AI for small businesses: what actually changed in June 2026

Most of what gets written about AI for small businesses is a list of tools. June 2026 was not a tool month. It was the month the American government told private companies who was allowed to use their products, and that is where Aziz Musa spends the episode, because it decides what you will be able to build on next year.

The month opened normally enough. OpenAI released 5.6 X, faster and cleverer than anything before it, particularly at coding. He explains why every one of these companies puts coding first: a model that codes well can code itself, and he says 90 per cent of Claude is now written by Claude rather than by a team of programmers. After coding come biology and cyber security. The jump was large enough that the American government immediately restricted the release. Named people only, until we understand what it can do. His verdict on that is mild. They are getting ahead of a problem before it happens.

Then Claude released Fable 5 and Mythos 5. Aziz used Fable 5 for about a day and a half before the American government stopped it: nobody outside America, and nobody who is not American, could use either model. No reason was published at the time. What came out afterwards, in his account, is that the models were finding cyber security holes in the American government's own systems. In April, Mythos had been set on the kind of large systems Facebook, Meta, Instagram and Google run, turned up a quantity of vulnerabilities, and the intention had been to give it to those companies free to fix their own security.

He does the arithmetic out loud, because it is the fastest way to show what happened. If you take a company public, what people pay depends on how many customers you could reach. Claude's number was seven billion people. The restriction made it the two hundred million in America, and Claude had filed to go public in October on the basis of a model for the world. He expects the gap to show in the share price.

The specific problem was solved. Changes were made after the government's request, a new version of Fable came out later in the month, and you can use Fable again. Mythos 5 has been released, but only to companies working in cyber security.

What is left is not about what happened. It is about what can now happen. A government has established that it can stop a private AI company at any moment, for any reason it likes. He calls it a war running in front of our eyes that most people cannot see, notes that the Chinese have their own versions of all of it, and says the plain thing: we are not ready for this.

The weaker your base, the more AI gives you the illusion that you are strong.

His answer to the whole month is not a tool recommendation. AI is an addition to your base knowledge, and it behaves according to what it is added to. If you are capable and your foundation is strong, it makes you excellent. If you have no foundation, or a weak one, it eats your time and produces what he calls slop: work out of a machine that nobody understands and that carries no value.

He says it to the graduates at his company and he says it to his daughters. Do not spend your time learning how to use AI. Spend it building the base. Learn both if you can, but give 70 per cent of the time to the foundation: read, take courses, do an MBA if you can afford one, and apply what you learn. The last line of the episode is the same argument in one sentence: if you do not understand how to ask and what to ask, AI will not help you.

Years ago in Sudan a man rang his mother. Mona, it is Mohamed, I am at the airport, there is a problem with customs, send me 5,000. She asked which Mohamed. Mohamed from Atbara. She said she did not know any Mohamed from Atbara and put the phone down. That scam was everywhere at the time, and it depended on a voice you might half recognise.

The same call now comes in a voice you cannot tell from a real one. Android has shipped voice clone detection on its phones, which tells you whether what you are hearing is a person or a model, and he expects it on iPhone next. He had been worried about this since February, when he said the danger was a call that feels like your bank, from someone holding all your personal details. Deepfakes are in voice, in video and in email, and the people using them to steal move forward at the same speed as everyone else, so you are permanently a step behind.

The change with the most immediate use is at Meta. Any business with a Facebook page, an Instagram account, WhatsApp or Messenger can now train an AI agent to answer customers in the company's own voice, instead of the fixed reply telling people you will get back to them within 24 hours. Aziz gives his own reason for wanting it: people write to him in Arabic, he cannot read Arabic, and the messages run into the thousands. A trained agent can at least handle the repeated questions.

The other one is a retry. Around 2012, Facebook launched a search that let you find the best restaurant near you based on what your friends had said about it, and it failed completely. They are building it again with AI, along the lines of the AI mode in Google, with the knowledge base being what people do on Facebook, Instagram and Messenger. They say WhatsApp and Messenger are not read, and his position is to take them at their word and watch whether the thing works.

The item he enjoys most is from science. A researcher had spent three years on a question about how glucose affects T cells, and had tried a quantity of models against it. ChatGPT Pro produced the best result and the best reasoning on it inside two hours, and resolved a set of problems in immunity along the way.

That is the direction he wants: scientists and pharmacists and doctors moving their own fields, rather than the capable tool ending up with whoever wants to use it for harm. Every month brings something new. Learn AI, and come to it carrying something of your own.