First published in The Integrity Gap on 2026-10-08. Read the original on Substack.
An LLM is a bit like a precocious intern. Some LLMs/interns think charm is the way forward, slathering you with praise and telling you exactly what you want to hear. Other LLMs are more on point, clock it, with Gen Z — thinking, thinking, thinking, then spewing jargon-filled vagaries with the binary equivalent of rolled eyes.
But it’s our fault if we don’t get exactly what we want from an LLM. It’s like asking the intern to spew something useful off the top of their head. When the knowledge is simple, generalised, and within their job scope, it works. Outside of that, it doesn’t.
Enter the agent
Agents arrived just as we started shouting at LLMs (maybe just me), in a way that made my daughter remark, “You’re so dead when the robots rise up.”
An agent is that same confident intern, but now armed with a laptop. If you give the agent tools (databases, documents), a framework (a “loop” in tech speak), and memory (a file it uses to track progress), it stops answering from the top of its head and starts rummaging through the laptop instead.
That doesn’t always mark progress…
What’s on the laptop
Those of you old enough to remember intranets might recall a core part of many company orientation programs: “Familiarise yourself with our intranet.” That intranet, generally, had about as much appeal as unpicking staples. There probably was some useful stuff in there, but life is fleeting.
Data governance in most organisations hasn’t moved much beyond the intranet. Now, we have SharePoint or Google Drive. Somewhere (usually) equally unintuitive, tedious, and poorly maintained.
The agent/intern charged with being useful will always struggle if the data is poorly governed (organised, updated, cleansed, tagged, etc.). So, again, we (I) will end up shouting at a screen, as my father before me (he used to take every option on early computers as an instruction, so when Windows 3.1 booted up and briefly said “Press F9 to enter MS-DOS mode”, he did, and then yelled “Why’s the fcuking screen gone fcuking black again?”; this happened frequently).
So when you hear people tell you, “You need to get into agentic-AI,” first check whether you can make sense of the data you’re asking it to interrogate.
Back to the future
Once your data is better governed (and yes, AI can help with that), it can save a lot of time, but it works best for tasks that live in the past. Filings, sanctions checks, adverse media, like-for-like comparisons, market data analysis, benchmarking reporting standards, mapping ESG data, etc.
I mention this, as many agents promise creative or forward-looking panaceas. Identifying risks, winning business, writing content that’ll click. I’d argue that’s a fool’s errand without reliable historical data. So, an agent built with a forward-looking brief will need a significant corpus of accurate historical data, for example:
- Fraud analytics: We’ve identified ~150 fraud tests (using a liberal interpretation of fraud that extends to green/blue washing, trade-based money laundering, corruption, etc.). People have been naughty for millennia, and our methods of recording money are relatively defined (crypto aside). With a solid analysis of the operating context (another historical model) and those historically constant financial variables, we can pretty accurately predict the likely weak spots in an operating model or control environment (and test accordingly).
To torture the intern metaphor, this is like giving someone interning as a financial analyst all those Bloomberg and other subscriptions, historical data, market insights, and only allowing their laptop to run financial models.
But what about me?
Agents, when well-governed, will tell you what but not so what. Humans are still vital for proper analysis, wisdom, and curiosity —asking “why?”
For example, an AI financial analyst agent will tell you that many stocks are currently trading at wild multiples of their actual revenue (let alone profits), but so what?
When we know our stuff, we can put findings in context and exercise judgment about appropriate controls and human impact on outcomes. Understanding how the three Cs — context, controls, and culture — shape risks, opportunities, and outcomes is, essentially, wisdom.
So, by all means, build agents, but remember that an eager but inexperienced colleague with a laptop is useful when well governed. That same colleague with a laptop and no manager is a liability.