Your AI Is Working. Is the Process Behind It Working Too?

A PE-owned food packaging company used AI to pressure-test ideas during a two-day workshop and landed a 58-page strategy book on the PE firm's board table. An employee-owned dairy company ran three versions of the same innovation sprint and cut the process from five days to 2.5. Canada's largest transit agency embedded the same tool inside its stage-gated pipeline and pushed 25 pilots through faster than any Excel tracking ever allowed.

All three presented at Innov8rs in Toronto this June, each walking through how they added AI to their innovation process. And all — despite different outcomes — arrived at the same conclusion: the competitive edge isn’t AI itself, but the process around it: where the tools enter, what stays with people, and who owns the output when it reaches a decision-maker who isn’t in the room. This process work is exactly what gets ideas funded, killed, or scaled.

Today, most innovation teams still turn a blind eye to this. AI gets dropped into the existing workflow, the team runs the same sprint or the same pipeline with a new tool bolted on, and the output gets faster — but not stronger. The funnel fills up with more ideas, more slides, more concepts, and none of it changes the quality of what reaches leadership, the speed at which decisions actually get made, or the trust that the work holds up under scrutiny.

The three companies above recognized one thing early enough: no matter how many AI tools you adopt, if the process stays the same, so will the results. What each of them did differently with their own process, and what it produced once they did, is what follows.

Last year, Matt Parthasarthy, Director of Strategic Marketing at Novolex, sat through three AI-enabled ideation workshops. Each time a workshop finished, he came away with worse strategy than the old-fashioned whiteboard. AI generated plenty of ideas, but of very poor quality. Instead of making things better, it was creating more bottlenecks than before.

Novolex is owned by Apollo Capital, and the output of any strategic planning travels to the PE firm's board. When the company's Chief Innovation Officer proposed using AI for this year’s strategic planning process, the stakes were too high for a repeat of the previous failed workshops.

So Matt and his team made a design decision to avoid falling into the same trap: humans would generate the ideas, and AI would challenge them. Matt spent roughly nine months building a process around that division of labor. The split came from an early lesson during the workshops, when some smaller teams used AI to generate ideas, and watched the AI go down rabbit holes — producing 20 to 40 concepts from two or three narrow areas. Although the output looked impressive in a spreadsheet, it was useless in a planning meeting.

So that experience shaped every design choice that followed. The process, split in four phases, ran over two days in Cincinnati. Around 50 people were there — innovation directors, product managers, product engineers, and commercial leads from Novolex's major retail channels (QSR, food service, grocery, mass merch).

The first morning, ten megatrend briefs were posted as large banners in a gallery walk. Each brief followed the same structure: a Then/Now/Next trajectory showing where the trend came from, where it sits today, and where it's heading in five years, backed by 3 to 5 hard data points, the macro drivers accelerating the trend, and the specific risk to Novolex if the company gets the response wrong. The ten trends were organized into social, technology, and environmental clusters, selected based on signal strength across multiple research sources.

By midday, the teams moved into problem definition, the phase Matt feels that “workshops typically collapse”. Each sub-team took two megatrends and had to write a specific, ownable problem statement, before anyone was allowed to propose a solution. For example, "The aging population is growing" is a vague statement. "33% of our packaging touches food consumed by people over 65, and none of it was designed for reduced dexterity" is a problem a product team can solve. Another example that passed the test: "EPR fees will penalize untracked materials in 30+ states by 2027, and we have no traceability infrastructure." The specificity of the problem statement determines the quality of everything that follows.

Only after the problem statements were locked did teams begin brainstorming solutions. The teams scored their own ideas on Impact (1 to 4) and Effort (1 to 4) before any AI touched them. Then, each scored idea, with its problem statement and rationale, went into an AI tool the team had configured as an adversarial pressure-testing engine.

AI challenged the teams' assumptions, flagged contradictions, and used synthetic consumer personas to test whether the proposed solutions would hold up against real market conditions. Teams had to respond with evidence (customer data, market size, Novolex capability, competitive rationale) or abandon the idea. Ideas that couldn't survive the pressure test were cut.

The surviving ideas moved to a Shark Tank-style pitch on Day 2. Each team presented its strongest idea to Novolex's three innovation directors, covering the problem, the solution, the target customer, why Novolex wins, and the resource ask. The Q&A followed, also allowing participants to express their opposition to the idea, and rankings happened in an open session. Eventually, five ideas were pitched. Matt expects three to be championed and carried into the strategic plan this year.

