Most conversations about AI risk are about what AI says. Once AI can act — update a record, send a reply, change a schedule — the question that matters is what it’s allowed to do on its own. Our answer is a short list of commitments we build into every system we make for clients. Here they are in plain language, along with how we work to keep them.
You approve changes before they happen
In what we build, the AI can look, work things out and suggest. Making a change is a separate step, and that step waits for a person’s OK. You see what the AI is proposing in plain language, then approve it or turn it down. Nothing changes until you say yes.
This matters most for actions that can’t be taken back. A sent message stays sent, so it waits for your OK before it goes out, not after. The exception is an answer you’ve approved in advance: an agent can send it on its own, changing nothing but the customer’s details, and it’s logged like everything else.
Approving doesn’t mean approving the same routine step over and over. Once you approve a recurring job — a weekly report, say — it runs on its schedule without asking each time, and it stays logged like everything else.
Everything is logged
Every change is recorded: what changed, and when. That gives you a clear trail to check any time someone asks “what happened here?” It’s also what makes undo possible, because you can’t reverse a change you can’t see.
We write down the bigger decisions too — why we chose one approach over another — so the reasoning isn’t lost later, when someone wants to change how things work.
Changes can be undone wherever possible
We build changes so they can be rolled back to how things were. That applies to the software we ship as well as to the changes our systems make: if a release goes wrong, we can go back to the version before it.
“Wherever possible” is there on purpose. Some things can’t be undone, and we’d rather say so than pretend otherwise. For those actions, your approval is the safeguard, and it comes before the action rather than after it.
Every build is independently checked before launch
Whoever builds something is badly placed to judge it. That’s true of people, and it’s even more true of AI, which sounds fluent and confident even when it’s wrong. A confident report that something works isn’t proof that it does.
So before anything goes live, a separate reviewer — not the builder — tests it against the checklist we agreed at the start. If it doesn’t pass, it isn’t done, and it doesn’t launch.
How we work
Those commitments shape every project, from the first call to long after launch:
- A free strategy call. We learn where your team’s time goes and where AI could really help — and where it shouldn’t.
- A plain-English spec. Before we build, we write down what the finished system must do as a checklist of pass/fail tests. You read it, and you sign off on it.
- A stress-test. Before any code is written, separate reviewers look for ways the plan could fail, and we fix the plan while changes are still cheap.
- The build. We build in small, reversible steps, around the tools you already use.
- An independent check. A separate reviewer tests the finished work against the checklist from the spec.
- Launch and care. You approve the launch. Then we stay on for Ongoing care — monitoring, fixes, improvements and new automations — or hand it over for you to run.
Whichever you choose, you own what we build — the code, the accounts it runs in, and your data.
We run our own business this way
Bluecurl runs its own online business on the same kind of AI systems we build for clients: a growing set of production systems that monitor, report, schedule and recommend what to do next. In our own operation, every change waits for approval, gets logged, and can be rolled back where possible.
Questions to ask any AI builder
Whoever you work with, these are fair questions to ask before anything is built:
- Who approves changes before they happen, and how?
- Where is the log, and can I read it myself?
- What can be undone, and what can’t?
- Who checks the work, apart from the person who built it?
- Who owns the code, the accounts it runs in, and the data?
If the answers are vague, that tells you something. Ours are on this page, and we’re happy to go through them on the call.
What safe by design doesn’t mean
It doesn’t mean AI is mistake-free. It isn’t. The point is that mistakes are caught before they happen where it matters (your approval), visible when they do happen (the log), fixable wherever possible (undo), and less likely to reach you in the first place (the independent check).
It also doesn’t mean you need to understand how any of it works under the hood. You approve the plan and the changes, and we explain the rest in plain language.
In short: you approve changes before they happen, everything is logged, changes can be undone wherever possible, and every build is independently checked before launch.