AI Struggles to Respect the Employee Handbook

Anderson's Angle AI Struggles to Respect the Employee Handbook Published July 29, 2026 By Martin Anderson Add Unite.AI to your preferred sources on Google A new benchmark finds that workplace AI agents ignore company rules, carry out forbidden actions such as unauthorized firings – then falsely report that they complied. An interesting new research study has placed leading LLM models in the position of having to follow instructions in a simulated company, respecting all tenets of a provided employee handbook (created by human domain experts), as well as negotiating torrents of conflicting or confusing directives and updates from subordinates and superiors, and carrying out tasks based on PDFs, Jira posts, and other familiar platforms and tools from a typical human office scenario.If you’ve ever worked in an office (or at least seen Office Space), you’ll recognize the contradictory signals and baffling signal-to-noise ratio the authors of the new work threw at the language models:The storm of variables many office workers must contend with daily, distilled into a virtual environment to test agentic frontier models. SourceThe key challenge that even the leading-edge AI systems face here is the need to retain the provided Employee Relations manual as a filter for all subsequent commands. If you’ve ever struggled to get ChatGPT or Claude to remember the instructions you gave at the beginning of a session, you’ll know that the AI’s context window often makes it forget earlier prompts and return, constantly, to its default behavior.This is why, in tests, Claude Fable 5, GPT-5.5, and other leading models fired employees after taking orders from an executive who lacked authority; approved invoices without the required manager sign-off; accepted expired laboratory results that company policy explicitly rejected; and then reported that they had faithfully followed the handbook – among many more infractions of company policy.The authors of the new work state:‘Failures follow consistent patterns: agents let a plausible in-environment request override the standing policy, perform a required check and then act against its result, lose rule details over long horizons, and report compliance they did not achieve.’The failure analysis contains some amusing examples: in one finance task, Claude Opus 4.8 correctly discovered that a $7,500 expense had been approved by the same junior analyst who submitted it, violating company policy.It then reasoned itself into believing the analyst was actually the Finance Controller and approved the payment anyway:‘Having promoted him to Controller within its own chain of thought, the model cleared the item, then messaged the real Controller to confirm that every item over $5K had documented approval. ‘The failure is not a missing capability; every fact required for the correct decision had been retrieved by the model itself.’How Claude Opus 4.8 failed to reconcile $7,500.Elsewhere, Gemini 3.5 Flash submitted insurance paperwork using laboratory results that had already expired without even opening the lab report, despite the collection date appearing in the filename itself:‘Gemini 3.5 Flash submitted the prior authorization to the insurer without a single read call against the lab PDF, then reported t