Post by Alistair S.

HOW TO MAKE AI DO FORENSIC ACCOUNTING PROPERLY. Asking generic questions gets you generic answers. For example, if you ask AI: “Do forensic accounting on this company", you’ll usually get a generic checklist back. Receivables, Inventory, Working capital, Adjusted earnings, Exceptional items. Debt, Cash conversion etc etc. These are all useful things to check, but on their own, they’re not enough. The better way to do this is to properly explain how you’d actually think about the company as an analyst (which is harder than it sounds). If you were doing the work yourself, you wouldn’t look at every company in the same way. A software company, a retailer, a housebuilder, an insurer and an industrial business all have different things that can go wrong in the accounts. They make money differently, they recognise revenue differently, cash comes in at different times, working capital behaves differently and management has different places where judgement can creep in. So before you ask AI to look for red flags, you need to make it understand the business. What does the company sell? When does revenue turn into cash? Where does working capital normally build? Which costs are fixed? Where can management judgement affect the numbers? What would normal cash conversion look like for this type of company? Only then should it start asking whether the accounts make sense. And even then, the job isn’t just to list red flags. A good analyst spots something odd, then asks whether it’s actually worth exploring. Is this a real warning sign, a normal feature of the business, a timing issue, a disclosure quirk, or just a rabbit hole that’ll waste half a day? This is super important because you can't harness to productivity gains of AI if you still have to go down every rabbit hole and ironically, if you can now do forensic accounting on all 50 companies you look at, that's a lot more holes to go down than when you didn't have that capability. I’ve built a forensic accounting prompt / skill for investment research that starts with the business model, then works through a broad range of flags and it also forces the model to decide which flags deserve deeper work and which probably don’t. One thing that's really important to remember when writing skills for LLMs generally, DON'T MAKE THEM TOO RIGID. You want to harness the creativity of the LLM so there's always a balance to strike of giving it a rigid checklist vs broad guidance. It works in ChatGPT or Claude if you upload the right documents, although like all generic LLM workflows, the output is only as good as the information you give it (and the quality of the agent of course). You can try now for free in Primer (www.primerapp.com). Make sure we're connected and comment FORENSIC and I’ll send it over.