AI and ML

Majority report AI-related security incidents or vulnerabilities

The majority of companies that deploy AI systems end up shooting themselves in the foot with security, according to DigiCert.

Seventy-eight percent of enterprises report “experiencing AI-related security incidents or identifying AI-related vulnerabilities,” the digital identity biz said in a commissioned survey.

Among respondents, 27.7 percent experienced one incident, 21.9 percent experienced multiple incidents, and 28.4 percent had no incidents but identified vulnerabilities, a company spokesperson told The Register. Incident details were not disclosed, but they were caused by AI agents that were unauthorized or misconfigured rather than flaws arising from AI-generated code.

Consistent with its business focus, DigiCert attributes the survey’s findings to lack of AI governance.

“We wouldn’t allow an employee to operate without a verified identity,” said DigiCert CEO Amit Sinha in a statement. “AI agents should be no different.”

That’s become a common refrain. There are several initiatives underway to establish identifiers for bots, such as Private Access Control Tokens (PACTs), Estonia’s digital IDs for agents, and Microsoft’s Agent ID. But bot badging infrastructure remains a work-in-progress, leaving AI agents to run amok in many organizations.

DigiCert’s findings [PDF] echo a similar report two weeks ago from Spacelift that found 93 percent of organizations experienced AI-caused infrastructure incidents while only 19 percent had a governance plan in place. 

The survey stands in stark contrast with picks-and-shovels seller Nvidia’s State of AI 2026 report, which gushes, “Across every industry, AI is helping increase annual revenue and drive down annual costs while boosting productivity.” 

The DigiCert Q&A involves responses from 1,001 IT and cybersecurity leaders in the US, UK, and Australia, from various businesses. The survey shows that businesses are deploying AI first and asking questions later.

While 90 percent of organizations surveyed have discussed AI governance at the board level, just 50 percent have dedicated AI governance budgets and formal governance programs. This allows operational blind spots to persist. Just 53 percent of respondents said their organization could trace AI decisions back to the models and source data that produced those results.

“That becomes a problem the moment an AI system produces an unexpected or controversial result,” the report says. “Customers, executives, and regulators will all ask, ‘Why did it do that?’”

And perhaps at some point, companies will ask, why did we deploy that? ®

China’s open AI models have been a gift to developers everywhere. Now Beijing may pull them back in.

Chinese officials have discussed limiting who outside the country can use the nation’s best AI models, Reuters reports. The Ministry of Commerce ran the meetings over the past month, and Alibaba, ByteDance, and the startup Z.ai took part. The talks cover the most capable models, including some not yet out.

What is on the table

The plans reach past a simple export ban. They would also catch open-weight models, the freely downloadable systems that made Chinese AI popular abroad, alongside closed ones. Alibaba’s Qwen, ByteDance’s Doubao, and Z.ai’s GLM-5.2 all count among them.

Two other ideas surfaced. One would treat the leak or theft of proprietary AI as a national security crime. The other would limit which investors can fund homegrown AI firms. The sources cautioned that officials have decided nothing yet, that any curbs might apply only to future models, and that no timeline exists.

Reuters could not learn how the curbs would work. One panel of Chinese legal scholars has floated a tiered scheme: a light filing for basic tools, security reviews for stronger ones, and a domestic-only lockdown for the most sensitive models.

Why it matters

The move would mark a sharp turn. China won global goodwill by giving its models away, and European developers leaned on cheap open weights from firms like DeepSeek as an alternative to pricey American systems. Curbing them would thin that supply, and Reuters notes costs could climb for the many businesses that lean on them.

The shift mirrors Washington. The US has moved to stop China copying its models and recently restricted Anthropic’s frontier systems on security grounds, the very thing Beijing now fears in reverse. China has already built its own walls, grounding its AI researchers and steering who can back its startups. Treating models as state assets is the next brick.

Whether any of it becomes law remains unclear. If it does, the era of freely downloadable Chinese AI could quietly close.

The job of a soundbar used to be simple: make TVs sound better without complicated component systems, multiple speakers and wires strung all over your living room. Over time, soundbars have evolved into complex multi-speaker systems capable of competing with component-based systems in both features and sound quality. But have they become too complicated for their own good? 

With the current flagship soundbar system, the HW-Q990H, Samsung believes they can walk the line between simplicity and advanced functionality while offering performance that approaches that of a component system. Are they successful? Let’s find out.

