AI agents run the early stages of innovation work in hours, not months. But one question keeps innovation leaders up at night: when do you stop optimizing the funnel you have, and rebuild it from scratch?

A hundred concepts generated before the coffee's gone cold. Prototypes by afternoon, decks before lunch. A year ago none of this was possible: the AI models were too shaky to hand real work to, and any output needed so much cleanup you might as well have done it yourself.

That changed. A team can now do more in a week than it once managed in a quarter, and every stage of the funnel runs faster than a team can work through it.

Innovation teams will eventually rebuild around agentic AI. The real question is not whether to tear up the funnel, but when to: rebuild now, or keep speeding up the one in place for the time being?

The answer depends on your context. Some have room to spin up an experiment without asking anyone, while others have to raise an IT ticket and wait three weeks for a resolution.

Below, one team rebuilt from scratch around agents because it had nothing to protect. The other kept its funnel standing because 153,000 people depend on it.

Ben Lewis founded three companies and spent twenty years working at the crossroads of human, animal, and planetary health. Before that, he was an Olympic kayaker, with training from Dartmouth and Wharton along the way. He had ticked off every box on the dream-job checklist, sold his last company, and was done building.

Then he saw what a venture studio could do: instead of running one company for years, he could build many at once, shaping each idea, bringing in a team, and handing it off before starting the next. That sparked curiosity.

Today, Ben runs One Health Studio at Alloy Partners. Alloy builds ventures for large corporate clients, and One Health Studio is the one it launched for Elanco, the animal-health company that funds the work. The studio generates concepts, tests them hard, and drops the ones that don't hold up, handing the survivors to an entrepreneur in residence.

Most of the concepts in the pipeline started in Ben's own head. The first fifteen are ideas he carried for five or ten years and never got to build. As a single-company founder he was always too deep in one thing to chase the rest, and no investor wants to watch a founder run fifteen companies at once, so they piled up. The studio is the first place he can work through the whole backlog at once. He shapes the core of each idea, runs the deep research, builds the nucleus. For the first time, none of them has to wait.

But recently, Ben has been hitting a wall. What happens when you design the company, but you can’t vet the concept, or staff the company, at the depth each one needs?

Both gaps come down to the same thing: the studio needs expertise it can't afford. Every concept needs a top reviewer to pressure-test it; the kind of expert who kills the weak ideas before they soak up all the budget. Every surviving concept then needs a full team to build it out. In reality, most studios can't put a world-class reviewer on every concept or staff a team fifteen times over.

Instead of hiring that team and spending months looking for experts, the studio builds it out of AI agents.

The method reconstructs a whole company, with an org chart, defined roles, and reporting lines, so the agents work like a real organization rather than a set of disconnected tools. The method has five steps:

The org-chart-of-agents idea traces to James Zou, a Stanford professor whose "Virtual Lab" assembles a team of AI scientist agents that hold their own research meetings, debate each other, and report up to a lead agent. Zou's lab used it to design new antibody candidates against recent COVID variants, which were then validated in the wet lab.

By Ben's account, the lab now generates the overwhelming majority of its output through the agents, with roughly one part in a hundred coming from the human researchers. The studio runs the same pattern, pointed at building companies instead of biology.

The studio runs around 2,000 agents, arranged in roughly seven layers of reporting - just like a real company has junior staff reporting up to managers, and managers up to directors. The trick is feeding the agent a very small amount of information.

A single agent covers one specific topic and nothing else; the next covers the topic beside it. An agent with too much to hold overruns the model's working memory, and its answers get less reliable. By keeping the scope narrow, each agent returns work that's about 90 percent ready the first time.

Also, because everything the studio produces feeds back into the agents, emails, messages, and meeting notes alike, the company's knowledge stops living in one person's head.

When a founder is out sick, or eventually leaves with the spun-out company, the reasoning behind each decision stays in a record anyone with the right access can query later.

The institutional memory that normally walks out the door, becomes a canon the next team can build on.

Those experts the studio couldn't hire now work as an AI review panel. Take a new drug concept, for example: one agent carries the FDA's drug-evaluation rules, another holds ten years of patent filings, the record of what's already protected in that drug class, and a third pairs a biochemical-engineering PhD with a law degree.

