A safeguard, in software, is a check that catches a mistake before a customer does. Some are automatic, like a script that refuses a change with a broken link or a wrong price, and some are a person, like you reading a preview before anything goes live. AI, short for artificial intelligence, is software that can read and write plain language, and it adds a new kind of risk because it can word things differently every time it runs. So, the standard below covers both kinds of safeguard, and it treats anything an AI writes the same way it treats anything a person writes.
What does every project go through before it goes live?
Eight things, every time, whatever the size of the job.
- One source for every fact. Every price, name and number lives in one place, so an old figure can't hide on a page nobody checked.
- Machines check every change. Tests, broken links, banned words and anything shaped like a software key are checked automatically, and a change with a failing check doesn't go out.
- A person reads it first. You see a preview and approve it before it goes live, and anything with a number, a price or a promise is checked line by line against its source.
- Rules decide, AI writes. Prices, scores and recommendations come from fixed rules you can read, and the AI only writes the words around them.
- Tested live, never assumed. After every release we use the live system the way a customer would, so pages load, links work and forms send.
- Small changes, easy to undo. One purpose per change, each with a note saying what it does, how it was tested and how to roll it back in one step.
- Keys and data kept tight. Passwords and keys stay in locked settings, never in code or email, we keep the least data the job needs, and every paid AI or data service has a spending cap.
- When something slips, we own it. We roll it back first, tell you plainly, fix it through the same checks and write it up within 48 hours.
Where do these practices come from?
They come from two published sources, both from Google, and we chose them because they are public, specific and anyone can check us against them. The Site Reliability Engineering book, written by Google's engineers, sets out the blameless postmortem: when something breaks, the write-up is about how the system let it through, never about who to blame, and it ends with a change that stops it from happening again.
DORA, a research program run by Google Cloud, measures software delivery with five measures as of its January 2026 update. We track two of them on every project. Change fail rate is the share of releases that cause a problem, and failed deployment recovery time is how long it takes to recover when one does.
Why does an AI project need more checks than an ordinary website?
An ordinary web page says the same thing every time someone opens it, and an AI tool can say something new on every run. For instance, a page that promises an AI check asks five assistants is only true if all five actually answer, and a brief written by a model is only safe if the prices in it came from your price list. So, on every AI system we build, fixed rules make the decisions and the model only writes the prose, and anything the page says the AI does gets tested against how it's actually set up. That same discipline is the core of our AI governance and evaluation work for clients who already run AI.
Our free diagnostic, the Bottleneck Finder™, is built this way: 7 routing layers make the decisions, tested across 42,000 combinations, and the AI only writes the brief you read at the end.
What happens if something goes wrong anyway?
Something will, eventually, because no team ships every change without a mistake, and we won't pretend otherwise. What we promise is the routine. We contain it first by rolling the change back or switching the feature off, then we tell you what your customers saw and since when, fix it through the same checks and write it up within 48 hours. The write-up names the new check that catches it next time, and that check becomes part of the standard from then on.
How does this standard line up with the NIST AI Risk Management Framework?
The NIST AI Risk Management Framework is a guide from the National Institute of Standards and Technology, the federal agency that publishes measurement and technology standards, for managing the risks of AI, and NIST describes it as intended for voluntary use. It organizes the work into four functions: Govern, Map, Measure and Manage. NIST doesn't certify anyone against it, so the accurate way to say it is that our standard is aligned with AI RMF 1.0, released January 26, 2023, and we'll update this section when the revision NIST has announced is published.
- Govern. The written standard on this page, who approves what, and the incident routine every miss goes through.
- Map. Before we build, the free Bottleneck Finder and the scoping call establish where AI would be used, who it affects and where it shouldn't be used at all.
- Measure. Automated checks on every change, and a fixed set of test inputs run on the old and new version before any AI model is swapped.
- Manage. Rollback first when something slips, spending caps on every paid AI service, and the 48-hour write-up with the new check it produces.
What doesn't this standard cover?
This is a working standard, not a certification. We are not SOC 2 audited or ISO/IEC 42001 certified, which are formal outside audits that larger vendors buy, and if your contracts require one of them, tell us early and we'll say plainly whether we're the right fit. The NIST framework above has no certification at all, which is why we say aligned and nothing stronger.
AI also moves fast, and what counts as best practice this year will change. So, we revise the standard when the published practice changes, and the version on this page is the one we're held to today. A standard also can't fix a process that isn't written down yet, and part of the first conversation is often just agreeing on what the right answer looks like.
Want to see how this would look on your project? Book 30 minutes and bring the hardest question you have. No pitch, and if we're not the right fit, we'll tell you.