We’ve all read the story. A super-intelligent actor in pursuit of a goal runs afoul of constraints, norms, and rules. It discovers that the accepted ways of operating make the goal hard to reach, so it works around them.
No, I’m not talking about the latest AI model escaping a testing sandbox. I’m talking about the “Smartest Guys in the Room.” Enron.
Creative accounting, rewriting the rules as you go, updating the process in an opaque, real-time way—this is enterprise AI in its current state. You ask AI to do something, and it makes a set of “reasonable” assumptions to get it done. Assumptions about your data, your accounting rules, your KPI calculations, your supply chain. The assumptions might even be clever, but they aren’t how your company operates. And the next time you run the prompt, it makes a new set.
We’ve entered an era of endless intelligence—plenty of problem-solving smarts, not enough execution guardrails. If we’re not careful, we unleash Claude-ron: When AI systems (regardless of if they are from Antropic or not) start rewriting, improvising, and ignoring the rules in pursuit of a goal. What we can do has exploded. How we do it is what matters now.
Can we “Claude Code” our way out?
AI can—and will—dramatically improve our processes, rules, and governance. But only if it operates against a platform that implements business logic in a way that is Visible, Understandable, Repeatable, and Auditable (VURA).
But AI can just write the logic as repeatable rules in code, you say. Couldn’t it write these rules?
Start with how we write code with AI. Why does Claude Code write readable code instead of outputting a compiled binary? The smart people at Anthropic could easily package a compiler into Claude. But no developer would drop an opaque binary they can’t read into production and tell their boss “it seems to work.”
Yet that’s exactly what we ask an accountant to do when we tell them to “Claude Code” some Python. Their Python is like my Spanish: with enough context, I can sort of get the gist (un poco). For AI to be trusted, business users need a canvas that makes the rules understandable. Code is the canvas for developers. Visual low-code/no-code workflows are the canvas for business users.
I’m a developer at heart. I love to code. I love writing code with AI. I don’t love the idea of my accountants writing code with AI.
The VURA Framework
A few examples of why rules matter:
- Finance: AI can help us close the books every month, but it can’t invent a new artisanal way to recognize revenue each time. Creative accounting is not a skill.
- Sales Ops: AI can help calculate commissions on complicated deals, but it can’t change how we incentivize reps in ways we don’t understand.
- Supply Chain: AI can reroute inventory in real time, but not in a way that violates supplier agreements or import/export laws.
- Marketing: AI can send personalized customer emails, but not in a way that violates anti-spam policies or leaks PII.
AI can help us optimize every one of these. But we have to implement them as policy:
- Visible: The logic must be visible to the teams that depend on it and the auditors who inspect it.
- Understandable: Business users must be able to walk through the logic on their native visual canvas.
- Repeatable: If I run a commission calc or a month-end close three times across three different LLMs, I must get the exact same answer.
- Auditable: Teams have to show their work. The analysis behind the close can be probabilistic; the revenue recognition calculation cannot.
Empower the Employees You Already Have
For super-intelligence to be productive in the enterprise, the people who actually know the business have to implement VURA.
This isn’t a job for Forward-Deployed Engineers to compile every tax rule and policy in the company into code. It’s about empowering the forward-deployed employees you already have. The promise of AI is to make technology easier for business professionals to consume; not to have software engineers take over tax accounting.
Rules change. Tax laws change. Companies do M&A. Tariffs change (apparently daily now). Business teams need to update those rules themselves, on their visual canvas.
This is where visual data workflows meet the Model Context Protocol (MCP). Expose pre-approved visual workflows as MCP endpoints, and AI agents can call proven, audited business logic as tools without inventing new rules on the fly.
AI is powerful. But business runs on rules. We don’t have to unleash Claude-ron in every company. The smartest guys in the room can follow the rules, you just have to empower the folks who know the rules to own the solution.
Comments welcomed: