The most valuable AI implementations I've built are genuinely boring. That's not a criticism — it's the whole point. Here's why proven, predictable AI beats cutting-edge every time for law firms and insurance carriers.
Founder of Grayhaven
Founder of Grayhaven
There's a phrase in software engineering circles: "choose boring technology." It comes from a 2015 essay by Dan McKinley, then an engineer at Etsy. His argument was simple. Every technology choice carries a cost. Novel technology has unknown failure modes — surprises you haven't encountered yet, edge cases nobody has documented, behaviors that only emerge under production load. You have a limited number of what he called "innovation tokens." Spend them carefully, on the problems where novelty actually earns its cost. Use proven, well-understood tools everywhere else.
"Boring" isn't an insult in this framing. It means the technology has been tested by thousands of teams across millions of deployments. You know how it fails. You know how to fix it. The docs are good. The answers are on Stack Overflow.
The opposite of boring is exciting. And exciting, in production systems, means untested, surprising, and occasionally catastrophic.
I think about this constantly when I'm talking to law firms and insurance carriers about AI.
In most industries, "exciting" technology is a calculated risk you take in exchange for competitive advantage. In legal and insurance work, the calculus is different.
Your claims processing has regulatory requirements attached to it. Your compliance reporting gets audited. The outputs of your intake workflows feed into decisions that affect policyholders, litigants, and regulators. A surprising failure in that context isn't just an engineering inconvenience. It's a liability.
You don't want exciting in your claims pipeline. You don't want novel in your compliance reporting. You want reliable, auditable, and predictable. You want to be able to explain to an auditor, or a partner, or a regulator, exactly what the system did and why.
Boring AI is the only kind that can make that promise.
The most valuable AI implementations I've built are not glamorous. They are:
That's it. Nothing revolutionary. No autonomous agents reasoning over a knowledge graph. No multi-modal inference pipelines. Just: get data out of documents, figure out what they are, send them where they need to go, and ask a human when the system isn't sure.
These four steps, implemented reliably and integrated into the tools your team already uses, will eliminate more manual work than anything a vendor with a "transformation" pitch is selling you. I've seen it happen. The boring version wins.
McKinley's "innovation tokens" concept translates directly to professional services firms.
You don't have unlimited organizational capacity for change. Every new system requires training. Every new workflow requires someone to own it. Every integration creates a dependency. Your IT team, your ops team, your managing partners — they all have limited bandwidth for absorbing new things. If you spend that bandwidth on an exotic AI architecture that requires specialized expertise to maintain, you've made a bet that will come due in six months when the vendor's onboarding consultant is gone and your team is left holding something nobody fully understands.
The better bet: spend your organizational change capacity on getting one workflow automated well. Something that works. Something your team understands. Something that integrates cleanly with what you already have. Then expand from there.
You get a limited number of innovation tokens. Don't burn them on novelty. Spend them on results.
I want to tell you about two law firms. I'm not going to name them, for obvious reasons, but I've seen versions of this story more than once.
The first firm spent $400K on an AI-powered knowledge management platform. Semantic search across the case archive. Automatic tagging and classification of documents. Recommendation engines that surfaced related precedents. It was genuinely impressive technology. The demo was outstanding.
A year later, most of the attorneys had stopped using it. The search surface enough noise alongside the signal that people didn't trust it. The auto-tagging was inconsistent in ways that were hard to predict, which meant you couldn't rely on it for anything that mattered. The recommendations were sometimes useful and sometimes bizarre, and nobody had time to figure out which was which before a deadline.
The second firm spent $15K automating intake document extraction. Incoming documents hit a shared inbox, an extraction pipeline pulled out the relevant fields, a confidence score determined whether it went straight into the case management system or landed in a review queue, and the paralegal who previously spent four to five hours per day on this work now spent forty-five minutes reviewing the flagged exceptions.
Twelve months in, the second firm is still running the extraction workflow. The first firm is back to keyword search and sticky notes.
The boring solution shipped, worked, and stayed. The impressive solution created complexity that the team couldn't sustain.
I want to give you a heuristic that has served me well. When evaluating any AI solution, ask yourself: is this boring?
If the vendor spends the first thirty minutes of a pitch explaining a proprietary approach — novel architecture, proprietary models, unique methodology that no one else has — treat that as a risk factor. It means you're betting on something unproven. You don't know the failure modes. You don't know what happens when you need support and the vendor's roadmap shifts. You're holding exotic technology with no fallback.
The boring answer to most legal and insurance AI problems is not exotic. It is:
These pieces have been assembled thousands of times. The failure modes are documented. The implementation patterns are understood. When something breaks, the fix is usually findable.
If a vendor's pitch requires you to believe that their novel approach is categorically better than proven components assembled well, ask them to show you the receipts. How long has it been in production? How many clients are using it in workflows that have real consequences? What happens when it's wrong?
Novelty is easy to demo. Production reliability is not.
There's a question that almost never comes up in AI vendor evaluations, and it's the most important one: what does this look like on a random Tuesday, six months after we go live?
Not the demo. Not the proof of concept. Not the case studies about the ideal deployment. What does day-to-day operation actually look like? Who monitors it? What happens when a batch of unusual documents comes through and the model starts making different mistakes than the ones you tested for? Who fixes it? How long does that take?
Boring technology answers these questions easily, because the answers are all standard. You monitor it the same way you monitor everything else. You debug it the same way you debug any software problem. You improve it through the same feedback loops you'd use for any data pipeline.
Exciting technology often can't answer these questions cleanly. The support model is vague. The operational runbook doesn't exist yet. The team that built it has moved on to the next project.
In legal and insurance work, "we'll figure it out" is not an acceptable operational posture.
I'm not arguing that AI should be timid or incremental. I'm arguing that the most ambitious outcome — AI that genuinely transforms how your firm or carrier operates — is most reliably reached through boring choices applied consistently, not through a single bet on cutting-edge technology.
Start with document extraction. It's boring. It works. It will free up more hours of human labor than almost anything else you could build. Once that's running and your team trusts it, you have a foundation to build on. Add classification. Add routing. Add exception handling that improves over time as you accumulate feedback.
That's the path. Not because it's cautious, but because it's the one that actually gets you there.
If you're evaluating AI vendors or trying to figure out where to start, I'm happy to talk through what the boring-but-effective version looks like for your specific workflows. The contact form on this site goes directly to me.
More insights on AI engineering and production systems
Simon Willison laid out his LLM predictions for 2026. I found myself thinking about how they land differently when your AI tools are processing claims documents and privilege logs instead of pull requests. Here's what actually matters for law firms and insurance carriers.
The technical case for AI is settled. The hard part is internal. Here's how to build the business case, address the real objections, and get your organization moving.
A quarter of businesses that have tried AI report limited or no results. If you're in that group, the problem almost certainly wasn't the technology. Here's what went wrong and how to fix it.
We help legal, insurance, and compliance teams implement AI that saves time and reduces risk. Let's talk about your needs.
