Daily accounting in most situations comes down to invoice processing and payslip work, and we wanted to know whether autonomous agents could handle it. The system we built has been running internally since 2025, processing invoices from email identification through to ledger assignment. The agents verify validity, confirm we actually need to pay the supplier, extract the data, upload to the accounting system, and assign line items to the correct accounts, following the accounting rules of whichever country applies along with internal policies. They're not guessing, and extreme accuracy is possible because the rules are followed rather than approximated.

Auditing

Financial auditing requires expert judgment and is full of edge cases that feel more like investigation than procedure. But auditing also follows exhaustive, clearly written rules (IFRS, ISA, EU audit regulations). Building a full auditor agent was tempting, but we decided to start with foundations. The systems we've built handle autonomous data processing, visualization, outlier detection, and situation explanation, all derived from the standards and the audit firm's best practices. The infrastructure for auditing has to exist before the auditor agent can.

xOC (Operations Centers)

The benefit of multi-industry project and research experience is the ability to recognize patterns across seemingly very different industries and tasks. One pattern we identified is that most operation centers follow the same structure regardless of whether it is IT security, network operations at a telecom, or cloud operations at a hosting company. Since these patterns and procedures are very similar, we found it attractive to work on this area, hoping that if we can crack it in one industry, it becomes implementable in other industries at much lower cost and higher efficiency.

Security Operation Centers

We started with security operation centers and teamed up with a Germany-based managed security service provider to test our hypothesis in practice. The research ended up as an MVP in late 2024. We built a synthetic SOC analyst and a threat hunting system, connected to a SIEM system with full read rights. The SOC analyst acts as support for human analysts, while the threat hunter is proactive, doing hunts on its own. For context, threat hunting is one of the most sophisticated jobs at a SOC, requiring the most advanced skills and most experience.

The basic problem with operation centers is that you have people watching screens for things to go wrong, but modern systems generate so many alerts that nobody can actually look at all of them properly, which means analysts either cut corners or they get tired and start missing things, and on top of that finding good people for this work is expensive and difficult because it's stressful shift work that burns people out.

What agents change is they take all the repetitive investigative work off the analysts' plates, so the agent checks every alert and investigates it properly and only involves a human when something actually requires human judgment, which means the analysts can spend their time on the genuinely interesting problems instead of drowning in routine noise that exhausts them before they even get to the important stuff.

And once you're not constantly overwhelmed just keeping up with the queue, proactive work becomes possible for the first time, because most security operation centers barely have time for threat hunting since everyone is too busy fighting fires all day, but with agents handling the routine you can actually go looking for problems instead of sitting there waiting for alerts to tell you something is wrong.

Legal

Legal and compliance appears in the Gen AI use cases list, with Legal content analytics highlighted as a specific use case.

Compliance

Compliance appears in the Gen AI use cases list as part of Legal and compliance.