⚡ Breaking
$2.5T projected enterprise AI spend in 2026 MIT: 95% of enterprise AI pilots fail Lineage hits 99% touchless shipments, saves ~$10M Mallory Alexander: 99.9% customs filing accuracy Prism Logistics grows revenue 20%, no new headcount Winners report 3x–10x ROI on AI spend
Keynote · Enterprise AI · Supply Chain

The $2.5 Trillion Bonfire: Why 95 Cents of Every AI Dollar Goes Up in Smoke

Everybody bought the demo. Almost nobody read the fine print. Meet the man selling supply chains a way out of the 95% club.

Pallet, AI Logistics PALLET
Share: in · LinkedIn 𝕏 · Twitter f · Facebook ◎ · Instagram
$2.5T
AI spend, 2026
95%
Pilots that fail (MIT)
20x
Spend growth since 2020
3–10x
ROI for the winners

There is a number quietly haunting every boardroom that ever green-lit an "AI initiative," and it is not the cost. It is the casualty rate. Two and a half trillion dollars will pour into enterprise AI in 2026 — roughly twice the entire economy of the Netherlands — and according to a now-infamous MIT study, ninety-five out of every hundred pilots will quietly die on the vine.

Sushanth Raman walked onto the stage with that statistic and refused to let anyone look away from it. The founder and CEO of Pallet — a San Francisco company that builds an AI workforce for the unglamorous machinery of global logistics — opened not with a product, but with an autopsy. "Despite the two and a half trillion that's gone into AI investments," he told the room, "95% of enterprise AI pilots have not been successful."

It is the kind of line that should empty a conference hall. Instead, Raman used it as a hinge. Because his argument is not that AI doesn't work. It is the opposite, and far more uncomfortable: the technology works fine, and the buyers are the problem.

"It's not because frontier models are not delivering results," he said. "It's really because the enterprises haven't thought through the right ways on how to go identify how to deploy this technology, how to deal with change management." The models, in other words, are not the bottleneck. The org chart is.

01The Tyranny of the Slick Demo

Every failed deployment, Raman argues, begins with the same seduction. You wander a conference floor. A voice agent holds a flawless conversation. A document parser swallows a stack of invoices without blinking. The demo is gorgeous. You sign. And then nothing works.

"There's a huge difference between what you have in a demo to what it takes to go and productionize these deployments," he said. "And that's where most people are falling short." Real deployments are not clean. They plug into on-prem systems that were never designed to be spoken to. They run on tribal knowledge that lives in the heads of three veteran dispatchers and was never written down. They demand change management nobody budgeted for.

The demo is a magic trick performed in a vacuum. Production is the same trick performed in a hurricane, blindfolded, while the audience throws edge cases.

"There's a huge difference between what you have in a demo to what it takes to productionize these deployments. That's where most people are falling short."

— Sushanth Raman, Founder & CEO, Pallet

02Why "We'll Just Build It Ourselves" Is a Trap

Every CIO eventually asks the obvious question: why pay a vendor when we have engineers? Raman's answer is patient and a little brutal. First, the models are non-deterministic. "When you ask ChatGPT the same question 10 times, it oftentimes does not give the same response," he noted. Charming in a chatbot; catastrophic in a customs filing. "90% it's not good enough. You need it to be consistent and accurate." Getting there can swallow six to twelve months.

Second, there is no finish line. Your in-house tool may sing on day one, but six months later you've onboarded five new customers, each with their own operating procedures, and your engineers are now full-time babysitters instead of builders. The maintenance bill nobody forecasted comes due every quarter.

Third — and this is the cruelest one — the ground keeps moving. Every time OpenAI, Anthropic, Google, or an open-source lab ships a new model, the entire solution must be re-tested. Does the new model improve accuracy? Does it cause regressions? "On top of running your core business, you're left with a lot of time just maintaining." Build-it-yourself, Raman suggests, is less a project than a treadmill that speeds up on its own.

"When you ask ChatGPT the same question 10 times, it oftentimes does not give the same response. In supply chain, 90% is not good enough."

— On why non-determinism kills naïve deployments

03How to Tell the Real From the Hype

So you're back where you started: a floor full of vendors, all armed with beautiful demos. "How do you tell who's full of it," Raman asked, "and who's actually delivering ROI?" His framework has three legs, and he insists you need all three.