Those two days produced a 58-page book called "Markets of the Future," with 10 megatrend chapters, each structured as a strategic argument. The book was distributed to senior leadership, the board, and Apollo Capital, and was designed for external distribution to customers. The output feeds directly into the annual strategic planning cycle, which begins three months after the workshop, giving innovation directors time to build out the ideas before committing them to a three-year plan.

Novolex's process had four pieces of discipline in place before AI entered the picture: problem definition before ideation, self-scoring before external input, public defense in front of judges, and a predetermined home in the strategic planning cycle. Each of those disciplines shaped what the AI tools received as input and what it was asked to do with that input. By the time an idea reached the AI, it already had a sharp problem statement, a team-assigned score, and a rationale the team was prepared to defend. The AI simply amplified a process that was already rigorous. Two days, 50 people, and one AI platform later, Apollo Capital's board had a 58-page strategy book on the table, with three ideas moving into a three-year plan.

Melissa Pierson, Innovation Programs Manager at Schreiber Foods, started from a different place. Schreiber is an employee-owned dairy company based in Green Bay, Wisconsin, with manufacturing and distribution across five continents. Melissa's team runs design sprints combining elements of Google Sprint, Jobs to Be Done, and Strategyzer into a five-day process that starts with empathic interviews, moves through pain point immersion and JTBD statement development, tests ideas overnight with real consumers, and ends with an open house where C-suite executives walk through the work in progress.

Up to now, the fully human sprint worked. According to Melissa, it builds true innovation culture: Schreiber tracks a 95% repeat rate (the percentage of participants who would sprint again if asked) and NPS scores well above industry standard. Sprint participants leave the room with genuine ownership of the ideas they develop.

With that in mind, Melissa and her team ran an experiment. They took the same five-day sprint process, rewrote each step as AI prompts, and ran it almost entirely on AI. The AI sprint was fast and resource-light, and the ideas that came out were bolder than anything the human sprints had produced. Nevertheless, nobody in the room felt like the ideas belonged to them.

Melissa's team identified three things that disappeared when AI took over:

The ideas from the fully AI sprint had no champion to fight for them. When it came time to assign ownership, nobody stepped forward. The ideas were technically strong — but organizationally orphaned.

Rather than completely abandoning AI, the team designed a third model: a hybrid sprint. As Melissa says, "We went back to the AI and said: you took the soul out of the room. Help us fix that."

The hypothesis was that the team could run a full sprint in a fraction of the time, with fewer people, by designing a joint human-and-AI team, as long as two conditions held: equivalent output quality and equivalent culture impact. The hybrid sprint runs over 2.5 days instead of five, and the design centers on which parts of the process stay human and which parts move to AI.

Empathy stays on the human side. The team conducts real interviews with consumers, then has sprinters sit side by side with AI-generated personas and ask the same questions. When the personas return the same answers the real interviews produced, trust in the AI tool goes up.

For example, it synthesizes the pain point data into a six-minute video that kicks off Day 1, regrounding the team in the customer's experience. Facilitators build in pause points throughout the sprint: moments where the team stops and asks what solving the problem would mean for the specific people they interviewed.

AI also handles the tasks that drain cognitive energy without adding insight. Writing JTBD statements, theming large volumes of pain point data, developing stimulus material for ideation sessions, and opportunity sizing all move to AI. The innovation team also built a standardized prompt library so that every facilitator across the company uses the same validated prompts.

Finally, AI enters the sprint at the canvassing stage, where teams map out what must be true to bring an idea to life. The tools host consumer personas that sprinters can interview in real time. If a team starts drifting on a product concept, the facilitator pulls up the persona and asks how that consumer thinks about protein content when choosing a product. The persona's response pulls the idea back to the original need.

So far, the hybrid sprint has delivered real results: time has dropped from five days to 2.5, output quality matches the fully human sprint, and culture impact matches too. Melissa tracks conviction through three metrics:

The say/do ratio measures who claims ownership of an idea at the end of the sprint versus who leans back. That single moment tells you whether the sprint worked.

The 30-day survival tracks whether the idea has moved forward a month after the sprint ends. And the watered-down effect measures whether the idea's core value proposition stays intact as the idea travels through the organization or gets diluted to make it palatable.

