics, landed costs, and historical tech packs.
Unit economics, margin requirements, product-line P&Ls, investment thresholds, and historical revenue performance.
Generic AI could generate a product concept. That was not the objective. The objective was to let AI reason about a potential product using information a competitor's model would not have: this company's customers, products, factories, economics, brand constraints, historical decisions, and market performance.
The foundation model was broadly available. The enterprise context was not.
The agents did not operate as independent chatbots. Outputs from one part of the system became structured inputs into another.
The economics architecture illustrates how. A launch recommendation might require: product definition, then pricing, then cost structure, then marketing assumptions, then conversion assumptions, then demand scenarios, then customer-acquisition economics, then margin, then scale economics, then a three-year P&L.
Inputs came from different domains. Product defined the use case and positioning. Supply chain supplied material and manufacturing economics. Growth supplied media and distribution assumptions. Ecommerce supplied conversion and return assumptions. Financial agents reconciled those inputs into scenarios and an investment recommendation.
When a function surfaced new evidence contradicting an earlier conclusion, the system did not simply average the disagreement away. The new information could be pushed back through the workflow. Relevant agents reconsidered the original assumptions, recalculated the economics, and surfaced the trade-offs again.
Humans still owned the consequential decisions. But substantially more of the work required to reach those decisions could happen continuously.
Perhaps the most revealing test was not whether the agents could produce an attractive product. It was whether they could recommend against one.
The system developed multiple AI-generated product opportunities and compared them with a product concept that had been developed through the company's conventional innovation process. The portfolio analysis did not simply rank the AI-generated ideas first.
Instead, it separated the opportunities by strategic role and economic readiness. One AI-generated concept was recommended as the lower-risk market-entry opportunity. The conventionally developed concept was judged to have the strongest long-term revenue potential. Another AI-generated concept was held back because important customer-acquisition, conversion, and economic assumptions had not yet been sufficiently underwritten.
The system was not being rewarded for producing more concepts. It was designed to ask: which concepts deserve investment?
Leadership could compare opportunities using a common view of demand, customer economics, margins, risk, readiness, and three-year P&L — rather than waiting for separate functional analyses to be reconciled afterward.
The clearest result was not the production of an AI-generated product. It was a different way of making innovation decisions.
The AI-generated concepts were well received by leadership, particularly because their recommendations could be connected back to real market signals and enterprise-specific information. Leadership accepted the strategic capability and chose to continue operationalizing the agentic approach as part of the innovation workflow.
Leadership estimated that approximately 12 months of planning, analysis, follow-up questions, and cross-functional coordination could potentially be compressed into a period of weeks.
This was not a claim that an 18-month physical product-development cycle had suddenly become a few weeks. It was a more specific and more important claim: much of the organizational work required to decide what deserved to enter the expensive physical-development process could happen dramatically faster.
Recent company-reported experience elsewhere in consumer products supports the direction of that estimate. Reuters reported in July 2026 that L'Oréal says AI is helping it develop some new products roughly four times faster, while Mondelez says AI has compressed parts of product development from months to weeks and years to months. Those are company-reported outcomes, not controlled benchmarks, but they show that substantial compression in product-development knowledge work is no longer theoretical.3
The experiment also clarified where AI stopped creating the same degree of leverage.
Agents could contribute to product specifications, bills of materials, tech packs, compliance analysis, sourcing options, and preparation for prototyping and sampling. But the physical system still contained constraints that intelligence alone could not eliminate: factory scheduling, factory capacity, tooling, physical material sourcing, physical testing, and quality assurance.
AI had accelerated the information economy of innovation. The physical economy still had its own clock.
The strategic implication was not that AI solved the 18-month process. It was that AI exposed which parts of that process still genuinely required 18-month-era constraints — and which did not.
A second limitation emerged only after the system became faster. Human decision gates began arriving much more quickly. A function that previously had days or weeks between rounds of analysis could suddenly receive an updated recommendation, revised economics, or a new trade-off far sooner. The organization now had to make decisions at a cadence its traditional operating model had never required.
AI increased the supply of decision-ready analysis. It had not automatically increased management's capacity to absorb and act on it.
Governance also became more consequential. As agents challenged assumptions, incorporated new evidence, and revised prior conclusions, leadership needed to understand which data had changed, which assumptions had shifted, which agents had responded, why a recommendation changed, and where human judgment had entered the process. The audit trail was not administrative overhead. It was part of the operating architecture.
The alternative would have been much easier. Give product teams copilots. Give marketing AI tools. Let consumer insights automate research. Let the supply chain experiment separately. Everyone becomes individually faster.
Yet the company's core problem would remain. The same information would still sit in different functions. The same handoffs would still exist. The Innovation Office would still reconcile competing assumptions. And the investment decision would still depend on assembling the complete picture afterward.
Individual productivity does not automatically create organizational throughput. The advantage came from redesigning how the work connected — so the unit of AI transformation was not the employee. It was the innovation system.
The company began with a technology question: how much can agentic AI actually do? It ended with an organizational question: how should innovation work if analysis and coordination that once required months can happen in weeks?
The emerging model was straightforward. Functions own the expertise. Humans own the consequential decisions. Agents perform and challenge the analysis. The Innovation Office orchestrates the system.
The strategic advantage was not 40 agents. It was lowering the organizational cost of getting from a market signal to an investment-grade product decision.
That changes more than cycle time. It changes how many opportunities a company can afford to investigate before committing scarce people, manufacturing capacity, and capital — which may ultimately be the more consequential source of competitive advantage.
We are packaging the orchestration architecture, the enterprise-context design, and the disagreement and governance patterns from this engagement into a practitioner playbook. Coming soon.
Widespread AI use, no material change in innovation throughput. Employees get faster. Functions generate more analysis. More ideas surface. But the same organizational machinery still decides how quickly an idea becomes a funded product opportunity.
Functions keep expertise and authority. A 40+ agent layer researches, challenges, and economically underwrites concepts against proprietary data. Leadership compares opportunities on one shared view of demand, economics, risk, and P&L.
The risk was not failing to adopt AI. It was putting AI inside a process designed before AI existed.
The unit of AI transformation is not the employee — it is the system the work moves through. Redesign how consumer insight, product, manufacturing, marketing, and finance connect around a decision, ground the agents in your own data, and the cost of getting from a market signal to an investment-grade product bet drops sharply.
We help COOs, innovation leaders, and product leaders design AI operating models — the agent architecture, the enterprise-context layer, the decision rights, and the governance — then run them against real work.
Start a Conversation