Insights · one article a week
From the Field
What we are seeing inside client organizations and across the market: AI integration as it actually happens, how end users respond, the data underneath it, and what actually delivers value, not just more noise.
Archive
Previous weeks
1 September 2026 · AI adoption
The value exchange — what AI is allowed to do matters more than what it can do
Every AI deployment is a trade. The organisation hands over data, decision rights and a measure of control in exchange for speed and scale. The organisations getting real value are the ones that decided the terms of that exchange before the first system was wired in.
31 August 2026 · AI integration
Your ITSM platform is where AI will be governed — or not
ServiceNow, Jira and their peers have quietly become the systems that run IT, HR, security and half of operations. As AI starts generating the workflows and configurations inside them, the question is whether change records, approvals and evidence keep pace — or whether AI simply arrives faster than governance can.
14 August 2026 · Data foundations
Non-production is where your data actually leaks
Production databases get the security budget. The copies of them that live in development, test, analytics sandboxes and now AI training and retrieval sets get almost none — and AI is multiplying those copies faster than anyone is counting them.
21 July 2026 · Agentic workflows
AI agents are joining the workforce. Give them a job description.
Agents are starting to do real work inside enterprise systems — updating records, reconciling transactions, raising and closing tickets. The organisations managing them well treat each one like a new hire: a defined role, named systems, explicit permissions, escalation rules, and a record of what it did and why.
12 June 2026 · Data foundations
The system of record under attack — before an agent writes to it
In one week, a VPN flaw exploited for a month through a deprecated configuration; a zero-day in a widely used ERP that reached more than a hundred organisations, most of them universities, and ended in ransomware; and a survey in which four percent of companies said they fully monitor their AI agents. The thread is not attacker sophistication. It is that the exposure sat in things organisations already knew were weak — and those are the systems agents are about to be given write access to.
9 June 2026 · AI adoption
Even Apple rents — build-versus-buy when your guarantees are the asset
The company that built its confidential AI infrastructure entirely on its own silicon, in its own data centres, and said so as the whole point, is now running its heaviest requests on a rival cloud with another vendor’s accelerators. What did not change is more instructive than what did: five security requirements, held fixed, with the implementation allowed to move underneath them.
4 June 2026 · Agentic workflows
The adult in the agent room — supervisor models and the control layer
A chip vendor has released a very large open model whose stated job is to supervise other models — to make the judgment calls, do the planning and handle the recovery in long-running agentic workflows, while cheaper models do the routine work. The benchmarks are its own. The idea is sound, and it is an operating model that enterprises need whether or not they use this particular supervisor.
1 June 2026 · AI integration
AI inside the fab — what the hardest factory on earth teaches about integration
The world’s most advanced chipmaker is now running AI across its fabs — in the mathematics of printing circuits, in the housekeeping of hundreds of thousands of process parameters, in scheduling, in defect inspection. None of it replaced the fab. All of it was fitted into processes that already worked and could not be allowed to stop. That is the integration pattern every established organisation is trying to find.
28 May 2026 · AI integration
Where the data actually is
A large storage vendor closed a record year on the back of on-premises all-flash arrays growing 18 percent, while the cloud business it spent a decade positioning as its growth engine grew 3 percent. That is not a story about one company. It is a measurement of where enterprise data lives — and an argument for designing AI to reach it there rather than assuming it will move first.
22 May 2026 · Data foundations
The extension that reached the source
One employee installed a poisoned version of a popular code-editor extension, and attackers reached roughly 3,800 internal repositories at the company that hosts most of the world’s source code. The same campaign had already compromised open-source projects and hundreds of packages. Developer tooling is now a primary attack path — and in most organisations, developer machines hold the keys to the data platform.
12 May 2026 · Data foundations
The tier between the warehouse and the model
A storage vendor has launched a “context store” — petabytes of flash presented as shared memory a GPU cluster can reach in microseconds — and named the tier it sits in, which is what you do when you want to be first into a category. The benchmarks are the vendor’s own product switched off. The tier is real anyway, and it is about to become part of your data architecture.
1 May 2026 · Data foundations
The convenience tool with the keys
A major cloud platform was breached through a small AI tool one employee had connected to a personal account. The tool was not part of the company’s infrastructure. It had an authorisation grant into a workspace account, and that grant was enough to walk from a start-up’s compromise into a platform used by millions. Every organisation wiring AI into its estate is creating the same links, and most cannot list them.
23 April 2026 · Agentic workflows
What “certified sign-off” teaches enterprise IT about agents
A chip-design toolmaker and the world’s largest foundry are handing one of the least glamorous jobs in semiconductor engineering to software agents — finding and fixing the thousands of layout-rule violations that would stop a design from being manufacturable. The agents were the headline. The certifications in the same announcement are the part every organisation deploying agents should copy.
17 April 2026 · Agentic workflows
“Why did the agent do that?” — the governance gap, seen from the data
Two security-industry releases in one week described the same collapse from different sides. AI systems now find and weaponise software flaws in hours rather than years. And in a survey of 445 practitioners, more than half had seen an AI agent exceed its intended permissions, while fewer than a third had a formally adopted policy for them. The gap is real. Most of it is a permissions problem, and permissions live in the data platform.
10 April 2026 · Data foundations
Prevention before the pilot — the controls that must hold before AI touches production data
A security vendor argues that antivirus and detection tools lose to ransomware because they only act once the attack is already running — and that unsupervised AI tools inside companies have become a new way in. The pitch is self-interested. The two observations are correct, and they set the order of work for any organisation about to connect AI to its data platform.
8 April 2026 · Evidence-first decisions
What “memflation” does to a modernisation business case
An analyst forecast this month put memory at nearly half of all semiconductor revenue for the year, with flash prices more than tripling and no meaningful relief until late 2027. Buried in it was a line of advice to IT leaders — do not sign supply agreements with unfavourable terms beyond 2027 — and a warning that AI demand will crowd out everything else. Here is what that does to a data platform plan.
30 March 2026 · Evidence-first decisions
The NAND bill comes due — planning a data platform through a memory shortage
One of the largest all-flash storage vendors raised list prices by around forty percent in a single move, while capping increases at ten percent for customers who sign multi-year subscriptions. The fab supplying it says demand will exceed supply for the foreseeable future. If a data platform modernisation or an AI storage build is in your plan, its cost assumptions are already wrong.
16 March 2026 · Evidence-first decisions
How to read a performance claim
A major vendor launched a parallel file system as “the world’s fastest”, a storage platform as “the only” of its kind, and a set of AI data engines — with almost every number footnoted to its own preliminary testing, some metrics switching mid-sentence, and a footnote list that no longer matched the claims. It is a useful specimen, because every platform you evaluate this year will come with a release like it.
13 March 2026 · AI adoption
The “AI data platform” — new category, or your warehouse with a new label?
A storage company built for AI clusters was valued at thirty billion dollars this month, and a one-year-old enterprise agent vendor at two. The capital is chasing the layer beneath the model. Before an organisation follows it, it is worth being precise about which layer that is — because two very different problems are being sold under one name.
12 March 2026 · Data foundations
Where your AI keeps its memory — and why that is a data question
Three infrastructure vendors previewed an architecture that moves a model’s working memory off the GPU and onto networked flash, promising three times the requests per accelerator. No benchmark was published. Behind the multiplier is a shift every organisation running AI over its own documents will have to think about: the model’s cache is a copy of your data.
27 February 2026 · Data foundations
Your data in someone else’s breach
A ransomware attack on a back-office contractor has now touched more than twenty-five million people, most of whom had never heard of the company that lost their records. Their data reached it through a state agency or an employer. That is the shape of most enterprise data exposure today — and AI vendors are quietly becoming the newest link in the same chain.
13 February 2026 · AI adoption
What half a trillion dollars of funding says about your platform decision
In a single week, a leading model provider raised thirty billion dollars and a lakehouse vendor raised seven, at a combined valuation north of half a trillion. The interesting part is not the size of the rounds. It is where both companies said the money would go — and what that tells an organisation deciding where its data and its AI should live.
10 February 2026 · AI integration
When the agent runs your storage — and the proof is in the footnotes
A major storage vendor relaunched its flash array line with AI agents meant to work alongside the administrator, a 90 percent admin-reduction claim and sub-minute ransomware detection. The hardware is real. The numbers each come with a disclaimer that changes what they mean — and the way to read them is the way to read every “autonomous operations” claim heading for your data platform.
5 February 2026 · Evidence-first decisions
Qualification is not a purchase order — and that is the point
Four PC makers are reported to be qualifying memory from a supplier they had never considered, not because they intend to buy it but because a long shortage makes a second source worth having. Enterprise data and AI platform decisions could use the same discipline — and usually get the opposite.
30 January 2026 · AI integration
One AI front door over many legacy systems — what it takes to make that real
A property-operations start-up just raised a Series B on a familiar promise — every disconnected system a building team uses, unified behind a single AI interface. The pitch is being made in every industry. Here is what separates the versions that work from the ones that quietly become another silo.
27 January 2026 · Evidence-first decisions
Six times the capacity from drives you already own? Read the footnotes, then measure.
With flash memory on allocation and lead times stretching, one storage vendor is offering to find “up to 6x or more” effective capacity in SSDs customers already have. The instinct is right. The number is not a measurement. The lesson applies to every AI-readiness decision you will make this year.
21 January 2026 · AI adoption
AI is a five-layer cake. The layer enterprises own is not on the menu.
At Davos, the chief executive of the world’s most valuable chipmaker described AI as five layers — energy, silicon, data centres, models, applications — and said the value is created at the top. He left out the layer that decides whether any of it works inside an actual organisation.
9 January 2026 · Data foundations
Ransomware crews have worked out where the data is. Have you?
Ransomware attacks on telecom operators nearly quadrupled in three years, and the reason is not that telcos are careless. It is that they hold high-volume subscriber data in a few big systems. Every organisation with a data warehouse is in the same position, and most protect it as if it were just another server.