What We’re Bringing Back from the Plug and Play Silicon Valley Summit

Thread AI Editorial · November 9, 2026

“Everything still feels the same, except we’re spending more money.” That’s how a technology leader at Starbucks described enterprise AI, three years in, on stage at last week’s Plug and Play Silicon Valley Summit in Sunnyvale. It’s an event close to home for us, since Plug and Play is one of our investors and the Summit is where we first met our customer Brightstar.

Over three days, DHL Supply Chain, Flight Centre Travel Group, Smartbird (formerly Allbirds), and others kept coming back to the same practical questions: who should own AI, how to tell whether it’s paying off, and how to keep it under control as agents take on more of the work.

Here are five signals we’re bringing home:

1. 01Companies are running AI as one central platform
2. 02The metric that matters is cost per outcome, not tokens
3. 03The model is becoming a commodity
4. 04Companies are quick to green-light pilots, but slower to invest in scaling them
5. 05Reliability comes first, and oversight has to be real

1. Companies are running AI as one central platform, with business units owning the use cases

The most organized companies split AI ownership in two. One central team owns what every AI project shares: vendor contracts, the budget, which models are approved, and the security and quality rules. Business units own the rest. They pick the problems worth solving, fund them from their own P&L, and answer for the results.

This setup also makes it easier to use the right model for each job. Instead of relying on one chatbot, teams are moving to a mix of models, from large frontier models for the hardest reasoning to small, cheaper ones for simple tasks like sorting requests. They all sit behind one central gateway.

This is something that we’ve been saying for a while: run AI like infrastructure, and let the teams closest to the work build on top of it. Lemma works the same way. It connects to the systems you already run and works with any model, so you’re never locked into one provider. Teams can use the best or most cost-efficient model for each step, even within a single workflow.

2. The metric that matters is cost per outcome, not tokens

The most mature teams budget AI per use case rather than per person, so spending is tied to what each use case is meant to deliver. They track cost per request handled or per hour saved. As one slide put it, “Tokens are an input, not a success metric.”

Presenters were candid that this is the hardest change to make inside an organization, but the payoff can be big. Flight Centre’s AI assistant handles 81% of requests without human help and finishes routine tasks 2.4 times faster, and DHL Supply Chain plans to double revenue in five years without doubling its workforce. Across sessions, the goal sounded the same: more output from the same teams, with roles changing rather than disappearing.

Patience is short, too. One bar we heard for any new tool: measurable impact on the P&L in less than 60 days, repeated sprint after sprint. Outside the sessions, ROI came up in almost every conversation we had. These are the issues to be raising. Counting tokens tells you what you spent, not what you got for it.

3. The model is becoming a commodity, and the advantage is what you own around it

A recurring theme was that models are becoming interchangeable, and the lasting advantage is everything around them: your data, your experts’ judgment, and the corrections they make every day. When an underwriter overrides a recommendation or a lawyer rejects a clause, that decision has value. If it isn’t captured, it improves a vendor’s product instead of yours. That’s why many executives we meet now want to own and run their AI workflows themselves.

One example: Bridgewater fine-tuned a smaller model on a handful of narrow tasks and beat the best frontier model’s accuracy at a fraction of the cost. The question we heard that sums it up: if your AI vendor disappeared tomorrow, how much of the intelligence would survive?

Companies are also rethinking how they work with AI vendors. Rather than buying a tool and running it as is, they want to co-develop systems they own and can keep improving. One speaker’s warning: if a vendor fixes one problem and leaves, and the team can’t run the system on its own, expect it to be replaced within a year or two, since ripping out software has never been easier.

We’d go further: your team’s expertise is the asset, and it should stay with you no matter which model or vendor you use.

4. Companies are quick to green-light pilots, but slower to invest in scaling them

Many AI pilots get approved as exceptions, skipping the usual review by security, the CIO’s office, and finance. That gets them started fast, but it leaves them without the backing they need to scale.

The fix we heard: once AI is strategic, meaning it’s on the board’s roadmap, touches sensitive data, or carries real costs, treat it as an investment measured by ROI and bring those teams in early. Then fund the internal champions and power users who drive adoption, because it doesn’t happen on its own. Companies that scale also use a clear filter. DHL Supply Chain backs an idea only when it has a big business need, mature technology, and a real payback. Ideas that pass get productized and rolled out across its network.

There’s a technical side, too. In The AI Ship Date Dilemma, our research on 132 enterprise AI teams, only 5% had committed to a ship date outside their own company. Getting to production is more work than most teams expect.

5. Reliability comes first, and oversight has to be real

The teams running AI at scale don’t reach for an agent first. If a simple, rules-based tool like robotic process automation can do the job, they use it, because it gives the same result every time and is cheaper to run and maintain. Agents come in where rules fall short, like pulling demand and staffing data from several systems to tell a warehouse manager where to move people when trucks run late. As one logistics leader put it, “Not every solution is an LLM hammer.”

When agents act on their own decisions, there’s often no pause for a person to catch a mistake, so the checks around them matter more. Several speakers stressed a principle we’ve always held: a person stays accountable for the outcome. That person needs to be able to step in before an agent acts, not just sign off afterward, and accuracy and fairness need watching all the time, not once a quarter. Getting this right is a strategic advantage, and getting it wrong is a strategic risk.

We agree, and it’s how we built Lemma. It runs rules-based steps and AI agents together in one platform, so each task gets the right tool, and its human-in-the-loop controls let people review and approve the steps that matter most. We call this Controlled Autonomy: agents do the work, and people make the calls that matter.

What we’re taking home

Across travel, logistics, retail, and finance, the hardest problems we heard about had little to do with the models themselves. They came down to ownership, measurement, and control: who runs the platform, how spending ties to results, how expert decisions get captured, and who answers when something goes wrong. The companies making real progress treat AI as core infrastructure, with clear owners, outcome-based budgets, and checks they can trust.

We left Sunnyvale with plenty to think about. Thanks to Plug and Play for three great days. If you’re working through any of this, we’d love to compare notes.