What Is It? 

Samsung offers a wide selection of soundbars, from simple one-piece systems like the HW-QS90H to the more robust flagship soundbar system, the HW-Q990H, subject of our current review. The list price of the system is $1,999, but it is typically discounted to around $1,600. Unlike companies like Sony, who sell their flagship soundbar standalone, requiring customers to add expensive subwoofers and/or rear speakers for full performance, Samsung includes everything you need in the box with the Q990H: the soundbar itself, a pair of rear speakers, with both front and up-firing drivers and a compact powered subwoofer.

The system features a whopping total of 23 drivers, across all components. The soundbar itself includes fifteen individual drivers pointing in multiple directions: forward, to the sides and angled upward, to reflect height channel sounds off the ceiling. The rear channel speakers include three drivers each, pointing forward, to the side and up for reflective height channels. The system is completed by a very compact powered subwoofer, a cube which measures in at just under 10 inches on each side. The subwoofer uses dual 8-inch drivers in a push-pull configuration and weighs in at a fairly hefty 18.3 lbs. 

The HW-Q990H supports the two most common immersive surround sound formats – Dolby Atmos and DTS:X – and one uncommon one – Eclipsa Audio. Eclipsa Audio is an object-based immersive audio format developed by Samsung and Google (among others) as the IAMF (Immersive Audio Model and Formats) open audio standard.  Ecplisa Audio is currently the only immersive audio format supported on YouTube, and there is actually a growing collection of content available on that platform in the new format.

Get Connected

For many owners, the only input you’ll need is the HDMI/eARC port. Use an HDMI cable to connect this port to the corresponding HDMI/eARC on your TV and any devices connected to the TV, as well as the tuner built into your TV and any streaming apps built into your TV will automatically pass audio to the soundbar. If your TV doesn’t have “eARC” (extended Audio Return Channel) but does have an “ARC” (Audio Return Channel) HDMI port, then you can use that port, but just know that the sound quality is a bit limited on this older option. If your TV lacks both ARC and eARC HDMI ports, then Samsung still has you covered with a fiberoptic digital input, though, like ARC, sound quality over fiberoptic digital is limited: you’ll only get 2-channel PCM sound or 5.1 channel Dolby Digital.

The Q990H soundbar also includes not one but two additional HDMI ports. So if you like to listen to movies with DTS:X sound, but your TV doesn’t support DTS passthrough, you can plug a device like a UHD Blu-ray player into the bar directly, and the bar will output the video signal to the TV while directly decoding the audio signal.

The HDMI input jacks are a little tight, however, making it tricky if you want to plug in a streaming stick or if your HDMI cables are thicker than average. Still, I appreciate the flexibility here and wish more soundbar makers would follow suit.

It’s Got The Look

In terms of build quality and aesthetics, the bar itself feels and looks substantial and its contemporary design blends well with modern TVs. Its height is just under 3 inches which means it won’t interfere with the IR remote control sensor on the bottom of most TVs if placed on a console in front of the TV. The bar comes with a wall mount bracket, but be sure the top of the bar is not blocked by the TV or you will lose some or all of your height channel effects.

The rear speakers are fairly compact at roughly 5″ x 8″ by 5.5″ and include a standard 1/4″-20 UNC threaded mounting hole for a wall bracket or stand. A basic keyhole mount would have been nice but isn’t essential. Be sure to mount the speaker with sides and top unobstructed so the side and top-firing drivers can reflect off side walls and ceiling respectively.

The bar also supports both Bluetooth and WiFi inputs including Apple AirPlay 2 and Google Cast, though it does not support Dolby Atmos sound over Google Cast (few soundbars or components do). Spotify Connect and TIDAL Connect are both supported, but not Qobuz Connect. You’ll need to use Bluetooth or Google Cast if you’re a Qobuz user.

The Set-Up

Although you can set-up and use the bar without loading any apps, you’ll get some additional options and settings if you install Samsung’s SmartThings smart home control app and add the Q990H to Smart Things. The subwoofer and rear speakers are pre-paired to the bar, so if you simply connect an HDMI cable from the bar’s HDMI/eARC port to the HDMI/eARC port on your TV, and plug the bar, subwoofer and rear speakers into wall power, you should get sound from all of the speakers.  

