--- title: "Can your team trust what AI gives them? How to check it" description: "Why AI gets things wrong while sounding sure, three checks that catch it, and why the person who sends the work still owns it." url: "https://onetwogrowth.com/writing/can-your-team-trust-ai-output" canonical: "https://onetwogrowth.com/writing/can-your-team-trust-ai-output" --- # Can your team trust what AI gives them? How to check it Why AI gets things wrong while sounding sure, three checks that catch it, and why the person who sends the work still owns it. Nov 7, 2026 Harry Siggins [Subscribe](https://harrysiggins.substack.com/subscribe) There's a lot of talk about AI getting things wrong. An AI model that can handle hard problems miscounts how many times a letter appears in a word like "strawberry", or fumbles a simple sum, and the example goes viral. Moments like that unnerve people, and teams tend to react in one of two ways. Some stop trusting anything AI gives them. Others accept what it says without knowing whether they should. And then someone on your own team tries AI on a real task, gets a poor result, and stops trusting it too. So the question leaders bring to me is whether their team can trust what AI gives them, and how they'd know. My answer is that they can, once they set up the work so mistakes get caught, and as long as the person who sends the work on still owns it. Catching mistakes starts with knowing where they come from. ## Why AI gets things wrong and still sounds sure AI makes two kinds of mistakes worth knowing about, and each has a reason behind it. - **Hallucinations.** An AI model can state something that isn't true, with the same confidence as everything else, and without telling you it's guessing. Ask a general AI model for a private company's quarterly figures, which aren't public, and it can give you some anyway. It happens most when the model has no source to work from and fills the gap from what it learned in training. - **Math errors.** A model works with patterns in text, so it can get a sum wrong that a calculator would get right. You don't need the technical detail behind either one to work well with AI. If you want it, OpenAI has explained [why language models hallucinate](https://openai.com/index/why-language-models-hallucinate/). What matters here is that both kinds of mistake are known, so both can be caught, and three checks help. ## Point it to sources and make it show its work The first check follows from where hallucinations come from. If the AI makes things up when it relies on its training, give it something better to rely on, and ask it to show you what it used. - If the answer should come from your documents, your data or a particular source, point the AI there. - In your instructions, ask it to show where each answer came from. - If it's pulling quotes or research, ask it to quote the source back to you and explain how it got from the source to the answer. An answer you can trace is an answer you can check. It also helps to define your terms, because a phrase like "detect any anomaly" or "find the latest and greatest" can mean different things to an AI from one run to the next unless you write down what it means for your team. And access to sources makes a large difference. [TaxCalcBench](https://github.com/column-tax/tax-calc-bench) tests AI models on calculating realistic US tax returns. One model got 28% of the returns exactly right on its own. When it could search the web, it got 72% right. That's a big jump, and still not a return you'd file without a person checking it. Funnily enough, tax returns are mostly numbers, which brings us to the second check. ## Have it build the math For numbers, don't ask the AI to do the math. Ask it to build something that does the math. We don't trust people in the business to do complex calculations in their heads either, which is why we gave them spreadsheets. The AI's equivalent of a spreadsheet is code. It can write a short program that does the calculation, run it, and explain each step back to you in plain words, much as you'd check the formulas in a spreadsheet. And if you ask it to save the program it wrote, the same calculation runs the same way next time. Many AI tools now switch to code on their own when they see numbers, which helps, but wrong answers can still get through unchecked. So ask for it, and check that it happened. Or better still, have something else do the checking. ## Have a second agent help That's the third check, a second pair of eyes, and AI can supply it. An agent is AI that carries out a series of steps on its own, such as reading files, searching or running code, and many AI tools can now run two of them for you. One agent does the work, then hands it to a second agent, often called a subagent, whose only job is to check the steps and the result. You can ask for it in plain words, for example, "Before you show me, have a subagent check this against our latest revenue report." The second agent starts fresh, without the first one's assumptions, so it can catch what the first missed. And the more specific you are about what it checks against, the better, so point it to a named source rather than asking it to check in general. Asking the same AI to check its own work doesn't give you much confidence, because once it has done something one way, it tends to come back with the same answer. A second agent is closer to giving the task to two people and comparing what they bring back. Regardless, a person still has to look at what comes back. ## You still own what you send None of these checks moves the responsibility. Whatever you put around the AI, the person who hands the work on owns it. Producing work with AI is now fast, so the value is in the detail and care someone puts into making it what's needed. That speed cuts both ways, because it makes it easy to hand people more than they can use. So here's a rule of thumb. The time it takes someone to take in what you've made with AI should be far less than the time it took you to make it, and sending out AI output you haven't checked is the quickest way to break that rule, and tarnish some working relationships along the way. It matters even more when AI is checking AI, as in the second-agent check. When two AIs have already looked at the work, it's tempting to skim it. That's when a person needs to read it most closely, or work can go a long way in the wrong direction. ## Where to start this week You don't need all of this at once. Start with one piece of AI work your team relies on. 1. Make sure the AI is working from the right sources and shows where each answer came from. 2. If there are numbers in it, have the AI do the math in code and walk you through the steps. 3. Have a second agent check the result against a specific source before anyone sends it on. * * * Related: [Which work should your team hand to AI first?](https://onetwogrowth.com/writing/which-work-to-hand-to-ai-first) · [How to get your whole team using AI, not just the few who already do](https://onetwogrowth.com/writing/get-your-whole-team-using-ai) · [You've been handed the AI rollout. Where to start with your team](https://onetwogrowth.com/writing/handed-the-ai-rollout) · [Should your team do an AI workshop or an online course?](https://onetwogrowth.com/writing/ai-workshop-or-online-course) Back to [All writing](https://onetwogrowth.com/writing) Next [Should your team do an AI workshop or an online course?](https://onetwogrowth.com/writing/ai-workshop-or-online-course) ## Start with a call. Thirty minutes about your team and its work. If a workshop would help, Harry would be glad to put one together. [Book a call](https://cal.com/harry-onetwogrowth/discovery-conversation) · [Email Harry](mailto:harry@onetwogrowth.com) OneTwo Growth Studio is Harry Siggins’ AI enablement practice in Seattle, WA, running hands-on AI workshops for teams. [harry@onetwogrowth.com](mailto:harry@onetwogrowth.com) · [LinkedIn](https://www.linkedin.com/in/harrysiggins/) · [Harry on AI](https://harrysiggins.substack.com) · [Privacy](https://onetwogrowth.com/privacy)