Blog
I run an AI governance company and I still said yes to everything
A founder confession: this week I botched my own env variables, approved every tool prompt an agent showed me, and lost an evening reverse-engineering token thresholds in a compiled binary. I am the customer. Here's what that taught me about what developers will actually do.
July 19, 2026 · 4 min read
A confession from the founder. This week, while running a company whose entire pitch is provable AI governance, I: fumbled my own environment variables chasing a lower token bill, clicked yes on every permission prompt an AI agent showed me because reading them was slower than approving them, and spent an entire evening tracing a compiled binary to figure out why my coding tool's context meter only warns me after it's too late, then setting an undocumented env-variable override to fix it.
I am not telling you this to be charming. I'm telling you because I am the ICP, and if I behave this way, with maximum knowledge and maximum incentive to behave otherwise, then every governance product that depends on developers behaving better is already dead.
The three taxes
Every developer using AI heavily is paying three taxes nobody itemizes:
- The sommelier tax. A new cheap model lands every few weeks. Is the new DeepSeek better than Qwen for your workload? Is it better than the last DeepSeek? I genuinely do not want to know. Nobody wants to know. We want the cheapest model that can do the task, chosen correctly, per task, by something that isn't us.
- The accountant tax. Watching the meter. Estimating context windows. Digging through binaries for threshold overrides, which is a real thing I did this week, and the fact that it felt productive is the tell: managing AI spend has become its own engineering discipline, and it produces nothing.
- The yes tax. Agents ask for permissions and we grant them, because the whole point of the agent is speed and the prompt is friction. Keys go in env vars. Tools point straight at providers. Every yes is rational in the moment and none of them are written down anywhere. This is how shadow AI happens to good developers; I built a scanner to catch it and it caught me.
The lesson: safety has to arrive as a side effect
The uncomfortable conclusion from being my own worst user: developers will not do governance homework. Not because we're careless, but because every unit of attention spent on model selection, spend tracking, or permission review is stolen from the thing we're actually trying to build. Any design that asks the developer to be disciplined is a design that fails on the median Tuesday. The bill has to cut itself. The guardrails have to be on before the first request. The record has to write itself. The only acceptable cost is one decision, once.
So that's what RemKey is
This is the product I built to stop paying my own taxes: the AI layer you can prove to a regulator, priced at one decision. Swap one base URL and keep your existing OpenAI or Anthropic code. From the first request, a classifier picks the cheapest capable model per task (its accuracy, sample sizes, and misses are published on the benchmarks page), a sampled check verifies the cheap answer held up so the savings are a number rather than a hope, every prompt is screened for PII and injection fail-closed, agent tool-calls run against an allowlist, and all of it lands on a hash-chained, Ed25519-signed ledger. I never choose a model. I never watch a meter. The yes I click is governed whether I read the prompt or not. And when someone eventually asks what my AI did, the answer is an export an auditor can verify offline, produced by a system that never asked me to be better than I am.
We sell to regulated financial services first, because they're required to care. But the product was built by this developer, for this developer: cut my bill, keep me productive, keep me safe, and don't make me think about it. The free tier is 1M governed requests a month, no card, precisely because the whole thesis is that governance should cost one decision. Here's the decision: swap the base URL.