Sent through the panel, the concept comes back with its business-model, patent, manufacturing, and regulatory problems flagged, before it reaches a human reviewer or the incubation committee that votes each concept up or down. That review would take one person a whole quarter; the panel does it in an afternoon.

The review doesn't lean on a single model. Several run at once and argue, an approach the AI researcher Andrej Karpathy has described: each attacks the same concept and looks for reasons it will fail.

The models often disagree, and those disagreements catch flaws a single model would have missed.

In another example from the studio, three of the four models read a patent in the field as blocking the concept, and flagged it as a conflict. The fourth read the same patent and found a legal path around it, which the studio's human lawyers later confirmed as correct. Left to a single model, or a single reviewer reading it the same way the three did, the concept would have been dropped over a patent that turned out not to be a problem.

The agents also keep working overnight, on a daily cycle:

The nightly review also catches problems early, resolving them while they are still small enough to fix.

With a week of work waiting on every venture by morning, the constraint is not money or talent, but how fast a single human can review what the agents produce.

With so much work done overnight, Ben now feels like the bottleneck to the agent's output — rather than the other way round. He does, however, have three hard rules on what humans should still remain in charge of:

Critical numbers and regulatory claims get checked by hand before anyone builds on them.

Agents draft emails but do not send them, to avoid leaking confidential content.

Humans should still originate genuinely new concepts. The best source continues to be a person who has lived with a problem for years and is actively hunting a way through it, which is why a pipeline starts in human heads and points the agents at everything that comes after.

The studio now spends about $2,000 a month and gets back the work of 100 to 300 people. Hiring that many people would cost far more, and a desired outcome would take far longer. Now, every venture starts with its full team from day one, and a concept becomes a Series A-ready business plan within days.

That is a full company's worth of work for less than the cost of a single junior hire.

In another example from the studio, an agent went digging and found GFI 261, an FDA Center for Veterinary Medicine guidance document that governed whether the concept could ever reach the market. The agent flagged the specific provision, explained why it bound this concept, and laid out what it meant for the path to market. A human would have struggled to get there: tracking down a single buried guidance document means weeks, sometimes months, of a specialist combing through filings, and being paid hefty amounts. Also, that kind of detail usually surfaces too late, after the budget is committed and walking away from the concept costs the most.

No one at the studio knew the document existed. The agent surfaced it, a quick manual check confirmed it was correct, and the studio had a clear answer on whether the concept would survive before spending real money.

Overall, the method runs because the studio has nothing to protect. It can point agents at anything without risking a system that thousands of employees depend on. A bank, by contrast, cannot let agents reach across internal systems without creating a security exposure the moment they connect. "The freedom actually belongs to the studio, not to the method," Ben says.

That freedom is also why Ben can move first. He has no old process to protect, so he can rebuild from scratch straight away.

For a company whose systems thousands of people rely on, the choice is harder. ZF is not ready to tear up its whole funnel yet — as it risks 153,000 employees losing their footing. Instead, it leaves the funnel standing, builds agents into it, and watches out for the signals to push further.

ZF Friedrichshafen is a major automotive and industrial supplier: 153,000 employees and 3.3 billion euros a year are spent on research and development. Inside its commercial-vehicle division, Chi Zhang, AI Innovation Engineer, and Maarten Korz, who heads the division's innovation team, build agents under the constraints Ben does not have: confidential engineering data, an IT department that has to approve what they touch, and real fallout if an agent reaches the wrong system.

ZF keeps the agents away from any sensitive information. Confidential work runs on an internal, approved version of Copilot, while everything else, such as trend-watching or building customer personas, uses only outside data. To try a newer model, the team carves out a walled-off sandbox with IT rather than wait for blanket sign-off. ZF runs on Microsoft, so the bet is that whatever they prove out will eventually run on approved Microsoft tools. Their job is to test what's coming before it arrives and show where it fits.

