The month the plumbing got settled
May produced one genuinely structural change and a great deal of spectacle. Sorting the two is the whole job of reading this stuff.
Three large conferences, two frontier models, a governance reversal and a video model that generates its own soundtrack. May was loud. One thing in it will still matter in 2027.
The one that matters
On 19 May, Microsoft adopted the connection standard across its products. Combined with OpenAI in March, that is effectively every major platform on one specification, six months after publication.
This matters because integration cost is the reason most SME AI projects never start. Not model quality, not price — the question of whether the thing will talk to the ERP you bought in 2016. That question is becoming structural rather than a bespoke negotiation per vendor.
Nothing to do today. In two years it will be the difference between an AI project costing three weeks and costing three months.
The ones that were spectacle
A video model with synchronised audio is remarkable and changes nothing about how a Swiss SME runs its business this quarter. The March rule covers it.
Two new frontier models is a Tuesday now. The useful reading is the one from 28 February, unchanged: do not ask which model is best, ask whether you can change it.
What we wrote about instead
Looking back at the month, the posts likely to be read in a year are the unglamorous ones: what prompt injection actually is, what a realistic AI budget looks like, what has to be true for a tool to connect to your ERP, and why translation is the capability with the clearest business case in this country and the least attention.
None of those were prompted by a release. All of them were prompted by questions Swiss SMEs actually ask.
Into June
The 2 August deadline is nine weeks away. If you wrote the page in January and reviewed it in April, the July review is the one where you add the supplier answers from the 4 April post.
And the recommendation that has survived every month since October: one process, cleaned data, a named person who checks the output, run for a month before the second. Nothing in May changed it, which after a month this noisy is worth saying out loud.