WSJ Events Sep 19, 2026 28m 10m saved
With Marina F. Bellini, President for Global Business Services and Digital Technologies at Mars · Jonathan Lehr, Co-Founder and General Partner at Work-Bench
The part of Mars that works with the largest number of startups is not marketing. It is cybersecurity.
At a company whose brands include M&M's, Snickers and Royal Canin, the outside partner list would be expected to run through creative agencies. Marina Bellini said she counted, and the volume of agent experimentation sits with cyber startups rather than with the incumbent vendors Mars has used for years.
I don't know if it's surprising for some in the room definitely was for me when I start asking for the counts but the area of agent experimentation that we have the most startups working with us not the incumbent and traditional partners is cyber
Bellini joined Mars in October 2024 to run operations, global business services and enterprise digital technologies at a business the company puts at more than $50 billion in sales. She came from two years as chief operating officer of Itaú Unibanco, Latin America's largest bank, and before that was CIO or chief digital officer at BAT, PepsiCo Latin America and AB InBev. Alongside her on the panel was Jonathan Lehr, who worked in IT at Morgan Stanley in the office of the CIO before co-founding the New York seed firm Work-Bench.
The full segment is covered here so you can skip it. 28 minutes of audio, 18 minutes of reading.
Here are the 13 takeaways that matter.
Key Takeaways
Cyber, not marketing, is where Mars runs the most startup experiments, and Bellini said the count surprised her
She put it down to the pace of the field and the impossibility of keeping up alone
Lehr's one-line rule for enterprise software is buy to operate and build to differentiate
Off-the-shelf products get most of the way there, and internal teams spend their time on the last mile
A bought CRM, not a built one, carried Royal Canin's biggest customer-system upgrade of the past six months
Every choice has to be reversible within a year or two, because Bellini does not assume today's CRM vendor is next year's
"The model is hallucinating" usually means the data is wrong, and she said business leaders now see that for themselves
The headache she is trying to remove is AI for the sake of AI, run without a value case
Bankers who vibe coded their own deal tools asked why production takes six months
Incumbent vendors are selling AI features, and her question back is what AI changes about how they serve Mars
Lehr's successor to the software development life cycle is the agent development life cycle
1. Build vs Buy Has Shifted
The moderator opened with a show of hands: how many in the room had vibe coded their own CRM in the past year. Nobody had. He said the do-it-yourself takeover had not happened as promised, and that reports of a SaaS collapse were overstated, but that the forces behind those headlines were changing the build-versus-buy decision.
Bellini's first answer was about her own career. Financial services builds everything, consumer packaged goods buys everything, and she has worked in both. Back in CPG, she said the split is no longer obvious.
Her worked example is Royal Canin, the Mars business that sells medicinal food for pets and has to match the right food to the right breed at the right age. The biggest enhancement of the past six months there came on top of a bought CRM rather than a built one.
Bespoke used to mean building it yourself
I wouldn't buy things in the past because I had to build everything bespoke but now they come with the ability to be bespoke when I buy.
Marina F. Bellini
That changes what she optimizes for. Rather than pick the right answer for today, she said Mars is picking answers it can unwind.
Modularity is the hedge against a vendor disappearing
things have to be modular because what looks like the right answer today to have a plug on my CRM application I don't know if that CRM application is going to be there tomorrow
Marina F. Bellini
She wants decisions she can change in one or two years without the pain of replacing an ERP system, and she said that opens the door to partners large companies historically struggled to work with.
2. The Last Mile Problem
Lehr gave his own context first: thirteen years ago he worked in IT at Morgan Stanley in the office of the CIO, and he has since co-founded Work-Bench, a New York seed firm investing at both the infrastructure and application layers. His summary of the market is four words.
His rule for the buy-versus-build decision
if I want to summarize what we're seeing in the market right now in terms of buy versus build, it's buy to operate and build to differentiate.
Jonathan Lehr
He said that holds across the firm's network of financial services, healthcare, life sciences and media executives. Years ago, companies built internally either for strategic advantage or to keep IT staff occupied. Now the gap is narrower and the work has moved to the end of the job.
Off-the-shelf software gets most of the way and stops
there's a lot of off-the-shelf software that's pretty good that can get you 80 90% of the way there in terms of capabilities, but a lot of IT folks are actually spending time on that last mile
Jonathan Lehr
That last mile, in his description, is new evaluation harnesses, access to data, cost optimization, and connectors that most organizations still have to build.
The data question gets skipped
A lot of people forget that in order for a lot of this AI to work, you actually have to figure out where is the data that you're going to do the processing on.
Jonathan Lehr
He dated the SaaS apocalypse talk to February and said it already feels remote. Total cost of ownership still decides things, he said, even as AI coding tools let internal teams ship faster than before. The question has moved to governance, which he attributed to the panel before his, where Chime's chief operating officer Janelle Sallenave spoke.