Their innovation funnel runs through foresight, ideation, and acceleration before a concept becomes a project. Recently, they were noticing that early stages of the funnel were leaking: faint market signals were hard to spot and harder to track over time; real customer interest stayed murky until it was almost too late to act on; ideas piled up faster than anyone could judge them; and whether a real business sat behind an idea often went unanswered until far too late. These were the problems the team set out to fix with agentic AI.

Their answer is a set of agents they call InnoAI, one for each job in that early stretch: trend analysis, startup scouting, patent analysis, persona building, workshop facilitation, idea rating, competitor analysis, business design, and more. Each agent is a separate, swappable service, so when a better tool or model arrives, the team can replace one piece of the system without rebuilding the rest.

What matters more than the agents themselves is the connective layer underneath them. Real work rarely moves through the agents one after another in a fixed order; they have to hand off to each other as the task takes shape. So a coordinating layer sits on top and decides, step by step, which agent to call next and what to pass it. In a trend-analysis task, that layer chooses each next agent on its own instead of running a pre-set sequence. That is what turns a set of single-purpose tools into something that works like a team.

ZF didn’t get to this model in a day; they’ve been constantly iterating. It all began in 2023 with ChatGPT, moved to an internal Copilot once confidential data was involved, and by 2025 ran custom made agents in Python, with AI threaded into Mural, SAP, Office, and the idea platform itonics. In 2026 the team started piloting outside platforms, including agent-run startup scouting, and testing the newest agent frameworks, with one rule: nothing touches real work until it clears ZF's security review.

The best example of all this working together is a pipeline the team recently built by linking four agents into a single chain. It takes the messy front end of the funnel, raw market noise, and carries it through to a shortlist of vetted ideas, each agent handing its output to the next:

What comes out is a shortlist the agents have already screened, brainstormed, scored, and checked against existing patents. This work would normally be split across several people and several tools; the agents do it in a single run.

Across all of this, ZF keeps people at the start and the end, with agents working in between, a model Maarten calls a sandwich. A person sets up the problem; the agents do the work in the middle; a person makes the final call. Wherever one agent passes its output to the next, a human checks it first and decides what moves on. Some of that judgment now sits inside the agents themselves. The workshop agent, for instance, generates a batch of ideas and then shortlists them, based on how the team runs workshops in the room. As the models improve, Maarten expects the agents to take on more of the middle work, while people keep control of the setup at the front and the decision at the end.

Creating the agents was never the bottleneck. The most complicated part is corporate IT, which has to keep the company secure. Maarten compares this challenge to the early days of the web, when companies kept blacklists and whitelists of the sites staff could open; the way forward then was to carve out sandboxes where the team could test new models without putting anything at risk.

Education is equally as important for innovation leaders trying out agentic AI tools. At ZF, most of the workforce already uses AI at home, and the division runs an internal "AI program house" that trains them, with onboarding required before anyone can build an agent of their own.

From their journey so far, the ZF team share four golden rules for using agents in your innovation funnel:

Until now, the agents have run inside the process ZF already had. The question the team keeps returning to is whether that's enough. Speeding up a process built for people is useful, but it may not be the real prize. The bigger one could be a process designed for agents from the start. So they are weighing a clean sheet, built around how agents actually work, not how people have always done it.

Take this example: engineering design. Today a person draws the first concept, working from the constraints in their head. An agent handed the same goals and constraints can open up options nobody would have drawn. The same goes for generating ideas. The workshop agent runs SCAMPER, a brainstorming method built for people. Agents work differently; their version might drop SCAMPER completely and build its own way of generating ideas. At that point the team is no longer speeding up the old process, but building a new one.

ZF can see where this leads in its own products. Its commercial-vehicle fleet platform, ZF Scalar, could run the same way the innovation funnel now does: a fleet-manager agent coordinating specialist agents for fuel optimization, predictive maintenance, and driver-safety monitoring, orchestrating the whole thing itself. That is precisely the difference between using agents to speed up people's work and handing them a whole domain and letting them run it.

The work gets done while the office sleeps. Hand the funnel to agents and they run it faster than any team possibly could. Each stage clears overnight, and by morning a week of output is waiting on every venture. That is the early version of the future, and it has already arrived.