An example of what conviction looks like at Schreiber: during a hybrid sprint, a participant felt so strongly about a product idea that she went to the R&D pilot lab that evening, developed a formulation, drove home to get a piece of equipment she needed, and brought a working prototype back to the sprint room the next morning. That kind of ownership never showed up in the fully AI sprint, which used the same platform and the same methodology.

Schreiber ran three sprints on the same tools with the same innovation methodology. When the process handed everything to AI, that amplified the absence of human ownership: bolder ideas, zero champions. When the process kept humans in charge of empathy, sensemaking, and decisions, the AI amplified the discipline those humans brought: equivalent output in half the time, and people who kept building after the sprint ended. The tools performed identically in both versions; but it was the redesigned process that determined whether AI was making good work sharper or making shallow work faster.

Naina Dewan, Manager of New Technology and Innovation at the Toronto Transit Commission, took a more conservative approach. She and her team dropped AI inside the stage-gated pipeline the organization had already built.

The TTC set up its innovation function about two and a half years ago. Naina, who previously built innovation labs at Walmart Canada, Loblaws, and Scotiabank, designed the function with a board-approved Innovation and Sustainability Strategy (2024 to 2028) and a structured pipeline: idea sourcing from five channels (innovation challenges, workshops, trend scanning, leadership tables, and unsolicited vendor proposals), followed by initial screening, research and exploration, evaluation and business case development, pilot execution, and a go/no-go decision before scaling.

The pipeline has clear gates. Concepts scoring below 80% on desirability, feasibility, viability, and strategic fit don't advance. A business champion is required at the evaluation and business case stage. Without that champion, the idea has no organizational support and doesn't move forward.

In TTC’s case, the problem was tracking the sheer volume of ideas. As the pipeline matured, the team was tracking them across Excel spreadsheets, Miro boards, and email threads. Ideas didn't have IDs. Related concepts lived in different files. At 80+ active concepts across 35+ business units, the job was becoming unmanageable.

Therefore, AI plays a different role at the TTC: pipeline infrastructure. What used to be scattered across spreadsheets and boards now runs through a single system. Every idea gets an ID. Related concepts get clustered automatically. For the first time, Naina's team could show leadership exactly where every concept sat at any given moment.

The TTC's innovation function currently has 25 pilots in flight, 2,500+ engaged employees, and $5M+ in projected benefits and ROI. The pipeline already had the discipline to filter and prioritize before AI showed up. Because someone already has to stake their name on every idea that advances, AI can focus on what it does well at scale: clustering related concepts, flagging duplicates, and running market due diligence across dozens of ideas at the same time.

That combination of human sign-off and AI infrastructure is how Naina's team runs 25 pilots and 80+ active concepts with a team that would never have managed that volume on spreadsheets.

The problem most innovation teams are running into in 2026 is that AI adoption happened, but process design didn't follow. Teams added AI to workshops, sprints, and pipelines without redesigning how those processes work with AI inside them. The result is more output at every stage — more ideas scored, more concepts generated, more decks produced — but with no substance or scaling potential.

The three companies in this piece each took a different approach to the same question: where does this tool sit inside our existing process, and what does the process need to do differently now that the tool is there? Novolex spent nine months redesigning a strategic planning workshop around a single principle: AI challenges, humans generate. Schreiber ran three versions of the same sprint until the process design produced both speed and ownership. The TTC left its stage-gated pipeline intact and gave AI the infrastructure layer underneath it. The principle was the same: design the process first, then give the tool a role inside it.

The results: a 58-page strategy book traveling to a PE firm's board, a 2.5-day sprint producing the same output and culture impact as a five-day sprint, and 25 pilots managed across 35+ business units with $5M+ in projected ROI, all shaped by the process work each company did around the AI.

None of these companies used tools other innovation teams don't already have access to. The difference was that each one looked at the process it already had, figured out what needed to change now that AI was inside it, and made those changes before expecting better results.

The front of the innovation funnel has never moved faster. Research that took weeks now takes hours. Early-stage validation is cheaper and quicker than it has ever been. But what about the funding decisions, the kill decisions, the scaling decisions?

A new benchmark from Disruptive Edge, Aucctus AI, and Innov8rs, covering 119 innovation professionals across 20+ countries, puts numbers on what many teams are already feeling: 97% of innovation teams use AI, yet half report no measurable change in cycle time and fewer than 1 in 10 bring AI anywhere near the portfolio decisions.

The full report shows where the breakdown happens and what the teams pulling ahead are doing differently.

The sharpest finding in the report is how AI usage collapses across the innovation lifecycle. 88% of teams use AI in research, sensing, and discovery. That figure drops to 44% in experimentation and testing, and to 18% in scaling and go-to-market. From the first signal to market entry, AI usage falls by a staggering 70 percentage points.

The reason is straightforward: early-stage work is low-cost, individually driven, easy to redo — whereas scaling decisions commit capital, require cross-functional alignment, and carry consequences for the team's mandate. Most organizations have handed employees AI access without changing the governance, evidence standards, or decision processes that later-stage work depends on.

The result: the front of the funnel accelerates while the middle and back end run at the same speed, through the same bottlenecks.

Teams generate more research and early validation faster than their organizations can evaluate, fund, or advance any of those opportunities. AI can make the oldest problem in corporate innovation worse by increasing the volume of ideas without improving the organization's ability to move the right ones forward.

The report identifies four hurdles that keep teams from progressing along the Innovation Intelligence Curve, a four-stage model of how deeply AI is embedded in the innovation process: 42% of teams sit at Stage 1 (Individual Assistant), 47% at Stage 2 (Structured Workflows), 11% at Stage 3 (Proactive Intelligence). None have reached Stage 4 (Self-Learning Innovation). Only 4% of respondents cited executive skepticism. The real barriers sit in how the work around AI is structured:

Teams that have embedded AI into shared workflows, decision points, and operating cadence (Stage 3) are 6.2x more likely to report 25%+ cycle-time gains than teams at Stage 1. Among those embedded teams, 69% use AI to inform kill-or-continue decisions, compared with 20% of teams still operating at the assistant level. The same embedded teams report 2.8x fewer data quality or access barriers.

Changing the operating model matters as much as choosing the tools. Teams that changed at least one structural element alongside AI adoption (roles, cadence, stage gates, or KPIs) are 4.5x more likely to report meaningful speed gains. Among the deepest adopters, 75% changed their operating model, 50% use AI in decisions, and 63% measure AI's impact formally.

The report introduces the Embedding Method, a four-step process the strongest teams followed:

Schreiber Foods followed this sequence. The team mapped an innovation process that was already working well, identified where AI could play the strongest role within those workflows, and embedded AI into them. Cycle times dropped by more than 25%, and delivery of new products and business improvements increased. The gains came from augmenting a working process.

Mars embedded AI across its core innovation workflows, from brand intelligence to commercial assessment, and now treats AI use as an expectation. When a use case proves out, Mars redesigns the workflow around it, turning individual experiments into repeatable capability. Colgate-Palmolive scaled advanced AI adoption to over 50% of its white-collar workforce by deploying 240 cross-functional ambassador leads and integrating responsible-AI principles into its code of conduct.

EllisDon, a $10B+ construction firm, took a different angle. Rather than restricting employee experimentation, the company bought enterprise licenses, built sandboxed environments walled off from live project data, and published a governance framework. Tasks that once took four hours now take 30 minutes. Individual initiative became a governed, shared capability.

The full report also covers the Innovation Intelligence System architecture (a five-layer blueprint for connecting workflows, shared intelligence, and a continuous learning loop), a 70/20/10 investment model for balancing proven applications with emerging pilots and longer-term system investments, six cultural principles that appeared consistently in the teams furthest along, and additional case studies from Toronto Transit Commission and Beem Credit Union.

Download the full report free at disruptiveedge.com/the-next-frontier-in-corporate-innovation. We’re hosting an online discussion about the findings on Sept 22nd, 19:00 CEST/1pm EDT/10am PDT - RSVP for that session (it’s free) here.

Beyond the hype, I hope today’s piece helped you get real about integrating AI in innovation management.

What's one thing in your process AI still isn't allowed near — and who actually made that call? Hit reply and tell me.

In the next edition, we’ll look at the messy middle of the innovation process, which may in fact be causing the (increasing) gap between innovation ambitions and actual impact that exists in most orgs.

Hans Balmaekers
Founder, the Compass and Novum

PS- We’re keen to get your feedback about the Compass: what’s valuable, what do you like, and what’s missing? Please complete this short survey to share your input. We’ll review and work with all input to make this newsletter one of the few you actually look forward to receiving. Thanks!