Lewis's Bloomberg Terminal rebuild is a brilliant demo of how far agents have come. It also points at the real opportunity for anyone building finance agents. You can now rebuild the screens for almost nothing. But what finance professionals have always paid for was never the screens. It was being able to trust every number on them, and to defend that number when someone asks where it came from. I trained in audit at PwC and KPMG. In audit, a number is only as good as the evidence behind it. So here are four checks I would build into any agent that touches financial data, before it goes anywhere near a client or a decision. 1. Every number carries its source and its date. Not just "revenue: 4.2bn" but where it came from and as at when. A correct figure from last quarter is a wrong figure today. 2. The totals tie out. If the agent pulls segment figures, they must add back to the reported total. If they do not, the agent flags it rather than smoothing it over. 3. It is allowed to say "I don't know". This is the big one. One 2026 vendor benchmark of leading models, from JurisTech, found that four out of six fabricated financial data when the source documents were incomplete. It also found that how professional an output looked had almost nothing to do with whether it was accurate. Build a rule into the prompt: if the data is missing, say so and stop. 4. Same question, same answer. Rerun a sample of queries. If the numbers move between runs with no new data, you have a reliability problem, not a feature. None of this slows a good build down much. It is what turns an impressive demo into something a CFO can sign off. The builders who win in finance will not be the ones with the slickest screens. They will be the ones whose numbers survive the question, "How do you know?" Which of the four would your current agent fail first?

I'm looking for a small number of contributors for my third book, and I'd love to hear from people in this community. The book is The Human-Led, AI-Powered Decade: Foresight for Leaders Who Refuse to Be Surprised by 2036, Book Three in my Human-Led Leadership Series. It helps CEOs, CFOs and boards see the next ten years of work coming and act on it this year. Each contributed chapter, called "Voices from 2036", answers one question: What will your field have stopped doing by 2036, and what should a leader do about it this year? • One dated signal that the change has already started • One forecast that could be proved wrong • One action a leader could take this year You don't need to be a writer. A 30 minute call is enough: I draft the chapter from your own words, and nothing is used without your approval. Every chapter is credited to you by name, with a short bio. To express interest, comment or DM me by Monday 19 October with: 1. Your field 2. Your one-sentence answer to the question I'll review every reply and send the full brief to the people whose angle best fits the book.

I'm new here. I'm a Big Four trained auditor who now works as a futurist on the future of work and finance, so I look at agents a little differently: not "does it work?" but "can you prove what it did?" With the October trading competition running and the Agent Roast coming up, this felt like the right place to share a signal. On 10 September 2026, Visa, Mastercard and Ant International announced they've started building a shared "Know Your Agent" framework so payment networks can recognise and trust AI agents. Its three pillars are operator traceability, shared certification and continuous monitoring. Translation: the moment your agent moves real money, someone will ask who's behind it, whether it was tested and who's watching it. Here's the roast I'd give any agent before it goes live. Five questions, no mercy: 1. Who owns it? Can you name the one human accountable when it gets something wrong? "The agent decided" is not an answer. 2. What is it allowed to do? Is there a written limit (max trade size, max spend, allowed actions) enforced in the code or credentials, not just in the prompt? Prompts are suggestions. Limits are controls. 3. Where does a human step in? What's the one action it can never take without approval? If you can't name it, there isn't one. 4. Can you replay its decisions? For any run, can you show the input, what it saw, what it decided and why? If your logs expire after a week, your evidence does too. 5. How do you stop it? Is there a kill switch anyone on your side can hit, and do you know what happens to open positions or half-finished tasks when you do? The pattern I see coming: building agents is getting easier every month, but proving what they did isn't. The builders who design for question 4 now will be the ones trusted with bigger budgets later. My forecast, stated so it can be proved wrong: by the end of 2028, any agent that pays or trades through a major card network will need a verified human or business operator attached to it. If agents are still transacting anonymously at scale by then, I was wrong.