Whether to redraw the whole process is not the real dilemma — every team will get there eventually. The question is when. Ben started from a blank page because he had nothing to erase. ZF kept its funnel because wiping it would cost 153,000 people their footing, and built the connecting layer instead.

Speeding up the current funnel buys time, but it’s not the finish line. The advantage goes to those who take the safe path and read the signals for when to level up, leaving themselves room to get it wrong and fix it later.

Michael Munz runs new ventures at Thieme, a medical publisher with the same constraints ZF works under: confidential data, an IT function that guards what gets touched, no blank page to start from. He could have waited for budget, headcount, and IT sign-off. Instead, he kept the scope to what he controlled completely, his own work, and started there.

Michael shares four moves on how to build a personal agent setup with no team, no budget, and no IT ticket:

Most AI work is still happening one conversation at a time. A prompt resolves today’s problem, then evaporates. The next day, a new task opens in a blank window, with no context to build on. That is the ceiling on what a single person can produce; while the tools are powerful, every session still starts from scratch.

To get over this challenge, give the AI a “permanent home”; a place where it works 24/7, and where every task complements the last, instead of disappearing.

That “home” is built from four parts:

One of the saved routines circulating online is the so-called “research skill”. A single command (/research) sends the agent off to gather and analyze sources on a chosen topic, dig to whatever depth is asked, and format the findings the way they are needed. Then, it saves them straight to the right folder, with no step-by-step hand-holding. For example, Michael runs thirty such routines every day, which get more intelligent by the day. Over time, the setup becomes accustomed to how its owner works.

Your first reusable skill: select a task you run most often, set up a terminal agent (Claude Code is one option), point it at a single project folder, and save that task as a reusable skill it can run on one command. Write down the results of your experiment; how much time have you saved?

Trends, papers, patents, and posts arrive faster than anyone can read them. The feeds that should help, LinkedIn primarily, are crowded with hype. Finding genuinely useful items nowadays is practically impossible.

The answer to this is a filtering system that reviews all the sources but only lets the useful ones through, in a set order. The filtering begins with human selected sources: a short list of trusted feeds (newsletters, podcasts, research papers, influencers), so most of the worthless noise never makes it in.

The feeds collect in a single inbox. As each item lands, it is tagged with where it came from and who wrote it, so any idea can later be traced back to its source. The trail matters for accuracy and fidelity reasons.

Next, an AI model reads the whole inbox and keeps only what relates to the topic in focus, cutting the sources down to a handful. At this point, the AI’s job is very specific: surface the relevant material and label its source. What each piece of information means and what to do next is decided by a person.

Going deep on one of those topics is a separate job, handled by the so-called research routines. A research routine is a saved, repeatable process that takes one question and runs a full investigation on it, start to finish. Essentially, it is the same kind of skill with step one, built this time to dig deep. For example, one can be set up for patent research, another for sizing up a market.

The most elaborate of these skills is a deep-research routine, one that should be custom-made depending on the innovator’s needs. Michael built his entirely in-house and can easily be implemented by others:

Your first signal collection: list the five sources you trust most and route them into one place. Then ask an AI model to filter that feed down to only what touches the single topic on your desk right now, and check that each item still points back to where it came from. Now look at what is left: how much of the feed got cut, and which of your trusted sources mostly added noise?

It is crucial to set up a system that picks up the few ideas worth pursuing and dropping the rest early on.

A saved routine, like the ones in the steps above, handles the first major cut. Once a week, it reviews everything captured over the previous seven days, checks what is trending on GitHub and other relevant networks, and returns a shortlist of five to ten ideas worth a closer look. Michael gives every idea on that list a time-to-value score: a personal rating of how much effort it is worth, set against the payoff it might bring. When he is unsure how to rate, he makes a quick gut check with “the hammer test”.

The hammer test asks a single question: would this idea still be worth doing if AI models suddenly became ten times more capable? If the next generation of models could simply do the thing itself, there is nothing left to build a product around, and the idea is a dead end. Eventually, the ideas worth saving are those that hold their value even as the technology races ahead.

