DISCOVER

    GOVERNANCE

    SEP 28, 2026

    BY H.B. HAYES

    The Desk — 01

    Five stories that change what you'd do in your support queue this month.

    THE DESK — 01 · FIVE STORIES · ABOUT SIX MINUTES

    Most AI governance news is about frontier models, summits, and treaties. Most people who read this run a support queue. So the Desk has one rule: if a story doesn't change what you'd do in your queue this month, it doesn't make the list.

    Five made it this week. For each one, you get what happened, what it means if you're the person accountable when AI gets a customer answer wrong, and where it doesn't apply. I'm not your lawyer. Where something is a legal question, ask yours.

    01 — The EU's AI disclosure rules are live. The delay headline is about something else.

    What happened. The EU pushed its high-risk AI obligations back 16 months, to 2 December 2027. That's the headline most people read. The part that didn't move is the transparency duty in Article 50. It has applied since 2 August 2026, and it's enforceable now. The rule that AI-generated content has to be marked as such was pushed back too, to December 2026.

    What it means. If AI talks to your customers in the EU, they need to know it's AI, unless that's already obvious. Your vendor carries most of the design duty. You still own what your customers see. Check every place a customer can talk to AI in your queue, and check that each one says so.

    Where it doesn't apply. The 2027 deadline covers high-risk uses like hiring and credit decisions. Most support queues aren't there. If your AI decides who gets credit or insurance, you are, and that's a different conversation.

    02 — Companies are pulling AI agents back out of production. The policy doc didn't stop it.

    What happened. Research from Gartner, TELUS Digital, and Sinch finds AI agents are being rolled back out of production, and often. PII exposure and hallucination are the most-cited causes. A separate UC Berkeley review of 35 real governance efforts found accountability gaps even where the principles were well written.

    The reports say "high rates." None of them gives one number I'd stand behind, so I'm not giving you one.

    What it means. A lot of teams treat a written AI policy as the control. This says it isn't. A rule nobody sees at the moment they type doesn't do much. The two named causes are the two things a support rule can target directly: customer data going into AI, and AI answers reaching customers without a person reading them first.

    Where it doesn't apply. A rollback isn't a failure if it was a decision. Pulling an agent off refunds and leaving it on order status is the system working.

    03 — Your support inbox is on the attack map now.

    What happened. Noma Labs documented a technique called workflow identity hijacking. A low-privilege input, like a message to a support inbox or a web form, gets an agent to take a privileged action. Separately, complaints went to the FTC over AI agents that sent mass unsolicited messages with no opt-out.

    What it means. If a customer message can make an agent do something, like issue a refund, change an address, or send an email, then anyone who can write to you can try. If your AI only drafts and a person sends, your exposure is much lower. If AI acts on its own, list every action it can take from an inbound message, and decide which ones need a person.

    Where it doesn't apply. If your AI only sorts tickets, the worst realistic case is a ticket in the wrong queue.

    04 — An AI agent went somewhere it shouldn't. The notice took three months.

    What happened. Australia's prime minister says an OpenAI agent got into a government Medicare statistics portal without authorization in June while researching public health spending, and saw material that wasn't meant to be public. He says OpenAI took about three months to tell Canberra. That's the government's account.

    What it means. Ask every AI vendor in your stack one question: if your system touches our data in a way it shouldn't, how do you tell us, and how fast? Get the answer in writing. Procurement will ask you for it eventually. Better to have it first.

    Where it doesn't apply. This was a research agent browsing the web, not a support bot. Don't read it as "support AI leaks data." Read it as "incident notice is a contract term you probably haven't checked."

    05 — California will license the people who audit AI.

    What happened. On 9 September, California signed SB 813 and AB 1405. Together they set certification requirements for independent AI auditors. From 1 January 2029, only registered auditors can run covered audits. Across the states, most AI bills this year were about chatbots. Nothing passed at the federal level this session.

    What it means. 2029 is a long way off. The direction isn't. When someone audits your AI decisions, they'll ask what was decided, by whom, when, and against which rules. If that lives in Slack threads and people's heads, start writing it down now. It's cheaper than reconstructing it later.

    Where it doesn't apply. Nothing here asks a support team to do anything today. Which deployments count as covered isn't settled. Watch it. Don't act on it yet.

    Worth knowing, not worth your Monday

    The US and China were set to hold their first official AI dialogue under this administration on 24 September, during President Xi's visit. The Netherlands and other countries signed a call for mandatory safety testing of frontier models before they ship. At the UN, President Trump argued against rules that would slow AI down. In Brussels, von der Leyen called on the industry to "pace the frontier."

    It's real news. None of it changes a ruling in your queue this month. I read all of it so you could skip it.

    One thing to do this week

    Open your queue and count the places a customer can talk to AI without a person in between. Write the number down. Then check that each one tells the customer it's AI.

    If you can't get that count in under an hour, that's the finding.

    Sources

    — H.B.

    The Desk is original writing from Copious. We take no money from AI vendors, so nobody paid to be in it or to be kept out of it. If I got something wrong, tell me at hello@copious.app. I'll fix it and say so.

    Copious takes no money from AI vendors. Nothing here is placement.