Palantir, Zeta Stitch Data While AI Models Lose the Plot

So, What Actually Happened?

So, the loudest words in AI had a quiet week. Agentic AI and AI governance still filled the headlines, but underneath them the buying conversation drifted somewhere far less glamorous: data quality, data integration, and plain old compliance all gained ground. We scanned 190,000 articles this week so you don't have to, and the pattern was hard to miss. The money stopped chasing the model and started fixing the plumbing. AI quietly slipped out of the data center and onto the device in your pocket. Palantir and Zeta stitched their data together while consultants warned that poor data quality is why AI under-delivers. Even the security crowd moved down a layer, as Black Kite joined a push for agentic SecOps. The flashy layer kept the spotlight. The boring layer got the budget.

The Bottom Line: The model was never the hard part. This week the market admitted it, and started paying the people who fix the data underneath.

The Tracks That Matter

1. AI's Next Revolution Won't Live in a Data Center

Every week brings another headline about a giant new AI data center. This week brought the opposite story. As models get smaller and more efficient, the next computing shift is moving onto your own devices, the laptop on your desk and the phone in your hand. Big tech is still pouring hundreds of billions into hyperscale GPU clusters. Meanwhile the interesting action quietly relocated to silicon you already own.

The how is unglamorous and powerful. Quantization trims a model's math from 16-bit precision down to 4-bit, cutting memory needs with little loss in quality. Mixture-of-experts designs wake up only the few parameters a task actually needs. Dedicated neural chips, like Apple's Neural Engine and Qualcomm's Hexagon, run these workloads at a fraction of the power a GPU burns. Put those together and open models like Llama 3, Phi-3, and Gemma already run on a high-end laptop. The cloud trains the giant once. Your device handles the everyday answers.

Why care? Because the economics flip. Cloud AI charges a small tax on every query, and running locally drops that cost toward zero, trading a forever subscription for a one-time piece of hardware. For regulated industries the privacy math is sharper still: data that never leaves the device never has to be governed in transit. We have seen this movie before. Computing went from mainframes to PCs to smartphones, always drifting closer to the user. AI is taking the same walk home.

Here's what works: Audit which of your AI workloads actually need the cloud. The privacy-sensitive ones (medical notes, contracts, customer records) are exactly the ones that belong on-device, where inference is cheap and the data never travels. Start one small on-device pilot before your next cloud bill renews.

2. More AI, Less Value: The Gap Nobody Wants to Name

The uncomfortable number behind the AI boom is this: most companies are adding AI and not getting more value for it. A new analysis named it directly, the AI success gap, the widening distance between how much AI an organization deploys and how much it actually returns. More pilots, more copilots, more spend, and the needle on outcomes barely moves. The tools multiplied. The results did not.

Ask why and you keep landing on the same unglamorous culprit. A data-quality consultancy put it plainly this week: poor underlying data is the biggest reason AI and automation projects under-deliver. Inconsistent records, missing fields, duplicates, miscategorised entries, they quietly poison every model trained on top of them. This is the oldest lesson in the business wearing new clothes. They slap an LLM on messy data, call it AI-driven, then wonder why the magic never shows up on Monday.

The other half of the gap is people. A workforce study this week found readiness lagging well behind adoption, with teams handed AI tools faster than they were taught to use them. So you get the worst of both worlds: bad inputs feeding clever models, run by people nobody trained. The fix is not a better model. It is better data and a readier team, the two things no vendor can sell you in a box.

Here's what works: Before you approve the next model upgrade, run two scores. One on your data quality (completeness, accuracy, consistency), one on your team's readiness to actually use what you are buying. If either is weak, fix it first. A great model on bad data and untrained people is an expensive way to stand still.

3. The Quiet Land Grab Is for Your Messy Data

While the cameras pointed at model launches, a different kind of deal kept closing all week: companies racing to unify enterprise data so AI can finally use it. Palantir and Zeta paired up for data unification, bundling scattered systems into one usable layer. It is not a headline that trends. It is the kind of deal that decides whether any of the flashy AI on top actually works.