After doing the initial set-up, I installed the SmartThings app on my phone, added the bar to the app and hunted around for any sort of calibration option. Most high-end soundbars either use your phone or an external microphone to measure the speakers’ output in your room and adjust the levels and, in some cases, even phase and EQ, to optimize the sound for your specific room. But Samsung does things a little differently.

If you enable “SpaceFit Sound Pro” in Smart Things, the bar automatically adjusts its sound to fit your room, without the need to run through a calibration routine. It uses the microphone built into the bar to analyze reflections and frequency response and adjusts itself to correct for any deficiencies in the sound. While I can see the appeal of doing this whole process automatically, it bothers me to think the bar is adjusting sound while I’m listening to it. This makes me feel like maybe it’s not presenting the most accurate reproduction of the sound coming into it at all times.

I asked Samsung for details on how the SpaceFit Sound Pro feature works, but did not receive a reply in time for publication. If we get a more detailed explanation, I will be sure to update this review.

In any case, I will say the SpaceFit Sound Pro setting did appear to improve the sound in my listening room and there weren’t any egregious artifacts of its operation in normal use. I’d say most owners will benefit from having it enabled, and those who prefer not to use it can leave it off. The SmartThings app does offer manual EQ and level settings for each channel for those who wish to tweak the sound manually (though no built-in test tones).

For system review context, I connected the Q990H to a Samsung S95H QD-OLED TV. This way I was able to test out Samsung’s Q-Symphony feature which lets you use the TV speaker as part of the mix. In testing, I preferred the sound without Q-Symphony enabled as it changed the tonal balance of the sound coming out of the center channel. But owners an experiment with this in the TV’s Audio Settings menu. By using a Samsung TV, I was also able to play YouTube videos encoded in the Eclipsa Audio format, which the Q990H was only too happy to decode.

Listening Notes

Before I got too into music tracks and movie scenes, I decided to try a few test tracks, including channel test tone sequences for Dolby Atmos, DTS:X and Eclipsa Audio. All three were decoded properly on the Q990H with excellent test tone placement all around the room. Samsung bills the HW-Q990H as an “11.1.4” channel system, and while I didn’t have a Dolby Atmos 11.1.4 channel test patterns, the system did reproduce a Dolby Atmos 9.1.6-channel test tone sequence with excellent positioning. The phantom middle height channels were nicely positioned between front and rear height channels along the middle of my ceiling and the side surround channels did appear right around the middle of the room on each side, thanks to phantom channels created by wide side-firing drivers in both the soundbar and the rear channel speakers.

I will say the Q990H had the best image positioning of any of the bars I’ve tested to date, including the Klipsch Flexus Core 300 system and the Sony BRAVIA Bar 9 with its much-touted “360 Spatial Sound Mapping” feature. 

This performance on test tones carried over to real music. I listen to a lot of immersive music tracks in Dolby Atmos as you can find on my Amazon Music Dolby Atmos playlist. On the EDM track “Alive” by KX5/deadmau5, around 4 minutes in, when the snare roll rotates around the room starting in the front and traveling the full width and depth of the room, the circular motion was fairly seamless, though there were some minor tonal differences as it moved around the room from speaker to speaker. This type of precise motion falls apart on many lesser systems. And the tonal matching across the 20+ speaker drivers was pretty good overall. 

Another of my Dolby Atmos favorites, “Rocket Man” by Elton John starts mostly in the front of the room, but when the chorus kicks in, the music just explodes, taking over the entire listening space. This moment was particularly effective and dramatic on the Q990H as the music just inhabited the entire listening space, sucking you right in. Pulling out a few more classic rock tunes remixed in Dolby Atmos, The Who’s “Baba O’Riley” opens with a synth keyboard part that has been expanded to encompass the entire listening space, shimmering from side to side, and front to back, while vocals are placed more traditionally in the front and center of the soundstage. As the instrumental conclusion builds, instruments like violins, guitars and drums make full use of the space leading to a controlled chaos of sound. The Q900H maintains this chaotic build nicely. 

There’s also a nice selection of 90s grunge/alternative tracks remixed in Dolby Atmos. On Stone Temple Pilot’s “Creep,” Dolby Atmos is used more to open up the soundstage, rather than make instruments or voices spin around the room. On the Q990H, the simple opening instrumental section offers solid imaging and nice tight extended bass on the bass guitar and kick drum. When Scott Weiland joins in with the first verse, his voice is squarely placed front and center, filling out the wide, deep soundstage. Much better than the stereo mix, IMHO. 

More modern albums, like Justin Gray’s Grammy Award-winning album, “Immersed” take more liberties with instrumental placement. “Immersed” was specifically mixed to put you inside the mix, giving the listener a whole different perspective on the music. The track “Orion’s Belt” leads off with percussive elements starting at the back of the room and slowly swirling, building around the listener while layers of drums and horns bounce around the listener. Through the song, you’re right in the middle of the action, like a fly on the wall, only there is no wall. The immersive nature of “Immersed” is well captured on the HW-Q990H.

Moving onto Dolby Atmos movies and TV shows, I put on the opening scene of “Andor” Season 1. Rain falls convincingly from above as our antihero walks down a boardwalk in Morlana 1 in search of his lost sister. As he enters the night club/brothel, dialog is still clearly audible over the throbbing dance music in the background. And in “Dune” around 1:05 into the film as they are rescuing the doomed spice crawler, Paul Atreides is struck by the spice-infused wind and all sound drops out. The swirling spice, musical score, effects and voices of the Bene Gesserit witches build into a cacophony of sound. As it peaks with the line “Kwizatz Hadderach Awakes” the sound is captured well by the Q990H, not fractured and scattered as it can be on lesser systems.   

Testing DTS:X tracks, I loaded up the UHD Blu-ray of “The Blues Brothers.” The mall chase scene was even more chaotic than I remember, with squealing tires and smashing glass coming from all around as police cars chase the Blues Mobile through a crowded shopping mall. When the main performance begins with Cab Calloway, the DTS:X track fills the room with the sounds of a live concert hall. And as Jake’s psychotic ex-girlfriend tries repeatedly to murder him, bullets ricochet menacingly around the room (mostly in the rear). It’s effective use of space and the Q990H captures it well.

I mentioned Eclipsa Audio earlier. In order to get this to work on YouTube content, you do need to match the Q990H up with a recent (2025/2026) Samsung TV with support for Eclipsa Audio, or any TV with support for IAMF audio (the “non-branded” version). Also, make sure the TV’s eARC output is set to “Auto” and Digital Audio is set to “Auto” or “Bitstream” (not PCM). With these TV settings in place, and the TV connected to the Q990H on the HDMI eARC port, you should get Eclipsa Audio from compatible content. And if you have trouble, watch this quick tech tip video on YouTube.

By searching for “Eclipsa Audio” on YouTube (using the Samsung TV’s YouTube app), I was able to find several videos encoded in the format, all of which played back properly in Eclipsa Audio format on the Q990H with discrete sounds coming from all around and above me. If you’re feeling nostalgic for the classic “Deep Note” trailer from THX, you can check out this New Version of THX Deep Note “Spark” Encoded in Eclipsa Audio on YouTube. Other tracks include 4K demo videos and music videos, some of which are pretty entertaining. Over time, we would love to see this format adopted by more hardware vendors and more content creators.

Stereo Music and Notes on Listening Modes

Stereo music had a nice sense of space on the Q990H, particularly when using the Q990H sound modes. There are a few different listening modes from which to choose, which affect how the bar handles both 2-channel and multi-channel content. You can find these by hitting the “Sound Mode” button on the remote or by looking in the Smart Things app.

“Standard Mode” will keep your stereo music unadulterated, using the bar itself and the subwoofer to create a nicely balanced stereo soundstage. “Surround” mode sounds more like a “Multi-channel stereo” mode where sound from the front is cloned to the rear speakers so that the sound fills the room more evenly. This works well for a party when you want background music to fill the room. Personally, I find that the “AI Adaptive Sound” mode works well for most stereo material. It expands stereo music to use all of the speakers, with enhanced spatiality, but without sounding too forced or artificial. But this is a personal choice. It’s nice to have options.

In multi-channel listening (e.g., Dolby Digital, Dolby Atmos), the modes do similar things, but not precisely the same as with stereo sources. Standard mode will be a “pure” representation of the original 5.1 or 7.1.4 mix. In Standard mode, many of the speakers in the system will be silent (such as the front wide and rear side speakers) as a direct representation of the original 5.1 or 7.1.4 channel mix. In “Surround” mode, 5.1 or 7.1 or 7.1.4 content is “upmixed” to 11.1.4 to make full use of all speakers. This fills in the space between the speakers well and improves the overall spatiality of the sound.

For multi-channel sound sources, “AI Adaptive Sound” does what surround mode does, but adds AI-based EQ and localization to enhance the sound based on analysis of the type of content – or scene – being played. Action sequences may have the bass boosted while quiet scenes with whispered dialog will have slight emphasis added to the center speaker while reducing some of the ambient sounds. If you want to keep things “pure” for surround sound movie viewing and music listening, “Surround” mode is a good compromise as it makes full use of all the speakers in the system, without making any artistic decisions, based on AI analysis. Personally I found the AI Adaptive Sound mode to work pretty well overall on most material.

By the way, if you’re a gamer, you might want to check out the “Game Pro” mode, which accentuates some of the directionality of sounds, to make these more pronounced. It also prioritizes latency so you can hear things that might be important (like footsteps behind you) without delay. 

Overall, the sound from the Q990H was nicely balanced. Voices were presented naturally with good dimensionality and instruments like bass and snare drum had nice snap and attack. Overall bass response was solid and full, without sounding boomy, though the bass was not quite as deep or extended as I’ve heard it from larger, more powerful subwoofers.

The Bottom Line

When Samsung acquired the revered audio company Harman International several years ago, I was hopeful that this would improve the sound of the company’s audio products. And it seems like Harman’s influence is definitely rubbing off on the Korean tech giant. Earlier Samsung speakers and soundbars that I tested didn’t really stand out in sonic quality. But the HW-Q990H is different. With everything you’ll need in the box, all the essential codecs covered, and surround sound imaging that matches or exceeds the best competitive systems, the HW-Q990H gets a definite recommendation from me and earns our Editors’ Choice for 2026.

Pros:

• Supports Dolby Atmos, DTS:X and Eclipsa Audio
• Excellent spatial imaging on immersive movies and TV shows
• Solid performance on stereo music
• Independent speaker level adjustments and EQ
• Tight controlled bass with more oomph than you’d expect from a small subwoofer
• Fairly affordable for a flagship system with subwoofer and rear speakers included

Cons:

• Low bass not as extended as systems with larger subwoofer cabinets
• No manual calibration or room correction procedure (automatic “SpaceFit Sound Pro” mode only)
• Not IMAX Enhanced certified

Our Ratings

★★★★★★★★★★ Performance

★★★★★★★★★★ Usability

★★★★★★★★★★ Build Quality

★★★★★★★★★★ Value

Where to buy:

$1,997.99 $1,597.99 at Amazon | Best Buy | Crutchfield

Related Reading

Claude Cowork — Anthropic’s Claude Code-style agent for general knowledge work — is coming to your phone. 

Claude Cowork launched as a desktop app in January, but starting Tuesday it is available on web and mobile for Max subscribers. With the update, users can start a task from their desk, get status updates on their phone, and pick up the finished output later — even if their laptop is closed. 

The product expansion is a signal that Anthropic wants Cowork to feel less like a coding tool for dummies and more like an agentic administrative coworker: something that can work in the background, tag along across devices, and request human input when a decision pops up only the user can make. 

In other words: the coding agent wars are spilling into the rest of the office.

The move comes as AI firms try to push their products beyond chatbots into the everyday surfaces where work actually happens. OpenAI has made a similar move with Codex, which began as a software development tool but is increasingly being used by non-developers for reports, spreadsheets, presentations, research, data analysis and more. 

For both labs, the bet is that success will depend less on who has the best chatbot and more on who owns the space where work gets done.

That push also extends to other apps. Anthropic recently launched Claude Tag, an always-on Claude that lives in Slack and acts as an AI teammate.

Beyond the benefits of one specific interface, launching Cowork as a multi-platform app means that the agent can continue running tasks in the background without a device online, the company says.

One example from Anthropic reads: “Set Monday’s client prep for 6 am: Claude works through the email threads, transcripts, and recent news, builds the briefing doc, and leaves the follow-up email drafted but unsent. Review it over coffee.”

The desktop app will remain the place for deep work, where Claude can access local files and the browser. But bringing Cowork to web and mobile means people who didn’t install the app can also use it. Anthropic says chat and Cowork will be unified in web and desktop to start, with projects and artifacts living together across both.

Anthropic also released early Cowork data, which suggests the clearest use case for the tool is the “work around the work” that keeps companies functioning, handling what Anthropic calls the “tasks that are part of a broad swath of jobs, but are rarely a person’s core responsibility.”

The study sampled 1.2 million anonymized and aggregated Cowork sessions from more than 600,000 organizations over the last two weeks of May.

The largest category at 33.4% was business process operating: pulling scattered updates into a single report, building onboarding checklists, and reconciling spreadsheets. Anthropic said the tasks are common among roles in finance, HR, and administration. 

The next largest category at 16.4% was content creation and copywriting: tasks like drafts, slide decks, social posts, proposals, and other communications work that is usually performed by marketing and management positions. Software development, by comparison, only accounted for 8.7% of Cowork usage.

“While coding is still—understandably—one of the uses of AI that gets the most attention, the use of AI for everyday business work is on the rise, and the kinds of tasks people are finding it most helpful for are coming into focus,” Anthropic said in a blog post. “Our goal is to make this a reference point for people who are figuring out how to integrate AI products into their daily work, and to show where value is most concentrated.”

SYSTEMS

Everyone else must opt out manually if they don’t fancy settings data shipped off-device

Microsoft is enabling Windows Backup for Organizations by default in Windows 11 26H2 everywhere except the EU, meaning businesses elsewhere with sovereignty and privacy concerns will be forced opt out instead.

Now dubbed “Windows settings backup and restore,” the service backs up a device’s settings and a list of installed Microsoft Store apps, which can then be restored to a new device.

Microsoft gave a use case for the technology: “Imagine a lost laptop, a hardware refresh, or an unexpected reset. These are some of the moments when your users need backup most. And that’s rarely when anyone wants to discover that backup was never turned on.”

However, some organizations might not want it on. Perhaps those with strict privacy or data sovereignty requirements, or those regulated by the EU Digital Markets Act (DMA), for whom the default-on behavior won’t apply. Windows 11 25H2 and earlier are also excluded, as is any device with a backup policy that explicitly disables the setting. Everything else running Windows 11 26H1 will get switched on after a feature update, and the same applies to 26H2, currently with Windows Insiders in the Experimental channel.

Administrators might reasonably be wary of this being opt-out rather than opt-in. Backups are useful, but Microsoft is clear that this is not a comprehensive backup solution, calling it only “one step in a broader Windows resiliency effort.” The implications still need consideration.

An opt-out setting that quietly ships settings data off-device is exactly the sort of thing that adds to administrators’ workloads rather than lightening them.

Microsoft’s recommendation is to leave things as they are. “Eligible devices with the backup policy in a Not Configured state under Windows settings backup and restore will enable backup automatically at general availability of Windows 11, version 26H2.”

Anyone who doesn’t want that must explicitly disable the policy, which “always takes precedence over the default,” Microsoft added.

Before crediting Microsoft for making this feature default to on, consider its stated objectives for Windows Backup for Organizations: to “Help organizations accelerate PC refresh cycle or the transition to Windows 11 or deploying AI-powered PCs,” and to “Allow organizations to transition to a cloud-first approach for managing devices and user settings.” ®

• New Zealand’s Education Minister denies any plans to restrict or ban VPNs
• Reports previously alleged it was part of the teen social media ban package
• Prime Minister Christopher Luxon also confirmed “no plan to ban VPNs”

The New Zealand government has officially denied any plans to restrict or ban VPN apps as part of its upcoming under-16 social media ban, putting an end to intense speculation and a rapid backlash from digital privacy advocates.

The saga began following a report from The Post that Education Minister Erica Stanford said the government was considering any restrictions on VPNs as part of the country’s under-16 social media ban.

Imagine your refrigerator sits in another building, 100 metres from your kitchen. Every time you cook, you walk over for each ingredient, then walk back to check that you closed the fridge door. That could be another long walk back if you forgot the milk for your morning coffee.

Until the agentic era, this was the norm. Data could live in that fridge and get pulled when needed. Applications and humans didn’t need millisecond or even live data to make important decisions; humans can work on copies. But that era is ending. Agents think and act in instants, in context. And very soon billions of them will be working 24/7/365. They don’t pull a copy and decide later. They need to be governed in the moment, in the context of that moment, and they need to act fast and at reasonable cost. Agents cannot run to a lakehouse, or a fridge, and still meet those requirements.

That means intelligence has to be where the agents and data are acting.

Think about the exponential rise in digital fraud in payment systems, or the volume of retail returns from digital purchasing. We live in a more complex, integrated data world, and we expect real-time resolutions, solutions, and choices.

The lake was never built to run a business (or agents)

Your AI and your data need to be available the moment an agent acts: Petabyte-scale data, served in real time, live, not copies, without a walk to the fridge every time. You cannot retrofit a lakehouse to deliver that.

Everyone now agrees that the old separation between transactional and analytical systems had to end. The interesting question is what replaces it. This month Databricks offered its answer: LTAP, or Lake Transactional/Analytical Processing. Built on Lakebase, its serverless Postgres^®, LTAP puts transactions and analytics on a single copy of data in the lakehouse. It is interesting engineering, but built from the wrong end.

The reason is straightforward: the only gravity that matters is the data. Action happens at the data layer, governance has to happen at the data layer, so the data layer is where you build rather than somewhere you move data to. Pulling transactions up into the lakehouse is like moving the house and kitchen to the building with the fridge.

A lakehouse is, at bottom, built on a data lake, and the lake was built for analytical work: Large scans, append-heavy patterns, eventual consistency, object-storage economics. Transactions want the opposite: Low-latency reads and writes, strict consistency, row-level locking, the hard ACID guarantees that operational applications have relied on for 40 years. You can engineer a transactional layer onto object storage credibly enough, but you are swimming against the substrate the entire way. The lakehouse is a magnificent place to analyze data, but a strange place to run your order book.

The operational database is where transactions already live. It is consistent, governed, the system of record. Agents don’t act on copies. They act on the real thing, live and governed, right where it sits. The durable architecture doesn’t haul that into the analytical world and re-solve consistency from scratch. It starts from the operational core, the place the business actually runs, and extends analytics, vector search, and agents outward from there, against the same live, governed data, without moving it.

The destination everyone is describing is the same: one copy, no pipelines, one governed surface for every workload, whether OLTP, HTAP, or agents. The divergence is the starting point. A lakehouse-first model decides in advance that the data belongs in the lake, then pulls transactions up to meet it. Starting from the operational core presumes nothing: The data stays where it already is, and everything comes to it.

Opposite starting points compound: the gap between the two only widens the further you build.

For regulated enterprises, true sovereignty is nonnegotiable

A lakehouse is a cloud service, on the cloud’s object storage, under the cloud’s control. For an enormous share of the enterprises that most need agentic AI (banks, hospitals, telcos, governments), “move your transactional system of record into our cloud” is not a deployment detail. It is a nonstarter. These organizations operate under data-residency rules, sovereignty requirements, and, in some cases, air-gap mandates that no amount of elegant lakehouse architecture can make go away. You cannot regulate your way around where the bytes physically sit.

This is not only about what regulators permit. Where data does its work should be the enterprise’s choice in the moment, not a destination a vendor decided in advance.

The moment an autonomous agent can act on regulated data, sovereignty stops being a preference and becomes a constraint. An operational core built on open Postgres runs wherever the data has to be: on-premises, hybrid, across clouds, air-gapped if the regulator demands it. A lakehouse, cloud-bound by design, runs where the vendor’s cloud runs. For the regulated enterprise, that single fact settles the question before any benchmark is run.

Govern where the data is, not in a catalog above it

Governance works the same way. The lakehouse model governs through a catalog, a policy layer administered above a collection of engines. That is a reasonable design for a platform assembled from many parts. For an autonomous agent acting directly on data, a governance layer that lives somewhere other than the data is a governance layer with a path around it.

Governance has to be enforced by the database itself, through the same roles, row-level security, and audit trail that already govern human access. Govern where the data is, at the moment of action, not in a catalog hovering above it.

The market is already moving

This is the shift, and the market is confirming it. The most prominent name in the lakehouse world is now racing to embed an operational Postgres core, spending roughly $1 billion to acquire Neon to get there. When the company that defined the lakehousestarts building toward the operational database, the direction of travel is no longer in dispute. The only question left is which end you build from.

The enterprises that get this right will build from the operational core outward, on open Postgres, on infrastructure they own. Transactions, consistency, governance, and sovereignty are the hard constraints; analytics is the part that should come to them, not the reverse. Your AI, your data, your rules, on infrastructure you control.

Join the Era of Agentic AI with EDB Postgres AI. Watch the global digital event on demand: https://www.enterprisedb.com/join-the-era-of-agentic-ai-edb-postgres-ai