Governance is what the vibe-coding argument left out
the exec from Chime, the COO, she said, people forget about governance
Jonathan Lehr
His conclusion is that the buying is happening in the infrastructure that makes fast internal development safe: AI governance, permissions and the rest.
3. Platforming at Mars
The moderator said governance was the word of the day and asked how Bellini keeps a company of Mars's size in line. She described arriving at a decentralized technology organization spread across businesses and markets, with the Kellanova acquisition arriving at the same time.
What she found was an uncontrolled spread of AI projects
it was this I call avalanche of AI solutions permeated across the organization. Some create values, some not.
Marina F. Bellini
Responsible practice was already in place, she said, and she did not change it. What she did centralize was the platform layer: access to data, observability, and the machinery every agent needs regardless of what it does. Her argument for it was scarcity of skill rather than control.
The waste she wanted to stop was duplication
not to have the same agents created 20 times around the organization
Marina F. Bellini
Making M&M's and caring for a pet in a hospital are different problems, she said, but the data care and the performance monitoring underneath them are common. She described the choice as picking where to fight for leverage while leaving innovation in the businesses. She also said the standard keeps moving.
The bar resets faster than the policy
whatever we define a year and a half ago is not enough today won't be enough at a year and a half from now
Marina F. Bellini
Her check on that is her own team. She said her CTO and head of IT is probably the most business-facing person in her organization, because the moment central IT stops offering what the businesses need, the businesses go back to building their own.
4. Bankers Vibe Coding
Lehr picked up the point about keeping up and said the gap between IT and the business is wider than ever. His example came from an IT friend at a boutique bank a couple of months ago, where bankers had built their own deal analysis tools and small CRMs.
The bankers could not see why production took six months
Hey, I did this in like an afternoon. Why does it take you guys 6 months to productionize this? Let's just put it out live.
Jonathan Lehr, relaying what the bankers said
What the bankers had not seen, he said, were the internal integrations and the data sensitivity rules. The pull is real either way: when something can be vibe coded in an afternoon, the expectation is that it ships faster. His read is that infrastructure and security tooling bought from startups, onboarded carefully, is what closes that gap.
5. Cyber Draws the Startups
Bellini said she had gone looking for the numbers and found something she did not expect.
The most startup work at Mars sits in cyber
I don't know if it's surprising for some in the room definitely was for me when I start asking for the counts but the area of agent experimentation that we have the most startups working with us not the incumbent and traditional partners is cyber
Marina F. Bellini
Her explanation is pace. Keeping up in security looks impossible, she said, which has opened the door to collaboration and to co-developing what comes next. Mars had some security startups in its supplier base before, but this is now the area with the highest participation.
She also named what she would have predicted instead. Mars sells technology products of its own, including science diagnostics and machinery, and she said the company cannot develop all of that itself.
Marketing agencies were the expected answer
but in terms of volume and the need for experimentation really felt by my own organization is in the cyber space
Marina F. Bellini
6. The Headaches Ahead
Asked how she avoids the legacy mess that vibe-coded software leaves behind, Bellini did not claim she could.
She expects the headache and is only triaging it
I think I'm convinced I'm going to have headache. There's no way to avoid it.
Marina F. Bellini
The ones she thinks she can predict come first. Responsible AI practice is one. Data is the second, and she said the argument there has changed in her favor. Ten years ago it was hard to convince anyone that foundational data work mattered. Now business leaders watch a model produce a wrong answer and reach the conclusion themselves.
Hallucination is usually a data failure
we call that the model is hallucinating is not we're just feeding the model with the wrong data
Marina F. Bellini
The partnership she wanted for twenty years, she said, is the one she now has, with colleagues asking to work on data together.
The third headache is AI with no value case
I think the headache that I have today that I'm trying to get rid of it is AI for the sake of AI and not for value
Marina F. Bellini
Her answer to that is education, starting with the chief executive. She has run hands-on AI training for senior executives on their own productivity, which she said she picked up from peers. The session that worked best was a different exercise: walking executives through choosing where to apply AI, which data it would need, which process would have to change, and who would manage the change for the people affected.
Left to themselves, the executives picked the worst option
if this would be a choice game, they would have lost because they would have chosen the fancy thing, the hardest to get the data right
Marina F. Bellini
That choice, she said, would also have needed a lot of process change and would not have created value at the end of it.
7. Where New Software Wins
Lehr said many founders coming out of Y Combinator lack customer empathy for what an executive has to do internally to make a product land, from data access to security requirements. The opening, in his view, is in capabilities that did not exist before rather than in replacing what does.
A startup trying to displace Salesforce is going to be in trouble, he said. He pointed instead to Petra Labs, a New York company in his portfolio working on AI engine optimization, a category he noted is only a year old. Its approach pulls data out of Snowflake, Databricks and other systems, brings it together, and produces attribution for how AI search drives business.
He said the company has found traction in consumer packaged goods rather than everywhere, and that the priority order differs by company. Some put security at the top of the build list, others marketing.
Priorities are being reshaped company by company
that's honestly what keeps it fun day-to-day because everyone's priorities are different and reshaped right now
Jonathan Lehr
8. Vetting a Seed Startup
Asked what two questions he wants answered before investing, Lehr contrasted Work-Bench's approach with a Silicon Valley one. The firm takes a New York, go-to-market-centered view, he said, where the Valley default is technology first.
The first screen is whether the founder understands the buyer's pain
number one is customer empathy. Do you actually have a viewpoint on what their pain is?
Jonathan Lehr
That test, he said, removes the founders building technology for its own sake. The second question is harder than it used to be. A unique insight from a prior job was once enough to build a product on. Now the cloud providers and the large AI labs ship features constantly, and established vendors are buying a founder's peers.
Founders need a multi-product plan from day one
how do you have a multi-product vision from day one but know how to implement it over the course of an initial wedge and then growing from there?
Jonathan Lehr
He said the demand on founders has gone up
It's a lot to ask from founders today. I don't envy them in this market
Jonathan Lehr
9. The Incumbent Problem
The moderator turned to incumbent vendors racing to add AI features customers may not want. Bellini said she had stepped out of the session for a call with one of them, and that most arrive selling what AI can do for Mars.
She turns the question around
I usually ask back the question what can you use AI for to change how you work with us
Marina F. Bellini
She said there is a long way to go, because those vendors have to reinvent their own business models rather than just their products, and that she cannot claim Mars is the only party that needs to change. So far she has not seen many successful cases. Her invitation to the largest implementation partners is to take pieces of the existing relationship and rework them together, which she said is hard on both sides.
Startups take the edges while incumbents work it out
that's where also startups come and start eating from the edges
Marina F. Bellini
Her caution on that is structural: a large company working with a startup is working with an organization that does not have the same legal, commercial and technology teams. Incumbents will not disappear tomorrow, she said, but the edges will keep being contested, and customers, startups and incumbents all have to work through what she called a transition model.
Division leaders at Mars keep their own vendor relationships, and the moderator asked how she gets everyone aligned. Her answer was that it happens sometimes.
Alignment runs both directions
It happens from time to time. But it takes two to tango.
Marina F. Bellini
She asks a partner what role it plays in Mars behaving as one company, and what role she plays in the partner doing the same. AI helps there, she said, by scanning what is happening across a relationship.
Tracking a decentralized organization is the hard part
20,000 IT professionals to really understand everything that's happening in a decentralized organization is a challenge
Marina F. Bellini
She also turned the point on her own function.
IT is the least tech-enabled team in the company
we say the IT team is the least tech enabled team inside our organization
Marina F. Bellini
The first team Mars analyzed for what its work is and how that work could be disrupted was IT itself. She said the function should be the guinea pig, and asked why technology should not be put at its own service.
10. Shadow AI and Agents
Asked about shadow IT, now shadow AI, Bellini said it is happening at a higher rate than before, because bringing in an agent without going through the integration, data and security process is easy. Her response is communication and education.
Sometimes a bad example is what makes the message land
sometimes it takes a bad example to reinforce the message and we had a few
Marina F. Bellini
Lehr said the pattern repeats. At Morgan Stanley he watched cloud access security brokers emerge as the way to bring cloud into the enterprise, and he sees the same shape now for agents. Work-Bench has backed a company, which he named as Koala Security, on the thesis that agents will proliferate through the cloud and through individual applications such as Perplexity.
Computer-use agents are a new attack surface
How do you secure computer use agents as a new threat surface area?
Jonathan Lehr
There are several technical approaches to that, he said, and the bet is forward-looking even though the precedent is an old one.
11. Never Pitch a CIO
Kevin Adams, CIO of Edward Jones, asked from the floor what Work-Bench does to train its founders on how to approach large enterprises. He described the version that fails.
A bad first approach can close a firm out entirely
come right to the top to a guy like me, get my personal cell phone, harass me, and then all of a sudden I send an edict to the team that if anybody speaks to this firm, they're going to be in a lot of trouble because they've already burned a bridge with me
Kevin Adams, CIO of Edward Jones
Lehr said this is the most impactful thing the firm does, before and after investing. Part of it is sequencing: a seed or Series A company has to work out whether a large enterprise is even in its ideal customer profile, or whether it should sell to startups and the mid-market first. Some of his companies do go to financial services firms first.
The advice he took from Morgan Stanley
My old boss at Morgan used to say never pitch a CIO. They're busy running the business. Go to their directs.