One — pick a brutal metric. Decide, before any pilot, exactly how you will judge success: percentage of touchless shipments, number of customs orders filed, percentage of support tickets deflected. Then put three vendors through the same gauntlet against the same number. "When you do that, you're forming an objective grounding... you're not biasing yourself to a demo."

Two — demand embedding, not shipping. Off-the-shelf is a red flag. You want a partner who flies to your site, sits with your operators, and grinds through the messy integration work like a McKinsey or Cognizant engagement. "If someone's sitting there and telling you that they'll get you to a deployment in a day," he warned, "probably not going to work."

Three — codify the tribal knowledge. Your humans succeed because they carry context the AI doesn't have. The right vendor has a deliberate strategy for capturing that knowledge and documenting the SOPs that live in people's heads. Skip any of the three, Raman said flatly, and "good luck."

The Three-Part Buyer's Test

Raman's filter for separating ROI from theater

01

Pick a Brutal Metric

Define success before the pilot — touchless shipments, tickets deflected, orders filed — and hold every vendor to the exact same number. No demos, just evidence.

02

Demand Embedding

Off-the-shelf loses. Insist on a partner who comes on-site, manages change with your operators, and wires into your messy legacy systems like a consulting deployment.

03

Codify Tribal Knowledge

The AI lacks the context your veterans carry. The right vendor has a real plan to capture that knowledge and document the SOPs nobody ever wrote down.

The Numbers Behind the Bonfire

Where the money goes — and where it vanishes

Enterprise AI Spend Is Exploding

Projected annual spend · 2020 → 2028

2020
$150B
2026
$2.5T
2028
$5T

...But the Pilots Keep Dying

Enterprise AI pilot outcomes, per MIT study

95%
Failed pilots — 95%
Successful — 5%

Three Proofs the 5% Is Real

What success actually looks like on a balance sheet

Lineage · Cold Chain

$10M Saved in 49 Days

99%

The world's largest cold-chain operator — ~25% of the global refrigerated food supply flows through its warehouses — handed Pallet ~2M annual orders. Result: 99% touchless processing, ~$10M saved, and a roadmap toward $50M next year.

Prism Logistics · Freight

Growth Without Hiring

+20%

Told by its PE owner to grow revenue but add zero headcount, this freight broker let agents run the full order lifecycle — carrier booking to invoicing. Two enterprise customers signed, revenue up 20%, net margins up 10%.

Mallory Alexander · Customs

The High-Stakes One

99.9%

One wrong customs filing means $5K–$15K fines, with ~6 documents per shipment. A skeptical CEO let reasoning models — with guardrails — take over. The result: 99.9% accuracy and $1.8M added to EBITDA in six months.

Don't Boil the Ocean. Start With "Where's My Shipment?"

Raman's playbook for the first 12 months

  1. Pick one boring workflow. Comb the inbox for the most repetitive task — "where is my shipment," "reschedule my appointment." It feels lame. That's the point: nailing it builds internal trust.
  2. Run a real pilot, not a demo. Put several providers through a holistic evaluation against your one hard accuracy metric. "Do not trust the demo."
  3. Choose the one that clears the bar. Pick the vendor that actually hits your numbers, embeds with your team, and codifies your tribal knowledge.
  4. Build a roadmap of 10–15 agents. Then run the same loop again and again — metric, pilot, deploy — until, a year in, the ROI compounds into a genuine competitive edge.

What makes Raman's pitch land is its refusal of the usual AI sermon. He is not promising sentience or a revolution by Tuesday. He is promising that if you stop falling for demos, pick a number you can't fudge, and treat deployment like the grueling consulting engagement it actually is, the math eventually becomes absurd in your favor.

The roster he flashed — Estée Lauder, Lineage, and now the early days of work with FedEx — is less a brag than a thesis statement. These are not companies chasing novelty. They are companies counting cash. "For every dollar of AI spend," Raman said, "they're getting 3 to 10 dollars of cash flow back into their business."

So here is the question he left hanging over the room, and over anyone still nursing a stalled pilot: if there were exactly one workflow you could automate that would move real money, what would it be? The $2.5 trillion bonfire will keep burning. The only choice on offer is whether you're feeding it — or warming your hands.

"For every dollar of AI spend, they're getting 3 to 10 dollars of cash flow back into their business."

— The math the winning 5% live by

Watch the Full Keynote

Sushanth Raman lays out the framework — the metrics, the embedding, the tribal knowledge — that separates the 95% from the few who actually cash the check.

▶ Watch on YouTube →