It was not a one-off. The same days, NTT DATA and Cursor partnered to speed enterprise AI work, another bet that the bottleneck is plumbing, not intelligence. Step back and a cluster appears: the partnerships forming right now are about connection and cleanup, not new models. When several big players move on the same unglamorous problem in the same week, that is the market quietly telling you where the value moved.

The disease they are all treating has a name: integration sprawl. One firm just detailed a low-code fix for integration sprawl, the tangle of half-connected systems that every enterprise accumulates and nobody owns. Sprawl is where AI projects go to die: the data exists, but it lives in forty places that do not talk to each other. Unify it and the model on top suddenly looks brilliant. Skip it and the smartest AI in the world is guessing from fragments.

Here's what works: Map your own integration sprawl before you buy more AI. List every system that holds the data your AI needs, then count how many actually connect. The partnerships closing this week are a tell: the companies winning with AI fixed the connections first. The model was the last 10%, not the first.

4. Hospitals Want AI Agents That Can't Touch the Records

Healthcare is racing to put AI agents into real clinical work, and quietly drawing a hard line around them. The new guidance is blunt: AI agents should not run end-to-end workflows or write directly into systems of record. Adoption is already here. 71% of US hospitals ran predictive AI inside their health records by 2024, up from 66% a year before. The question stopped being whether to use agents. It became how to box them in.

The recommended shape is telling. Put a deterministic workflow engine in the middle as the single source of truth, and let narrow, single-purpose agents do small jobs with least-privilege access and their own identity. Without a shared workflow state, each app sees only part of the process, and you get duplicate claims, lost exceptions, and stale data written back into the record. The lesson travels far past hospitals: the agent is the easy part, the governed layer underneath is the actual product.

Researchers are flagging a subtler shift too. As patients carry AI health advice into the exam room, a triadic clinician-patient-AI dynamic replaces the old two-way trust, and accountability blurs the moment the algorithm contradicts the doctor. That is the real frontier: not whether the model is smart, but who is answerable when it is confidently wrong. Every regulated industry is about to hold the same conversation.

Here's what works: Put a governed orchestration layer between your AI agents and your systems of record, the way these hospitals are. Give each agent one job, one identity, and the least access it needs. Decide who is accountable for every automated action before it runs, not after an auditor asks.

5. AI Agents Became the Attack Surface Nobody Patched

Give software the power to act on its own and you also hand it to whoever breaks in. Security teams spent the week scrambling to catch up. Black Kite joined Torq's alliance for agentic SecOps, wiring autonomous agents into security operations so defenses can move at the speed attackers now do. The same agents that make AI useful make it a bigger target, and the industry just admitted the gap is real.

The threats are new in kind, not just degree. Practitioners at a software-security gathering argued that AI is rewriting the rules of software security outright, with whole new failure modes to defend. Chief among them is prompt injection, where poisoned input quietly hijacks a model's instructions. One team laid out how to defend against prompt injection before it becomes the decade's favorite exploit. You cannot firewall your way out of a model that can be sweet-talked into misbehaving.

And the job is spreading. A new report found the cyber skills gap moved to non-security teams, as the people building and running AI now own risks they were never trained for. That is the quiet danger: the attack surface grew faster than anyone's headcount. When every team ships agents, every team is now a security team, whether they signed up for it or not.

Here's what works: Treat every AI agent like a new employee with system access. Give it a unique identity, the least privilege it needs, and a logged trail of everything it touches. Then test it the way an attacker would, with adversarial prompts, before it touches anything that matters.

6. Canada's Pension Giant Bets CAD 1 Billion on Indian Data Centers

While model labs grabbed the spotlight, the patient money went somewhere physical. CtrlS landed a CAD 1 billion partnership with CPP Investment, Canada's pension fund, to build out data-center capacity in India. Pension capital is the opposite of hype: it moves slowly, hates risk, and parks where it expects steady returns for decades. When it commits a billion to concrete and cooling, that is a vote on where AI actually lives.