The hammer test. Run this test before committing real time to an idea. For each idea on your shortlist, answer three questions:

How you answer these tells you whether the idea is worth pursuing or worth killing, before it takes up serious time.

Scoring tells you which ideas deserve effort, but not whether they hold up. That is the final job: take the few that cleared the cut and push them far enough that you can actually judge them, then make the call yourself.

This used to be where ideas went to die. Building a prototype, sizing a market, checking the patent landscape each took weeks, so most ideas were decided on a hunch and a slide. That has collapsed. The same agent, repo, and skills that filtered the signals now do the work downstream, so a surviving idea can go from shortlist to something real in a day. The agent handles both moves, and Michael judges the result.

First, it builds a rough working version. Execution is cheap now, so instead of arguing over a one-pager you have something to look at by the afternoon, which usually shows you things the write-up missed.

Then it turns that into a presentation. Michael set up an output engine with his company's branding loaded in, so the agent outputs slides or designs that are ready to show, instead of him formatting them by hand.

Michael’s whole setup took around a year to grow — but one project repository, one skill, or one filtered feed is a day’s worth of work. Michael's advice is to build a small version first, show people what it does, because a personal setup that already runs makes the case for wider organizational adoption. In his words, "it's much easier to convince leadership with a running setup than just telling them what it does."

Most innovation teams are still using AI as a glorified assistant. The ones pulling ahead have figured out that this is not enough to survive.

Disruptive Edge, Aucctus, and Innov8rs surveyed 118 corporate innovation teams, backed by more than ten interviews with innovation leaders, to build the first benchmark of AI integration across the innovation function, measuring actual behavior to find out what separates the teams seeing compounding returns from the ones still waiting for them.

The average team uses AI in only 42% of their innovation activities, mostly in market research (73%), competitive intelligence (58%), customer insight (55%). Nearly every team member (96%) uses general-purpose AI assistants (ChatGPT, Claude, or Copilot).

Only 7% of innovation process areas have AI genuinely built in. Most importantly, 78% of teams have yet to embed AI deeply into even a single part of their process. That's nearly 4 in 5 teams still running AI on top of processes designed for people — one prompt at a time.

The report measures that depth on the Innovation Intelligence Curve, a four-stage model of how far AI has moved into a team's work.

Most teams currently sit in the first two stages. No team in the sample has reached the fourth.

Governance and policy rules are the top barrier to real AI implementation, cited by 26% of the participants, ahead of skills gaps at 17%. And 54% admit to “shadowing AI”, using unapproved tools to get around the blockers. These teams are what we call “value trapped”: they can see what AI makes possible, but they lack the access to use it properly.

Finally, the real gap shows between the teams that have rebuilt around AI and the ones still layering it on top. The report findings link deeper AI integration to three outcomes:

The takeaway? The gap is not about tools, since nearly every team has those; It is about how far AI reaches into the work. The deeper the teams rebuild around it, the larger the payoff, as shown across the three outcomes above.

Overall, AI is thinnest exactly where the stakes are highest: the decisions about what to fund, pause, or cut.

That is exactly where most teams hold back. Handing AI a role in those decisions takes a level of trust most have yet not reached. As one innovation leader in the study put it, "AI earns trust like a junior team member. You start by staying close, then only give it more responsibility once you see it can do the work.”

No team in the sample has reached full autonomy yet, but the teams which will survive the AI race are already giving more responsibility to it by the day. The full benchmark shows how far they have gone, and what it takes to get there.

Curious to discuss the findings with fellow innovation leaders? We’re hosting an online session on July 8th, 18:00 CEST/12pm EDT/9am PDT - RSVP here. The full report will be released on July 10th- if you sign up for the session, you’ll be the first to get a copy.

That’s it for today.

For sure, the conversation (and crusade?) of integrating AI in innovation management is far from over. Next time, we’ll look at changing team behaviours and fixing decision making to support (new) governance models.

Hans Balmaekers
Founder, the Compass and Chief @ Innov8rs

PS- ready to share your story in a next edition? Just ping me via [email protected]

PPS- feel free to forward this newsletter to all the innovation leaders in your company and network. Sharing is… indeed!