What to do when a client asks you to ship something unsafe
6 minute read. Updated 2026-08-08.
When a client asks you to ship something unsafe, put your concern in writing, name the specific risk, and offer a fix before you refuse anything. Escalate once through the client's own structure, with a deadline for a response. If they proceed without addressing it, invoice for work completed and walk. Never do free remediation to avoid conflict, and never let a verbal assurance replace a written record.
Name the risk precisely
Unsafe rarely arrives as one dramatic instruction. It shows up as a deadline that quietly swallows the evaluation step, a request to remove the disclaimer nobody reads anyway, or a decision to skip the red-teaming pass because the demo is booked for Thursday. Learn to recognise it at this stage. By the time it reaches you as a direct instruction, someone above you has usually already decided the risk is acceptable, and your job becomes harder from there.
The pattern repeats across the work. A chatbot given access to account data with nobody having reviewed what it might say back. A scoring model deployed without any check for uneven outcomes across groups of people. A system with no audit trail, so nobody can reconstruct six months later why it made a particular decision. None of these are exotic edge cases. This is the ordinary way AI work goes wrong, and it goes wrong quietly.
Put it in writing before you do anything else
The single most useful thing you can do is also the least dramatic. Write down what you are being asked to do, what the specific risk is, and what would need to change before you're comfortable proceeding. Send it to the person who gave the instruction, not to a chat channel that scrolls out of view within a day. Date it.
This is not about covering yourself for its own sake. It protects you if things go wrong later, and it protects the client, because it forces a decision that has real consequences to be made deliberately, by someone with the authority to make it, rather than drifting into place because nobody said no out loud.
Escalate once, properly
If the first response is unsatisfactory or there is no response, escalate once to the person's manager or to the commercial sponsor of the contract. Do it formally, with the same specificity as the original message: what the risk is, what you're asking for, and a reasonable deadline for a decision.
Do not repeat this through the same channel five times. Repeated escalation to the same person stops looking like documentation and starts looking like obstruction. One written escalation, one clear ask, one deadline. After that, you have the information you need to decide what happens next.
Decide what you can keep working on while you wait
There is a difference between work that can continue while you wait for a response and work that has to stop immediately. Building the missing evaluation, drafting the safer version, preparing the fix: none of that needs to pause. What has to stop is the specific unsafe step itself, flipping the switch that puts an unreviewed model in front of real customers, or removing the safeguard you were told to remove.
Holding this line while continuing everything else shows you're trying to solve the problem, not stall the contract. If the client addresses your concern, you haven't lost time. If they don't, you've done everything short of the one thing you were right to refuse, and that distinction matters when you explain your decision later.
When to walk, and how
If the client proceeds after reading your written concern without addressing it, you have your answer. Finish and invoice for what you've delivered, stop further work under the arrangement, and do not offer free remediation to smooth things over. A clean exit, with the paper trail already in place, looks nothing like storming out.
You need the money, and that is exactly why the paperwork matters. Walking away from a contract with no written record looks like a personality clash to anyone who hears about it afterwards. Walking away with a dated, specific, professionally worded concern on file looks like judgement, and that is what the next client hears about, whether directly or through people who already know you.
Where the client's risk stops being your risk
The client typically carries the formal compliance exposure, under UK GDPR for personal data and, on its own staged timetable, the EU AI Act. That does not touch your personal or professional exposure. Your name is on the delivery, and the people your next client asks about you will remember what you were like under pressure, not just what you built.
None of that can be delegated to the client's compliance function. You cannot outsource your own judgement, and a client's internal sign-off doesn't remove your name from the record of what was delivered. Keep your own copy of everything you wrote and everything you were told. It's the only record guaranteed to still exist when you need it two years from now.
What to do about it
- Put every safety concern in writing, dated, sent to the person who can act on it.
- Escalate once, formally, with a deadline for a response, not repeatedly through the same channel.
- Keep working on the parts you control, and refuse only the specific unsafe step.
- Invoice for work completed before you walk; never do free remediation to avoid conflict.
- Don't accept a verbal assurance that something will be fixed later, get it in writing.
- A documented, professional exit is a stronger reference than staying in a contract you can't defend.
Questions people also ask
What counts as unsafe in AI work if nothing has technically gone wrong yet?
Unsafe usually means a safeguard has been removed, not that harm has already happened: skipping the evaluation step, deploying without a human review point for high-stakes decisions, or removing the logging that would let anyone reconstruct a bad outcome later. If you can't explain, in one sentence, how the client would catch a serious failure before a customer does, treat that as your signal to raise it.
What if this is my only contract and I genuinely cannot afford to walk?
Then don't walk first, escalate first. A written concern with a proposed fix costs the client nothing to read and costs you nothing to send, and most clients would rather fix the problem than lose the person who flagged it. Reserve walking away for the point where the client has read your written concern and proceeded regardless. That is a different, much rarer, situation from a first-pass disagreement.
Isn't a written email overkill for something that might just be a misunderstanding?
No. A short, factual email costs nothing to write and resolves genuine misunderstandings faster than a verbal exchange, because it forces a specific written response rather than a vague reassurance. Save it either way. You may never need it again, but you cannot create it retroactively once something has already shipped and gone wrong.
Can a client refuse to pay me for raising a concern like this?
They can try, but a dated written record of what you delivered and what you flagged, sent before any dispute arose, is the strongest position you can be in if payment is contested. Invoice promptly for work completed, and keep the written concern and the invoice as separate, clearly timestamped items rather than bundled into one conversation.
Where the figures come from
Every rate and salary quoted in this article is a median or percentile of figures advertised in UK job postings over the six months to 8 August 2026. They are not rates paid, and the gap widens at the top of a range.
- IT Jobs Watch, UK contract rates, 6 months to 8 August 2026, read 2026-08-08.
- IT Jobs Watch, UK permanent salaries, 6 months to 8 August 2026, read 2026-08-08.