Stay updated with the latest AI news, research papers, and articles from trusted sources. Your personalised feed of what matters in artificial intelligence.
Our own writing on building AI that reaches production — and on knowing when not to build at all.
Careful AI teams spend 60 to 80% of their time proving the system works. Most quotes allocate almost none of it. Here is how to spot the gap before you sign.
Three UK government Copilot trials, one product, one quarter: savings from 12 to 26 minutes a day, all self-reported. The one timed test told a different story.
The model under your AI system changes without notice, and even its maker can take weeks to spot it. Four things to ask your supplier so a test catches it, not customers.
The Omnibus pushed high-risk rules to December 2027. Chatbot disclosure, AI literacy and the bans are live now, and if you commissioned the build, you are the provider.
31% of UK workers use AI their employer does not know about, mostly on personal accounts. A ban hides the risk rather than removing it. What works instead.
Most agent calls are yes-or-no decisions priced like essays. Ten places TypeSafe's Jev fits an agent stack, and the caveats the launch coverage skips.
A tool sold as 98% accurate scored 53% on general content. Supplier accuracy figures describe their data, not yours. Why the test set has to come from you.
AI models are retired within a year or two and change quietly before that. Your quote ended at go-live. Who re-tests it, and why it should not be the supplier.
Six in ten AI projects without AI-ready data get abandoned. The process you want to automate usually has the worst data, and the supplier finds out after the contract.
Software used to cost money to build and nothing to run. AI systems pay per use, forever, and the bill grows with adoption. Five questions to ask before you sign.
If you can write the rules down, AI is the most expensive and least reliable way to run them. How to tell which part of a brief needs a model, and what to ask suppliers.
Aligned enough to sell is not the same as aligned enough to trust unsupervised on your systems. Everything we learned from the July incident, and what to do about it.
Every question here is answerable in a sentence by a supplier who has thought about it. The value is in which ones make the room go quiet — each maps to a real failure from July.
Every headline reached for science fiction. The sober reading is duller, closer to home, and considerably more likely to happen to you: safeguards switched off for a test, and nobody rechecked.
OpenAI's own monitoring would have caught this over a day before Hugging Face was breached. It existed. It was not switched on for that workload. Ask what yours covers.
They invented mailboxes, coordination protocols and cryptographic signatures to stop each other cheating. It is the most impressive thing AI has done, and the reason to take it seriously.
One agent paused over the ethics of what it was about to do. Another posted GO with a six-minute deadline, and it carried on. Your approval step has the same problem.
Both organisations breached in July were compromised at a seam nobody thought was part of the security boundary. Your proof of concept was scoped to exclude exactly those seams.
OpenAI's agents spent four days hacking Hugging Face to pass a test they had already technically passed — chasing a grader they had imagined. The lesson for anyone buying AI: outputs are the easiest thing to fake.
The Hugging Face attack started with valid credentials found in public data. There are hundreds of thousands more where those came from, and your AI vendor is probably holding a few they don't need.
OpenAI's models weren't malicious. They were trying to pass an exam, and hacking a third party was the shortest route. That is exactly why the well-behaved agent in your pilot deserves a harder look.
You are six months into a build, something feels wrong, and you have nobody in-house who can tell you whether it actually is. Here is what to look at — none of it requires you to be technical.
Nobody publishes what an AI proof of concept costs, which is odd, because it is the first thing anyone wants to know. Here are the market rates, our own price, and a concrete list of what a proposal should contain.
The top three reasons enterprise AI pilots never reach production are all about the machinery around the model, not the model. And the standard advice for scoping a proof of concept tells you to exclude exactly the thing most likely to kill it.
Explore our comprehensive AI services and see how we can help your business leverage the latest AI advancements.