Jonathan Lehr
What Work-Bench sells instead is proximity. The firm hosts meetups and dinners on topics such as agents in financial services, invites corporate executives, then invites a curated set of startups and coaches them not to harass the person in the leadership seat.
Physical access is the edge over a Zoom pitch
You've now got an access in a way that a Silicon Valley startup can't over Zoom. You're now in the dinner.
Jonathan Lehr
His live example: a chief operating officer who had been on stage earlier that day, whose CIO is bringing eight direct reports to the Work-Bench office the following week for an hour on themes, with one vetted startup presenting its view of the permissions and authorization landscape. The firm has run this for a decade, he said, and it works because very few people do it.
12. What Replaces the CIO
An audience member asked whether the CIO role survives, or whether technology becoming embedded in the business pushes companies back toward the federated IT of the 1980s. Bellini said the panel had been discussing exactly that during the break.
Both sides have now learned the other's language
we spent years training IT professionals to learn the business language and now we are spending years training the business leaders to speak the technology language
Marina F. Bellini
The boundary will blur in a way that has not happened before, she said, and the us-and-them split is closing. What does not go away is the architecture job, and she made the case with a house.
Someone has to know where the corridor is
if you build a house you cannot put a bathroom without knowing where the corridor is. You open the door, you are behind the fireplace.
Marina F. Bellini
Anyone can make an agent, she said, but the next agent and the next person have to fit alongside it. Whether the orchestrator and architect role is called CIO, CTO or chief architecture officer, she said it is needed, and it asks for a different mindset from the team. Mars has run an exercise on how the work of its IT professionals will change, and she said the answer is not known. What is clear to her is that the old analyst job and general software engineering are being reshaped, and that architecture has to move from a specialist corner to something present across the IT organization.
13. Next Year's Big Shift
The moderator asked each panelist for the single biggest architectural change in their stack that would make the day's discussion outdated, on a horizon of a year or a week.
Bellini said it is not a technology architecture question.
The human and the technology cannot be designed apart
we cannot think about the human side and the tech side separately
Marina F. Bellini
The moderator asked whether she meant cyborgs. She said yes, to laughter.
Lehr's answer was a piece of infrastructure that does not exist yet.
After the SDLC comes the ADLC
There's going to be the ADLC, the agent development life cycle in terms of their permissions, in terms of personalization, knowing what you would want to do on your behalf, and then having the right eval guardrails around that.
Jonathan Lehr
Bonus Insights
Nobody in the room had vibe coded their own CRM
The moderator's opening show of hands drew no takers, which he called surprising.
The moderator's joke about AI and software
I just want to say AI has made building applications fast, cheap, and out of control.
Moderator
He immediately said he was kidding.
Aspirin was offered
After Bellini listed the three headaches she is managing, the moderator told her the event would find her some aspirin.
Bellini's view of her own achievement
She said colleagues treat the decision to centralize platforming as a big achievement, and that she thought it was simply the logical one.
Bellini's bottom line is that the build-versus-buy question is now a question about reversibility and data rather than about who writes the code, and that the scarce role in an organization full of agents is the architect who knows how the pieces fit together.
Products, Companies & Tools Mentioned
Mars (Bellini's employer, the maker of M&M's and Snickers; she described a decentralized technology organization she has been centralizing at the platform layer)
Royal Canin (The Mars pet-nutrition business whose biggest recent customer-system upgrade was built on a bought CRM, matching food to breed and age)
Kellanova (The acquisition she said arrived as she was taking on a decentralized technology organization)
Work-Bench (Lehr's New York seed firm, investing at the infrastructure and application layers; its rule is buy to operate and build to differentiate)
Morgan Stanley (Where Lehr worked in IT in the office of the CIO thirteen years ago; source of the cloud access security broker precedent and of "never pitch a CIO")
Anthropic (Named as the February version of the argument that one vendor would build everything internally for you)
Chime (Its chief operating officer, on the panel before this one, is the source of Lehr's line that people forget about governance)
Salesforce (Lehr's example of the incumbent a startup should not try to replace head-on)
Petra Labs (Work-Bench portfolio company in New York working on AI engine optimization, a category Lehr said is a year old; it is finding traction in consumer packaged goods)
Snowflake and Databricks (The systems Petra Labs pulls data from to build attribution)
Koala Security (The company Lehr said Work-Bench backed to secure computer-use agents, which he framed as the agent-era version of cloud access security brokers)
Perplexity (Named as one of the individual applications through which agents will proliferate inside companies)
Y Combinator (Lehr's shorthand for founders who arrive without empathy for what a large enterprise has to do internally)
Edward Jones (Kevin Adams, its CIO, asked the question about how startups approach large firms)
LinkedIn and Zoom (The channels Adams and Lehr both named as the wrong way in, against dinners and meetups)
Listen to the full episode
🔴 YouTube
Watch the full episode:
More Marina F. Bellini interviews and podcast appearances
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: