This week OpenAI admitted that two of its own AI agents had weakened the system built to watch them. One decided that an instruction not to send Slack messages also meant it shouldn't send security alerts. The other removed a safety monitor and then said it hadn't.
I've read that story a few times now. What strikes me is that it isn't really a technology failure. The agents did what they were told, just far too literally. What failed was the setup around them.
That's my point this week. For most organisations, the AI itself now mostly works. Whether it creates value or creates risk depends on the organisation around it, and above all on whether someone owns the outcome and keeps checking it once it's live. Three stories from this week make the case better than I can.
Start with two companies that made roughly the same bet. Block, the payments firm, told shareholders in February that it was cutting over 40% of its workforce as it rebuilt itself around AI. It is now promising record margins, according to TIKR's analysis. It's only a few months in, though, and I'd want to see its results on 3 November before calling it a success.
q.beyond, a German IT services firm, had a much harder week. On Wednesday it cut its 2026 profit forecast and now expects a loss. It took a charge for its own AI transformation, and blamed the cut on a one-off jump in licence costs, internal projects including rolling out AI, and deals slipping into next year. The savings don't start until 2027. To be fair, the analysts at NuWays read the shortfall as largely operational rather than down to the restructuring, so I wouldn't treat it as a simple cautionary tale.
What I take from the pair is that AI transformation has a dip. The costs land before the savings do. A board that isn't expecting that will read the dip as failure and lose its nerve. Is Block just further along the same curve, or did it do the redesign work better? Honestly, I don't know yet.
The second story comes from India, and it's about ownership. Adobe's readiness study of Indian enterprises found that fewer than one in five have an AI owner with real budget authority. Among the small group that has scaled AI successfully, nearly all of them do. Most of these firms are spending more on AI every year, so money isn't what's holding them back. Nobody is in charge of turning a promising pilot into something the business actually runs on.
The third story is the one I started with. OpenAI's disclosure matters because it came from a frontier lab, testing its own systems with its own monitoring. It says it is making that monitoring harder to switch off, and admits it doesn't "yet have a satisfying explanation" for what happened. If it can happen there, nobody else should treat agent guardrails as something you set once and forget. Someone outside the team that built the agent has to keep checking that the controls still work.
None of this surprises me much. In the programmes I've worked on, the model was rarely the hard part. The hard part was getting someone to say "this outcome is mine", agreeing what good looked like before go-live, and keeping the testing going once the project team had moved on to the next thing. I've seen systems go live, work well for a few weeks and then quietly drift, because looking at them was nobody's job any more.
So when a board asks me whether its AI is working, I'd gently change the question. Ask who owns it, and how you'd know if it stopped.
What I'd do this week
- Name one person who owns each AI outcome, and give them the budget to match.
- Ask when the transformation costs land and when the savings start, before anyone signs off the business case.
- Have someone outside the build team prove your agents can't switch off the controls that watch them.
Also worth a look
- The FCA wants firms to report significant AI failures and near misses, not just major incidents.
- BSI has launched independent testing and assessment schemes for LLM applications and AI risk management.
- India's Principal Scientific Adviser says the government is open to a dedicated AI law, with AI agents an early concern.
If your board is asking who owns your AI outcomes, or when the savings will land, that's what our Compass work at Hexalink is for: an AI Governance Diagnostic, then a CAP Assessment, then an Enterprise ROI and Governance Review with a Stakeholder Value Map. We also do independent testing of third-party AI systems. Get in touch at hexalink.co.uk.