The location is the signal. The AI conversation fixates on a handful of US hyperscalers, but the buildout is quietly globalising, and India is becoming a serious node in it. Sovereign and pension funds are treating compute capacity the way an earlier generation treated railways and ports: unglamorous infrastructure that everything else depends on. The companies renting that capacity tomorrow will care less about whose model is cleverest and more about who has power, space, and a data center nearby.

For anyone planning AI capacity, this reframes the map. Your compute geography is widening, and a second supplier on another continent is starting to mean lower latency and lower political risk, not just a backup line. The smart move is to watch where the slow money builds. It is rarely wrong about which infrastructure ends up mattering, even if it is always early and always boring.

Here's what works: When you plan AI capacity, look past the model launches to where long-horizon capital is pouring concrete. Pension and sovereign funds building data centers tell you which regions will have cheap, reliable compute in three years. Price that into your sourcing now, while everyone else is still arguing about models.

Signal vs. Noise

🟢 Signal: Data quality and data integration. While the buzzwords got the clicks, the unglamorous work of cleaning and connecting data climbed to the front of the buying conversation this week, alongside plain compliance. That is the audit committee and the data team deciding where money goes, not the demo crowd. Most coverage is still scoring model benchmarks and missing where the budget quietly moved.

🔴 Noise: Agentic AI. The phrase still pulled the heaviest headline volume on Friday, but its real influence slipped as the conversation broke into the specific layers companies actually buy: data, security, orchestration. Anyone still tracking ”agentic AI” as one big trend line is reading the brochure, not the procurement list.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

A consultancy declared poor data quality the real reason AI under-delivers, Palantir and Zeta partnered to unify enterprise data, and a major analysis named an ”AI success gap” between spend and value, all inside the same 48 hours.

Read alone, each lands on a different desk. The strategy desk takes the success gap. The infrastructure desk takes the Palantir deal. The data team takes the quality warning. Read them on the same morning and one sentence sits under all three: the industry just stopped blaming the model and started blaming the plumbing. For two years the story was bigger, smarter models. The quiet correction arriving now is that the model was rarely the bottleneck, the messy, disconnected, ungoverned data underneath always was. The companies pulling ahead are not the ones with the cleverest AI. They are the ones who finally did the boring work first. The move on Monday is to ask, before your next AI purchase, one unglamorous question: is the data this thing will run on actually clean, connected, and governed? If you cannot answer yes, a smarter model just buys you more confident mistakes.

By The Numbers

• 71% of US hospitals ran predictive AI inside their health records by 2024: up from 66% a year earlier. Adoption was never the problem. Trustworthy data and governed access are.

• CtrlS landed a CAD 1 billion data-center partnership with Canada's pension fund: patient capital is buying the physical AI buildout, and it is doing it in India, not Silicon Valley.

• Cloud data warehouses are now quoting $6.25 per terabyte: the storage layer under every AI project is commoditizing fast, which makes what you do with the data, not where you keep it, the real edge.

• The FDIC moved its resolution-filing threshold from $50 billion to $100 billion: the compliance perimeter for large institutions just shifted, and every AI system inside it has to move with it.

• More than 600 million people in Africa still lack electricity: every ”AI everywhere” roadmap eventually meets a power bill and a grid that has to exist first.

• See what's rising in our 190K-article corpus this week →

Deep Dive: Nobody Claps for the Soundcheck

When I was DJing, I learned fast that the gig is won or lost hours before the crowd shows up. The soundcheck. Nobody buys a ticket for it. Nobody films it. But if the levels are wrong, if one cable hums, if a monitor feeds back, the best set in the world falls apart at 1am in front of a packed room. The headliner gets the screaming. The soundcheck gets everybody home happy. AI just spent a week relearning that the hard way.

The headliner has been hogging the stage

For two years, the model was the headliner. Bigger parameters, flashier demos, another launch every week. The whole industry crowded the front of the stage to watch. And the value, for a lot of companies, never quite arrived. This week an analysis finally named it: a widening gap between how much AI gets deployed and how little it returns. The set looked incredible. The room still emptied.

