Everyone is talking about agentic AI adoption in enterprise:
"How will autonomous agents transform insurance businesses?"
"How will AI automate finance workflows?"
"Will back office operations finally become efficient?"
I think we're asking the wrong questions.
The biggest barrier to enterprise AI isn't the model, the interface or even the technology itself.
I think the real challenge is operational context.
Those unique business rules, process knowledge, system interactions, regulatory requirements and human judgement that determines how work actually gets done inside a business.
A bad business process doesn’t magically become a good process just because you wrapped AI around it.
This is the part that people aren't talking enough about.
Most enterprise environments are way more operationally complex than people realise. Legacy systems with legacy processes that have evolved over time across teams, tools, spreadsheets, emails, workarounds, and with IP that’s stored in people’s heads.
You often have the same data replicated across multiple systems, multiple stakeholders, and multiple interpretations of what the “right” process actually is.
In industries like insurance, which is the world we operate in, that complexity is amplified even further. You’ve got brokers, insurers, MGAs, banks, bordereaux, trust accounting, settlements, claims workflows, and regulatory obligations all interacting together.
But the broader point applies far beyond insurance. Most enterprises don’t have a single source of operational truth. And often, no single person fully understands how the entire process works end-to-end.
That becomes critically important when organisations start talking about deploying AI agents into operational workflows because agents are only as good as the context they’re given.
There’s a misconception emerging that organisations can simply deploy generic AI agents across the enterprise and expect transformational results. In reality, enterprise operations rely on years of embedded business rules, exceptions, judgement calls, unique cases and process knowledge that often isn’t documented anywhere.
This is the context that really matters when it comes to training AI.
I’d argue it’s businesses with strong operational knowledge and structured processes that will get the most value from AI over the next decade because they understand the operational environment they’re trying to automate.
That’s the bit I think a lot of organisations are underestimating right now. And there’s a real risk that enterprises recreate the same operational mess they’ve spent the last twenty years trying to clean up.
We’ve already lived through generations of spreadsheet macros, Access databases, disconnected workflows, and shadow processes sitting beneath critical business operations. AI has the potential to improve that. But only if organisations approach it carefully. If not, there’s a real risk they’ll create an entirely new layer of complexity on top.
Suddenly you’ve got AI agents, low-code workflows, MCP connections, and automations spread across the business with inconsistent governance, inconsistent security and unclear operational ownership. Today’s quick AI workaround could very easily become tomorrow’s tech debt. Especially in enterprise environments where financial, regulatory, and operational risk are involved.
That’s why I think the organisations getting the most value from AI are being far more pragmatic than the headlines suggest. They’re not trying to automate everything overnight. They’re focusing on specific operational bottlenecks, workflows and decision points.
They’re identifying where AI genuinely improves outcomes and where ‘a human in the loop’ is still needed. Despite all the discussion around autonomy, there are so many operational decisions where approximation isn’t good enough. Especially in financial systems.
You still need judgement and reasoning. You still need people who understand the operational context behind the output. And importantly, you still need to create the next generation of experts.
One of my favourite analogies I keep coming back to is that we still need people who know how to read a map, we can’t just blindly follow the GPS. If organisations lose operational understanding completely, they risk becoming dependent on systems they no longer fully understand themselves.
That, in my view, is the real enterprise AI challenge.
So before you deploy your next AI Agent ask yourself these four questions:
Is the process understood?
Is it documented?
Is the operational knowledge shared?
Is there governance?