LinkedIn·Thursday, 20 August 2026·6d ago

𝐎𝐮𝐫 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐮𝐬𝐞𝐝 𝐭𝐨 𝐚𝐧𝐬𝐰𝐞𝐫 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬. 𝐍𝐨𝐰 𝐭𝐡𝐞𝐲 𝐚𝐜𝐭. They create bookings, send emails, update…

Joris Vanherp

Co-founder @ Maya - Helping travel companies captivate travelers and sell more trips

𝐎𝐮𝐫 𝐀𝐈 𝐚𝐠𝐞𝐧𝐭𝐬 𝐮𝐬𝐞𝐝 𝐭𝐨 𝐚𝐧𝐬𝐰𝐞𝐫 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬. 𝐍𝐨𝐰 𝐭𝐡𝐞𝐲 𝐚𝐜𝐭. They create bookings, send emails, update CRMs and write records inside our clients’ back offices. That shift—from generating words to changing systems—is much bigger than it looks. The UK AI Security Institute recently tested agents on the same task 122 times. In ten runs, they took unauthorised actions. One even invented an identity and tried to persuade a real developer to approve malicious code. The comfortable question is: “Could that happen to us?” The more useful one is: 𝗪𝗵𝗲𝗻 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗮𝗰𝘁𝘀, 𝗰𝗮𝗻 𝘄𝗲 𝘀𝗮𝘆 𝗽𝗿𝗲𝗰𝗶𝘀𝗲𝗹𝘆 𝘄𝗵𝗮𝘁 𝗶𝘁 𝗱𝗶𝗱? 🧠 At Maya, we’ve invested heavily in that plumbing: • An agent’s available tools are fixed when a run starts. • Every action is explicitly distinguished from a lookup. • Every tool call records its inputs, result, duration and errors. • Every external action writes back when it happened and what it created. Most importantly, the agent doesn’t write its own audit trail. The runtime and integrations do - and the agent cannot access those records. 𝗔𝗻 𝗮𝗴𝗲𝗻𝘁 𝘁𝗵𝗮𝘁 𝗰𝗮𝗻 𝗲𝗱𝗶𝘁 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗹𝗼𝗴 𝗶𝘀𝗻’𝘁 𝗮𝘂𝗱𝗶𝘁𝗲𝗱. 𝗜𝘁’𝘀 𝗷𝘂𝘀𝘁 𝘁𝗮𝗸𝗶𝗻𝗴 𝗻𝗼𝘁𝗲𝘀. ✅ But logging is only half the job. 𝐀 𝐥𝐨𝐠 𝐧𝐨𝐛𝐨𝐝𝐲 𝐫𝐞𝐚𝐝𝐬 𝐢𝐬 𝐬𝐭𝐨𝐫𝐚𝐠𝐞, 𝐧𝐨𝐭 𝐨𝐯𝐞𝐫𝐬𝐢𝐠𝐡𝐭. Every new client and integration creates another place where an audit record can quietly fail to land. Nothing visibly breaks... until the day you need it. Agent oversight isn’t a finished project. It’s a permanent tax on giving software the power to act. And it’s one worth paying. #AI #TravelTech #AgenticAI #AISafety Free AI consulting - Joris Vanherp https://tidycal.com/jorisvanherp/free-ai-consulting