The hum was in the cables all along

Pull the story apart and the culprit is almost never the model. It is the data underneath: inconsistent, duplicated, sitting in forty systems that do not talk. A consultancy said it plainly this week, poor data quality is the number-one reason AI projects under-deliver. That is the cable nobody checked. You can headline the greatest model ever trained, and if the data feeding it hums, the audience just hears noise.

This week, the market started paying the sound engineer

Watch where the serious money moved. Companies pairing up to unify enterprise data. Hospitals insisting on a governed layer between agents and the records. Security teams giving every agent its own identity. None of it trends. All of it is soundcheck, the unglamorous work that decides whether the show actually plays. The budget is quietly leaving the headliner and going to the crew that makes the headliner possible.

What Actually Works

1. Check the levels before the show: Score your data quality (completeness, accuracy, consistency) before you buy the next model. A clean foundation beats a clever model every time.
2. Unify before you automate: Count how many of your systems actually connect. Fix the integration sprawl first, or your agents will guess from fragments.
3. Give the crew credentials: Every AI agent gets one identity, least-privilege access, and a logged trail. Govern the layer, not just the model.
4. Pay the sound engineer: Fund the boring data and governance work like it is the product, because for your customers, it is.

The crowd remembers the headliner. But the festival that sells out again next year is the one where the soundcheck was flawless and nobody in the room ever knew it happened. The model gets the applause. The data work was always the work.

What's Coming

On-Device AI Becomes a Procurement Line

The shift of AI onto personal devices is the early signal of a bigger move. As inference gets cheap on local silicon, expect ”what runs on-device” to turn from a nerdy spec into a real procurement question, especially anywhere privacy and latency matter. The next hardware refresh cycle will quietly become an AI strategy decision.

The Data-Unification Wave Keeps Building

This week's data-unification partnership is not the last. Expect more pairings aimed squarely at making enterprise data AI-ready, because that is where the value gap actually sits. The vendors who can connect and clean your data will start charging like the strategic layer they are, and they will be right to.

Governance Turns Into a Revenue Stream

One report this week asked whether AI governance could be a new revenue stream for security service providers. That is the tell. When the people who sell controls start pricing governance as a product, the boring layer officially became a market. Watch ”we govern your agents” become a line item, not a footnote.

For Your Team

Strategic purpose: This week belongs on the leadership table because it quietly moved the AI question from ”which model” to ”is our data ready for any model.” The headlines kept score on launches. The real story was the budget shifting to data quality, integration, and governance, the layer that decides whether AI returns anything at all.

Monday's meeting prompt: ”If we audited the data our AI actually runs on tomorrow, its completeness, its accuracy, how many systems it connects, would we trust the output enough to bet a customer decision on it? If not, what are we fixing before we buy another model?”

The Foundation-First Framework:

1. Score the data, not the model: Before any AI purchase, rate your data on completeness, accuracy, and consistency. A weak score means a smarter model just makes confident mistakes faster.
2. Count the connections: List the systems your AI needs and how many actually talk to each other. Integration sprawl is where AI value quietly dies.
3. Govern the agents: Give every AI agent one identity, least-privilege access, and a logged trail. The governed layer is the product, not the demo.
4. Ready the people: Match every tool you deploy with the training to use it. Adoption without readiness is spend without return.

Share-worthy stat: 71% of US hospitals were already running predictive AI inside their health records by 2024, up from 66% a year earlier. The technology is in the building. The question every sector now faces is whether the data and the guardrails came in with it.

Go deeper: Track where data, governance, and infrastructure are concentrating in real time →

The Track of the Day

”The intelligence revolution is no longer just moving to the cloud. It is coming home.”
AI World Journal

The whole week rhymed with that line. The big, centralized, expensive layer keeps the headlines, but the action is drifting back to the edges: onto your devices, into your data, down to the boring foundation nobody photographs. Same as it ever was. The flashy part gets the applause. The foundation gets the result.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: June 27, 2026 | Curated by Yves Mulkers @ Ins7ghts