Recently, Kitces.com published an article on vibe coding for financial advisors that struck a chord with me. The premise is compelling. Using AI-powered development tools, advisors can now describe what they want in plain English and get working code without ever learning Python or JavaScript. For advisors stuck choosing between clunky spreadsheets and software that doesn't fit their niche, it's genuinely exciting.
The barriers to building custom software have collapsed. Vibe coding is real. It works.
As a team of engineers and product leaders building AI agents specifically for financial advisors, I read the piece nodding along. But here's what I've learned from building production-grade AI tools: vibe coding gets you only 80% of the way there.
The remaining 20% - polish, testing, deployment, and compliance, is where real engineering discipline lives. In our industry, that 20% is often the difference between a clever prototype and a tool you'd stake your practice on.
What Vibe Coding Gets Right
The Kitces article walks through building a PMI prepayment ROI calculator using an AI assistant. No syntax memorization. No debugging cryptic error messages. Just a conversation: "Create a calculator that shows how ROI changes across different time horizons," and the AI generates working code.
This matters for three reasons:
- It enables everyone to build. Advisors have deep domain expertise. They understand tax planning nuances, retirement projection edge cases, and the specific needs of their client niche. Vibe coding lets them translate that expertise directly into tools.
- It enables speed. Ideas that would have taken weeks can be tested in hours.
- It solves niche problems. Software only gets built if enough people will pay for it. An advisor serving early-career physicians with PMI questions was never going to find an off-the-shelf solution. Now they can build it themselves.
Vibe coding is a legitimate capability. But I've also learnt the hard way where it breaks down.
The 80/20 Reality
The first 80% comes fast. Core logic that functions. A basic interface. Something you can demo. It feels like magic. The remaining 20% is where things get hard.
Prototype to production is a BIG chasm to cross
Iteration Fatigue
AI coding assistants have limited context windows. As you iterate - fixing bugs, adding features, refining the UI; the AI loses track of earlier decisions. Changes that worked three iterations ago suddenly break. The tool becomes inconsistent with itself and sometimes even starts contradicting what it built before! For a non-technical person, guiding the AI to find and fix the issue is equally difficult.
This creates a frustrating loop: you fix one thing, something else breaks, and the AI doesn't remember why you made the original choice. What started as a conversation becomes an argument.
Error Recovery
When vibe-coded tools break, they often break badly. AI assistants struggle to diagnose their own errors and revert to working code. Instead of rolling back cleanly, they layer fix upon fix, compounding the problem.
Professional developers use version control, testing suites, and systematic debugging. Vibe coding offers none of that infrastructure. When something goes wrong, you're often starting over. I've lived this a lot.
Testing and Validation
The Kitces article explicitly acknowledges that advisors have a fiduciary duty to validate results. AI-generated code can have subtle bugs invisible without systematic testing. What happens with negative numbers? Unusual time horizons? Edge cases the AI didn't anticipate?
LLMs are predictive by design, which makes them grossly inefficient at Maths; specific tools or models are required to handle this correctly. In financial services, calculation errors aren't just embarrassing, they're potential liability.
"I built it with AI and it seems to work" isn't an answer that inspires regulatory confidence. Financial data access controls and audit trails. The ability to demonstrate to regulators who accessed what data and when.
None of this comes out of the box with vibe coding. And we've not even touched upon integrations with various systems of record (many are walled gardens), maintenance of the code and providing support if something fails.
I'm not saying that vibe coding isn't valuable. Because RIA workflows demand rigor, prototyping isn't enough. The consequences of errors are real: regulatory scrutiny, client trust erosion, potential liability.
The Best of Both Worlds - Speed and Focus
Our approach combines AI speed with software engineering discipline. Our purpose-built agents are designed specifically for RIA and broker-dealer operations. They are not generic tools adapted after the fact. This is why you should consider PitCrew as your AI technology partner.
What does that look like?
- A Cash Availability Checker to identify cash shortfall risks before you send invoice debit requests to your custodian. Analyze data to ensure client cash is optimized & liquidity needs are fully met.
- A Form ADV Auditor that catches Part 1 vs Part 2 inconsistencies before examiners do. It automatically checks services, fees, and client types across all ADV sections.
- A Social Post Auditor that reviews social posts against SEC and FINRA rules before you publish. Flags promissory language, testimonials, and missing disclosures automatically.
Production-grade reliability. Audit-ready outputs. Consistent results. Integrations to existing systems. Documentation that holds up to scrutiny.
The Road Ahead
The lines between "buy" and "build" are blurring, and that's good for advisors. More options. More flexibility. More ability to solve problems software vendors have ignored.
But with flexibility comes responsibility. Understanding when a prototype is enough, and when you need production-grade engineering, will separate firms that thrive from those that stumble.
If you're curious what purpose-built AI agents can do for your compliance workflows, hit us up. We're happy to discuss your specific use cases and build AI agents to take the grunt work off your plate.
This article was originally published on LinkedIn. Read it on LinkedIn →