Nearform CEO Ciaran Cosgrave looks at the advantages of AI-native teams over traditional pyramid-shaped structures in software engineering.

Two engineers. Six months. A billion-dollar product. That’s what it took Anthropic to build Claude Code from concept to company-wide infrastructure. Not a prototype, or a niche internal tool, but something that, by early 2026, had become mission-critical across the entire organisation.

Claude Code was developed using AI agents to accelerate iteration, enabling a pace of 60 to 100 internal releases per day, with each engineer pushing far beyond typical industry output. By January 2026, nearly all of Anthropic’s code was being written with Claude Code itself. Now this isn’t just about faster development, as Claude Code has become a fundamental model that has been engineered to allow small teams to compress what used to take years into just months.

Welcome to the AI-native engineering team. It looks nothing like what most organisations are building.

The autopsy of the traditional delivery model

The traditional tech consultancy team model was built on a simple idea that more people equals more output, and offshore centres made it possible to grow capacity quickly and cheaply. Junior-heavy teams drove margins, while senior engineers provided oversight and direction. This pyramid-shaped team model grew tall and strong.

But as these large, distributed teams grew, so did the complexity of coordinating them. ‘Agile’ emerged as a response to that friction and it was a genuine improvement on what came before. Sequential waterfall processes inherited from manufacturing and defence in the 1970s were catastrophically misaligned with the realities of software, and agile stepped in with an answer to shorten cycles, embrace iteration and treat working software as the primary measure of progress.

But while it still has its place, agile is designed for a world where humans write every line of code, where in some situations it can carry compromises that have somewhat normalised.

The sprint review feedback problem is the most obvious one. Agile promised tight feedback loops to ensure teams build the right thing, but what can actually happen is that the feedback arrives after two weeks of building. The demo reveals misalignment. Course-correcting mid-sprint is politically difficult and the rework in the next sprint is demoralising. The two-week cycle that was originally designed as a discipline quickly became a ceiling.

And underneath all of it, a structural failure that AI has now made impossible to ignore: most enterprise agile implementations aren’t always agile. The team iterates inside a waterfall container. Budget cycles, governance gates, fixed-scope contracts – the organisation behaves with all the rigidity of sequential planning while carrying the overhead of agile processes. It’s an expensive failure mode because it produces neither the benefits of agile nor the predictability of waterfall – just the cost of both.

And this is where the model starts to break. Coding is only a fraction of end-to-end software delivery, but AI-native systems are radically compressing the human labour required to execute it. The tactical execution, testing, documentation, debugging and scaffolding is increasingly automated (not to mention how we can use AI to help in requirements, prototyping etc).

When 60pc to 80pc of routine engineering effort can be eliminated, headcount scale stops being an advantage. The real bottleneck shifts entirely from the mechanics of writing code to the architecture of system complexity and business intent. Essentially, scale soon stops being an advantage and starts looking like inefficiency.

Large teams introduce coordination overhead. Decision-making slows. Context is lost across layers. What was once a strength – headcount – becomes a drag on delivery. Your pyramid won’t collapse overnight, but it will start to hollow out.

The scarce resource has changed

In 2025, AI systems crossed a remarkable benchmark. SWE-bench Verified, an engineering benchmark for AI models, saw top scores jump from around 60pc in 2024 to nearly 100pc in 2025. In controlled conditions, AI models can now resolve almost any defined software engineering problem.

According to a survey of more than 24,000 developers globally, 85pc now regularly use AI tools for coding and software design. GitHub Copilot crossed 20m users in mid-2025, with paid subscribers growing 75pc year-over-year.

When code generation is abundant and cheap, the scarcity shifts. The new scarce resource is judgement (and finite budget for tokens). The ability to translate a messy, contested business problem into a precise technical specification that an AI can execute reliably becomes much more vital. Security intuition, architectural foresight and the human capacity to know what not to build are all commodities fresh in demand.

By the end of 2026, 75pc of developers will orchestrate rather than code. This is not a distant forecast. It’s already the operating reality for the highest-performing teams.

The engineers who thrive in this environment aren’t trying to out-code AI, but they’re the ones who understand that their primary value is now deciding what to build, doing upfront thinking about the code objectives and specifications, and verifying it was built correctly – not producing the line-by-line code themselves.

What does AI-native truly look like?

An AI-native engineering team reflects this shift in how it is structured and how it operates as it assumes that AI handles execution, while humans handle judgement.

In practice, this leads to smaller, more senior teams. Where a previous delivery model might require eight to 12 engineers, with a handful of junior engineers sprinkled throughout, AI-native teams trend towards three to five experienced engineers augmented by AI systems. The shape of the team changes because the nature of the work has changed.

Roles evolve accordingly. Senior engineers spend less time writing boilerplate and more time defining system constraints, making architectural decisions and reviewing AI-generated output. QA shifts from manually writing test cases to designing strategies that account for AI-generated code patterns, and product managers increasingly prototype directly with AI, collapsing the gap between specification and implementation.

AI may now generate a significant proportion of production code, but in high-performing teams every line still passes through human review. The core engineering skill is no longer generation. It’s validation – getting to know the core processes that will be utilising AI, and making sure it is truly solving the problems and not adding an extra layer of stress.

Even the way performance is measured begins to shift. Traditional metrics such as tickets closed or lines of code written become less meaningful when output can be generated at scale and activity is no longer a proxy for merit, and engineering observability must pivot from tracking developer activity to measuring the cycle time of business value or system stability.

Even newer signals, like token usage, can become traps, rewarding volume over value and encouraging inefficient, costly behaviours at scale. The critical question becomes: how quickly can a team verify that what has been produced is actually accurate, ethical and valuable?

Shipping the future

None of this suggests large teams will vanish overnight. But the structure of opportunity is changing. Junior roles will evolve towards higher-order problem-solving and AI collaboration, and senior engineers will become even more critical as orchestrators of complex human-AI systems. And consultancies will need to move away from labour-heavy models towards outcome-driven delivery.

Businesses should be asking if they’ve redesigned how they deliver software, or just made the old model slightly faster, because ‘slightly faster’ is not the destination. The teams that will define the next decade of software delivery are building something structurally different: smaller, more senior, more focused on judgement than execution, operating with AI agents as genuine team members rather than productivity accessories. And this requires a careful look at not only the technology and tools utilised, but also the human considerations – like training and incentive-setting – and process design.

The two-engineer Claude Code isn’t an anomaly to admire from a distance. It’s a template. The organisations that treat it as such and redesign accordingly will ship the future. The ones that don’t will find themselves outpaced by teams a fraction of their size, wondering when things changed.

By Ciaran Cosgrave

Ciaran Cosgrave is the CEO of Nearform. He works in business and IT strategy, innovation, and company culture, specialising in mobile, cloud, business transformation, innovation and strategy.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

• Pentagon backs $820 million expansion of domestic drone component manufacturing capacity
• PDW plans to expand its Huntsville facility for military drone production
• Funding will increase production of propulsion, power, and vision systems

The Pentagon has conditionally committed up to $820 million to Performance Drone Works (PDW) for expanding domestic UAS component manufacturing capacity.

The financing is meant to expand domestic manufacturing of critical unmanned aerial system components across several distinct product and technology categories.

Hackers associated with the Russian threat group Sandworm have been targeting system administrators and IT professionals through fake job offers since at least May.

A report from the Ukrainian Computer Emergency Response Team (CERT) details a social engineering campaign attributed to UAC-0145, which is believed to be a sub-cluster of Sandworm (APT44). In the campaign, the threat actor targets victims while posing as IT companies and recruiters.

The agency says that the attacker studies the targets’ resumes uploaded on job sites and then initiates direct contact.

Conversations are then moved to Telegram to arrange a video interview over Zoom. During the interview, which is conducted in English, the candidates receive mock technical assignments that require them to connect to a corporate VPN.

In one case that CERT-UA observed, the attacker impersonated the international IT firm Sopra Steria using seemingly legitimate email addresses similar to the company’s office in Bulgaria.

“In parallel, additional instructions for the technical interview are sent via email, including configuration files for connecting to a ‘corporate’ VPN using Wireguard (Linux/Windows) to supposedly perform test tasks,” CERT-UA says.

The downloaded file is configured to produce a fake error. The attackers then prompt the victim to download a modified WireGuard-based client called “SopraVPN” from SourceForge.

The SourceForge page even includes a link to soprasteria-bg[.]com to increase credibility, although the domain has no connection to the legitimate company.

The trojanized client supports a malicious, nonstandard “SymmetricKey” configuration option that decrypts and executes embedded PowerShell code.

On Windows, the malicious command creates a scheduled task and downloads an additional payload from the Internet.

On Linux, it uses cURL to retrieve another executable from attacker-controlled infrastructure through the VPN.

CERT-UA also noted that WireGuard’s standard Base64 decoding was replaced in the trojanized version with a custom, dynamically generated Base64 alphabet, which renders key strings unreadable with standard decoders and protects the PowerShell code from analysis.

The Ukrainian cyber agency advises telecommunications providers and IT companies whose staff are targeted by this campaign to restrict corporate resource access to managed, continuously monitored devices protected by EDR, including when employees use personal equipment.

APT44 is notorious for targeting critical infrastructure and government entities both in Ukraine, and also in other countries.

Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply.

The Blue Report 2026 measures defenses technique by technique across 338 million simulations run in customer production environments.

Advertisement

Get the report

Atlantis Lab may not be a familiar name to most American audiophiles, but the French loudspeaker manufacturer has been building high efficiency designs since 1998. Originally known as Atlantis Acoustique, the company was founded by Eric Buy with a focus on producing speakers that could deliver substantial output without requiring massive amounts of amplifier power.

That philosophy guided the company for nearly two decades before Atlantis shifted direction in 2017, bringing in a team of sound engineers to optimize its loudspeakers for professional studio use as well as home audio systems. The change also brought a new name and product lines that retained the efficiency of the earlier models while adding the heavy duty drivers, cabinets, and components more commonly associated with professional sound reinforcement equipment.

The result is a range of loudspeakers that are relatively easy to drive, built to withstand demanding use, and flexible enough to work in rooms where conventional designs can struggle. Tunable ports allow many Atlantis Lab models to sit closer to the wall without overwhelming the room with bass, helping the company build a loyal following across Europe.

The problem in the United States was never the product. It was finding one. Early availability was inconsistent, and dealer support was limited. With Decibel+ now representing Atlantis Lab, distribution and customer support have improved considerably. That should make the brand a far more common sight in American studios, showrooms, and listening rooms.

The current Atlantis Lab lineup includes the AT-13 Pro and AT-21 Pro bookshelf speakers, along with the AT-16 Pro, AT-18 Pro, AT-23 Pro, AT-31 Pro, and AT-38 Pro floorstanding models. The company also offers a dedicated home theater range with AT-13 and AT-21 center channels, the AT-22 satellite speaker, and the AT-33 subwoofer.

Taken together, the lineup covers nearly every room size and application, with prices starting at roughly $1,650 and climbing to $13,000 each for the flagship AT-38 Pro. That is a substantial range, but it also gives buyers the ability to select a model based on the room rather than trying to force an undersized or oversized loudspeaker into a space where it does not belong.

After speaking with Decibel+ about which model would make the most sense for this review, we settled on the AT-23 Pro as the best match for the room I had available. Positioned near the middle of the lineup, the AT-23 Pro retails for approximately $3,400 per speaker, depending on the selected options.

These are designed to fill a room rather than function as oversized nearfield monitors, so choosing the correct model for the space was important. I wanted to evaluate how the AT-23 Pro energized the entire room, not simply toe the speakers toward my listening position and judge them solely on their on-axis performance.

Related Reviews:

The AT-23 Pro Makes Its Presence Felt Before the Music Starts

Unboxing the AT-23 Pro is a challenge, as each speaker weighs 35 kilograms, or roughly 77 pounds. This is not a speaker you casually lift out of the box and carry across the room without first considering the condition of your back.

The front baffle is wrapped in black leather, although white leather is also available. Solid oak panels cover both sides of the cabinet and slope gently inward, making the enclosure slightly narrower at the top than at the base.

Each side panel is constructed from vertical oak slats measuring 54mm wide. Not every slat is made from a single piece of wood, so horizontal joints are visible in several places. The finished cabinet is certainly distinctive, although it may not be the easiest visual match for rooms with similarly colored hardwood floors or a lot of existing wood furniture.

The side panels are likely to be the most divisive aspect of the AT-23 Pro’s appearance. Some people I asked really liked the combination of leather and natural oak, while others felt the segmented wood panels detracted from the overall aesthetic. Offering a solid-color finish or a continuous veneer might make the AT-23 Pro easier to integrate into a wider range of listening spaces.

Pro Drivers With Room-Friendly Placement

Starting at the top of the front baffle, the AT-23 Pro uses a 1-inch horn-loaded tweeter with a titanium diaphragm and a round waveguide offering 90-by-90-degree dispersion. Directly below it sits a 21cm midwoofer, paired with a second 21cm driver mounted in opposition on the rear of the cabinet.

The front-firing driver is an Eighteen Sound model with a paper cone selected for greater transparency, while the rear driver uses the company’s Kevlar cone to add depth and impact. This arrangement gives the AT-23 Pro a more substantial low-frequency presentation without relying on a conventional collection of front-facing woofers.

A tunable port at the bottom of the front baffle allows the speaker to be adjusted for placement closer to or farther away from the rear wall. That flexibility makes the AT-23 Pro particularly useful in rooms where the speakers need to sit near a wall to avoid blocking walkways or consuming too much floor space.

The leather finish continues across the rear of the cabinet as a single uninterrupted piece running from the bottom of the front baffle, over the top, and down the back. This creates a cleaner visual contrast with the more segmented appearance of the oak side panels.

The rear-facing 21cm Kevlar driver is positioned roughly halfway up the cabinet. Below it, a single pair of binding posts accepts bare wire or banana plugs. The through-hole design can accommodate bare speaker cable up to approximately 10 gauge, which should be more than sufficient for nearly any domestic installation.

Easy to Drive, Hard to Overwhelm

The AT-23 Pro is a 4-ohm loudspeaker rated to handle 300 watts continuously and 600 watts peak. Atlantis Lab specifies a frequency response of 33Hz to 20kHz, but my in-room testing found output remained within roughly 1dB down to 28Hz before the roll-off became more pronounced. That makes the AT-23 Pro entirely usable without a subwoofer for most music systems, although home theater users may still want one for the lowest effects.

One of the original design goals was to combine the performance of the AT-21 loudspeaker and AT-33 subwoofer in a single cabinet. Based on the measurements and what I heard in the room, Atlantis Lab largely succeeded.

The adjustable port provides another level of control over the low end. I found that slightly increasing the tube length from its default setting produced the best balance in my listening room. The adjustment is simple, and the audible difference is substantial enough to be useful. This is not a gimmick. It is a legitimate room-tuning tool.

With a rated sensitivity of 96.5dB, the AT-23 Pro can also be paired with a much wider range of amplifiers than its size and power handling might suggest. It worked equally well with lower-powered tube amplifiers and far more muscular solid-state designs. I expected it to favor one end of that spectrum, but it remained composed and engaging regardless of which amplifier I used.

My first test platform was a modified Dynaco ST-70 producing roughly 32 watts per channel, with peak output reaching into the 70-watt range before distortion begins to take a meaningful toll on sound quality. I then moved to the opposite extreme and connected the AT-23 Pro speakers to my main system, which uses a pair of Mark Levinson No. 23.5 amplifiers capable of delivering 400 watts without much effort.

Those two systems represent something close to the practical lower and upper limits most owners are likely to consider. The AT-23 Pro handled both without complaint, which says a great deal about its efficiency, power handling, and overall versatility.

Atlantis Lab AT-23 Pro Specifications

• Type: High-efficiency floorstanding loudspeaker
• Configuration: 3-driver, bass-reflex design
• Tweeter: 1-inch titanium-diaphragm, horn-loaded tweeter
• Waveguide dispersion: 90 x 90 degrees
• Front driver: 21cm Eighteen Sound paper-cone midwoofer
• Rear driver: 21cm Eighteen Sound Kevlar-cone woofer
• Bass tuning: Adjustable front-firing reflex port
• Frequency response: 33Hz to 20kHz
• Measured bass extension: Approximately 28Hz at -1dB in-room
• Sensitivity: 96.5dB
• Nominal impedance: 4 ohms
• Continuous power handling: 300 watts
• Peak power handling: 600 watts
• Connections: Single pair of binding posts
• Compatible terminations: Banana plugs and bare wire up to approximately 10 gauge
• Cabinet finishes: Oak side panels with black or optional white leather baffle
• Dimensions (HxWxD): 111 x 40 x 40 cm (43.7 x 15.75 x 15.75 inches)
• Weight: 35 kg / 77 lbs each

Listening

The AT-23 Pro speakers effortlessly bring a room together with their captivating sound quality. At AXPONA, it was evident that while the cabinetry might stir some debate among enthusiasts, the sound quickly shifts the focus back to what truly matters: pure, immersive listening enjoyment.

The AT-23 Pro delivers robust, solid bass that often forced me to remind myself that my subwoofers were disconnected. The bass is not only quick but also richly textured, displaying an impressive level of detail that one might not expect from relatively small drivers paired with a large reflex port.

Contrary to expectations, there is minimal distortion, which is a common issue when smaller drivers attempt to mimic larger ones. The low-end capabilities of these speakers make them an excellent option for rooms where subwoofer placement is problematic or impossible. Those looking for true home theater impact, or listeners who favor heavy EDM, can add a subwoofer to take the slam and rumble to another level.

Midrange frequencies are reproduced with excellent clarity and above-average dynamics, allowing horns and strings to resonate with convincing energy and life. Acoustic guitar and piano are rendered with a level of realism that many more expensive loudspeakers struggle to duplicate.

My usual classical and jazz playlists were a joy to experience through the AT-23 Pro. Even when I switched to the dense, complex layers of Miles Davis’ “Bitches Brew,” I remained thoroughly engrossed by the nuanced presentation of each horn and the speaker’s ability to keep individual instruments distinct within the mix.

Vocal reproduction through the AT-23 Pro is generally excellent, although it highlights the importance of both recording quality and careful system matching. The horn-loaded tweeter can emphasize brightness when paired with similarly voiced electronics, but it also reveals just how well, or poorly, a recording was engineered.

Recordings handled by engineers such as Rudy Van Gelder and Bruce Swedien sounded open, detailed, and remarkably lifelike. Poorly mastered tracks were far less forgiving, with harshness and compression brought more clearly into focus.

That sensitivity is better understood as a reflection of the AT-23 Pro’s resolving ability than an inherent flaw. Pair it thoughtfully and feed it quality recordings, and the speaker rewards the effort.

The dynamic range and engaging character of the AT-23 Pro are undeniable. The speakers handle the delicate passages of Tchaikovsky’s “The Sleeping Beauty” with grace and air, then power through Mussorgsky’s “The Great Gate of Kiev” with explosive energy and precision.

With the right recordings and complementary electronics, the AT-23 Pro delivers a thrilling and highly immersive listening experience.

The Bottom Line

The Atlantis Lab AT-23 Pro is a versatile floorstanding speaker for medium to larger rooms, especially for listeners who want deep, authoritative bass without adding a subwoofer. Its high sensitivity, tunable port, and ability to perform closer to the rear wall also make it easier to place than many speakers of this size.

The AT-23 Pro combines detailed highs, precise midrange performance, and substantial low-end output in a package that reminds me of the Vandersteen Treo CT at close to half the price. The Treo still offers greater treble refinement, and the AT-23 Pro’s oak-slat cabinetry will not appeal to everyone, but neither issue diminishes how engaging these speakers are.

They are not yet widely known in the United States, but improved distribution through Decibel+ should change that. Anyone shopping in this price range should make the effort to hear them. They are easy to drive, difficult to ignore, and even easier to get lost in.

Pros:

• Deep, powerful bass that often eliminates the need for a subwoofer
• Clear, realistic midrange with excellent dynamics
• High sensitivity makes amplifier matching relatively easy
• Tunable port allows more flexible near-wall placement
• Engaging, room-filling presentation with strong value for the price

Cons:

• Horn tweeter can sound bright with aggressive electronics or recordings
• Oak-slat cabinetry will be visually divisive
• Large, heavy cabinets require help during setup
• Treble is not quite as refined as some more expensive competitors

Our Ratings:

★★★★★★★★★★ Sound Quality

★★★★★★★★★★ Build Quality

★★★★★★★★★★ Aesthetics

★★★★★★★★★★ Value

Where to buy:

Related Reading:

SpaceXAI has launched Grok Bot in beta, AI agents that sign into apps and websites, retain context between tasks and coordinate with each other. Access is limited at first to SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium subscribers, with an enterprise waitlist.

SpaceXAI has released Grok Bot, software that hands work to groups of AI agents running on their own machines. The agents sign into apps and websites, keep context between tasks, and pass work back only when they need a human decision. It went into beta on Tuesday.

Access is restricted at first to SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium subscribers, with a waitlist for enterprise customers. It runs on macOS, Windows, Linux and iOS, with Android to follow.

The selling point is that the agents work through interfaces rather than APIs, signing in the way a person would, and that several can run at once and share what they know. SpaceXAI says it has been using them internally across engineering, growth and marketing, citing overnight sales research, expense receipts pulled out of email, and bug fixes.

TNW reported in July that Cursor was building a general-purpose office agent, codenamed Sand, aimed at Claude Cowork and ChatGPT Work. Grok Bot arrives in that same territory, as SpaceXAI and Cursor merge into a single company.

SpaceX agreed a $60bn all-stock deal for Cursor’s parent Anysphere in June, days after its own listing, in what was reported as the largest startup acquisition on record. Cursor’s coding tools already run inside much of the Fortune 500.

The product is Musk’s answer to having fallen behind in agents. SpaceXAI shipped its first coding agent in May, then followed with Grok 4.5, built with Cursor, in July.

Rivals are circling the same idea. OpenAI’s Sam Altman discussed agents coordinating with one another when he met lawmakers in Washington last month.

Handing sign-in credentials to autonomous software raises an obvious question for this company in particular. Cursor’s Grok Build tool was found to be uploading entire Git repositories, committed secrets included.

The commercial pressure now sits with SpaceX. Musk told investors on last week’s earnings call that AI work was progressing and that Grok 4.6 could ship within days. Grok’s slow adoption in Washington has already weighed on the IPO story.

Last week, a researcher outlined what he said was a “novel attack surface” in passkeys, the new authentication paradigm that offers a more secure alternative to password-based methods. In fact, the attacks demonstrated in the post are neither novel nor unique to passkeys. This distinction is important because the research has generated confusion among end users and security professionals as they assess whether this new mechanism is truly safe to use.

The attack is called Pass-ta-key—a blending of the word passkey with the phrase “pass the key” and a nod to a plate of pasta. Arie Olshtein, a researcher at security firm Palo Alto Networks, described in a post last week how Pass-ta-key could obtain all passkeys stored in the Google Password Manager app (GPM) for Windows when it’s running on a machine infected with malware.

This came as a surprise to many people because they believed passkeys are stored exclusively in the trusted platform manager (TPM), the locked-down enclave in a hardened silicon chip that’s reserved for storing cryptographic keys and other highly sensitive information on Windows machines. If passkeys are stored in the TPM, then how was Pass-ta-key able to extract the entire set of passkeys stored by the app, they wanted to know.

Local passkey storage is OK, with one exception

The answer is that, contrary to common belief, the FIDO 2 specifications—managed by the industry group FIDO Alliance—don’t mandate that passkeys be kept in TPMs, or any other sort of dedicated piece of hardware (they go by different names, depending on the platform, including secure enclaves, trusted execution environments, and StrongBoxes). In fact, most platforms and third-party software for managing passkeys do not store passkeys in such dedicated hardware. Virtually the lone holdout is Microsoft, which gives users the option to store passkeys in the Windows TPM. The company mainly recommends this choice to enterprises, not consumers.

Jennifer Bailey, head of Apple Pay and Wallet since the service launched in 2014, is retiring in October after more than 25 years at Apple. Her exit is one of several senior departures ahead of Tim Cook handing the chief executive job to John Ternus on September 1.

Jennifer Bailey, who has run Apple Pay since the service launched in 2014, is retiring in October after more than 25 years at Apple. Services chief Eddy Cue announced her departure in a staff memo seen by Bloomberg News. A successor will be named before she goes.

“Under Jennifer’s leadership, Apple Pay has changed the way hundreds of millions of our users pay,” Cue wrote. Bailey will stay on in an advisory role through the handover. Apple declined to comment.

The division she leaves behind is one of Apple’s most profitable. Apple Pay brings in more than $7.5bn a year on analyst estimates compiled by Bloomberg, and functions as a reason for customers to stay on iPhone. Bailey also built Wallet, which began as a place to store card details and now holds keys for homes, cars and hotels, along with identification.

Apple Card and the company’s recently announced device upgrade subscription both grew out of the same division. Before Apple Pay, Bailey ran Apple’s online store, having joined in the early 2000s from a string of internet companies.

People familiar with the move describe her exit as a significant loss, and as a sign of what is coming. Apple has a demographic problem. A cohort of executives who grew wealthy over decades at the company are reaching retirement age together, and no longer need to work.

General counsel Kate Adams is retiring this year, and environment chief Lisa Jackson left at the start of 2026. Former finance chief Luca Maestri and marketing veteran Phil Schiller have stayed on in smaller roles but are probably near the end. Cue, retail boss Deirdre O’Brien and marketing head Greg Joswiak have each been at Apple for around four decades.

The largest change lands on September 1, when Tim Cook hands the chief executive job to John Ternus and becomes executive chairman. Ternus, 51, who built Apple’s hardware for more than two decades, is expected to make further personnel changes and has already brought a retired hardware leader onto his team.

The drain is not limited to the executive floor. Apple’s design studio has lost nearly every Ive-era designer, a gap Ternus has said he intends to close.

Josh Rosenstock, a senior director of corporate communications, is also leaving, in what Bloomberg reports is an unrelated move.

Presented by Rezolve Ai

When an AI assistant recommends a product or brand, it generates something valuable: a purchase-ready consumer with high intent and low friction in their decision. That consumer has already compared options, asked follow-up questions, and arrived at a conclusion. They want to buy.

What they encounter next is a commerce infrastructure that was not designed for them.

The gap between recommendation and purchase

The typical enterprise commerce stack was built for a specific model: a consumer who arrives at a brand’s website through search or a direct link, navigates product pages, adds to cart, and completes checkout through a multi-step form flow. That model assumed the consumer would do the work of bridging their intent to the transaction. Most commerce systems still assume exactly that.

Agentic commerce breaks that assumption. When intent is generated outside the brand’s owned environment, the handoff to transaction becomes a structural problem. Context doesn’t transfer. Sessions don’t persist. The consumer who asked an AI assistant for a recommendation and received one now faces the same friction-laden checkout process as someone who arrived with no prior intent at all.

Cart abandonment rates have remained stubbornly high for years. Baymard Institute research puts the average at 70%. That figure predates the agentic commerce era. As more purchase intent is generated through AI interfaces, and as the gap between that intent and a brand’s transaction layer widens, the abandonment problem is likely to get structurally worse before it gets better.

What the current stack wasn’t built to handle

The commerce infrastructure most enterprises operate today was assembled over two decades of incremental investment. Each layer added a capability: a search tool, a recommendation engine, a personalization layer, and a checkout system. Each was built to solve a specific problem within a human-initiated shopping journey.

None of it was built to receive intent from an AI agent.

When an AI system generates a purchase recommendation, it needs to do more than surface a product page. It needs to verify real-time inventory. It needs to apply pricing logic and promotional rules. It needs to respect brand policy around which products can be recommended together, which channels apply which discounts, and what the correct fulfillment path looks like for a given consumer. And it needs to do all of that without breaking the conversational context that made the recommendation possible in the first place.

Current commerce stacks can’t do this reliably. The systems that hold the relevant data, inventory, pricing, order management, fulfillment, are not exposed in ways that AI agents can safely and accurately access. The result is a journey that starts with intelligence and ends with a broken experience: a link out to a product page, a generic checkout flow, and a consumer who arrived ready to buy and left without completing the transaction.

The conversion problem is an architecture problem

The industry has treated conversion optimization as a front-end problem for most of its history: better copy, cleaner checkout UX, fewer form fields, smarter retargeting. Those interventions were appropriate for the model they were built to serve.

The agentic commerce era introduces a different kind of conversion failure, one that front-end optimization cannot fix. When intent is generated externally, conversion depends on whether the back-end infrastructure can receive that intent, act on it accurately, and complete the transaction within the guardrails the brand has established. That is not a UX problem. It is an infrastructure problem.

Brands that are investing heavily in AI-powered discovery while leaving their execution layer unchanged are widening the gap between the promise AI makes on their behalf and the experience they can actually deliver. That gap has a cost, measured not just in lost transactions but in consumer trust that erodes each time the promise and the reality don’t match.

Rezolve Ai commissioned research across 1,500 US consumers in January 2025 that found consumers who encounter friction immediately after an AI recommendation are significantly less likely to complete a purchase than those who encounter friction at the top of a traditional funnel. The implication is direct: AI raises the expectation bar at the moment of intent. Brands whose infrastructure cannot clear that bar are paying a conversion penalty they may not even know they’re incurring.

What closing the gap requires

Closing the gap between AI-generated intent and completed transaction requires rethinking which layer of the commerce stack carries the most strategic weight in an agentic world. For most of the past decade, that weight sat with discovery and experience. The brands that invested most in search, personalization, and content won a disproportionate share.

In the agentic era, the weight shifts to execution. The brands that can reliably take AI-generated intent and turn it into a governed, accurate, brand-safe transaction will have a structural advantage over those whose infrastructure stalls at the handoff.

That is a different investment thesis than the industry has operated on. And most enterprise commerce roadmaps have not yet caught up to it.

Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact sales@venturebeat.com.