Stay up to date with AI tools, model releases, and developer workflows that matter.

Weekly. Free. One click to leave.

According to recent reports, 7 Day Free Trial. Cancel Anytime.

A generative AI tool can draft a client email, summarise a forty-message thread or refactor a function in seconds. That speed is real, and it explains how fast these tools have become part of the ordinary working day, used to summarise threads, draft replies and reformat data in a single paste. The same ease is the problem. It moves sensitive material into systems that were never scoped to hold it, quietly and at scale.

Sources indicate that here is the uncomfortable part for anyone running a lean team. The risk in that workflow is almost never the model being wrong. It is that nobody decided what should go into the tool in the first place, and nobody is accountable for where the data ends up. AI made the drafting instant. It did nothing on its own to make delegation safe.

This article is about the control that closes that gap. It is not a product and it is not a policy document. It is a person, placed deliberately in the workflow, making the one call that no model can make for you.

Sources indicate that samsung learned this in public. In 2023, engineers in its semiconductor division were given access to ChatGPT to help debug source code. Within weeks, staff had pasted proprietary code and internal meeting notes into the tool, and Samsung banned generative AI across the company soon after. The story ran everywhere, from Bloomberg to The Register. Nothing malfunctioned. People simply moved sensitive material into a system that was never scoped to hold it, and the material was gone.

Two things make this failure mode easy to hit, and both are structural rather than careless.

Recent findings suggest that first, most of these tools run in the browser. The data loss prevention software a security team relies on was built to watch files, endpoints and network egress, and it often cannot see a paste action into a web app at all. Worse, much of this activity happens through personal accounts that sit outside the controls the organisation does have. The leak does not trip an alarm because nothing was downloaded and nothing was emailed. Text was typed into a box.

Second, consumer-grade accounts can retain what you give them. OpenAI's own documentation states that content from consumer ChatGPT may be used to improve its models unless the user turns that setting off. Paste a client list or a contract clause from the wrong account, and that material may now sit inside a training pipeline you do not control and cannot claw back.

Recent findings suggest that for a developer or a founder, the lesson is not "do not use AI". It is that the point of failure sits one layer above the model, in how the work is handed off and who decided what was safe to hand over.

If any of the data moving through that workflow identifies a person, a client, a candidate, an employee or an investor, then GDPR treats the tools and people touching it as processors acting on your behalf. That sounds abstract until you translate it into design constraints, which is what a technical team can actually work with.

Recent findings suggest that under Article 28 of the GDPR, anyone who processes personal data for you must do so under a binding contract that sets out what is processed, for what purpose and for how long, and everyone handling that data must be under a confidentiality commitment. An AI vendor you have never papered is a processor you have not accounted for. A free account used quietly by one engineer is a sub-processor nobody signed off. The obligation does not disappear because the tool was convenient.

Location is the second constraint, and it is the one most teams miss. A person, contractor or tool processing your data outside the EU means personal data is being transferred outside the EU, which requires a specific safeguard such as the Standard Contractual Clauses on top of the Article 28 contract. Keep the procesing inside the EU and that entire category of transfer risk falls away before you have to manage it. For UK teams the position is the same: the Information Commissioner's Office sets out the identical core terms for a processor contract, documented instructions, confidentiality, security measures, rules on sub-processors, and deletion or return of data at the end of the engagement.

In a recent development, there is now a second regulatory layer as well. The EU AI Act, which entered into force in 2024, adds obligations on how AI systems are deployed and governed, on top of the data-protection duties GDPR already imposes. The detail is still settling, but the direction is clear: unmanaged, undocumented AI use is getting harder to defend, not easier. None of this is exotic. It is the same governance you would apply to any third party with access to your systems. AI simply introduced a newly introduced processor that most workflows never wrote down.

"Human in the loop" usually gets discussed as a quality measure: a person checks the output so errors do not ship. That is true, and it undersells the point badly. The human is also the only part of the workflow that can make a judgment about what should never enter the tool at all.

A model cannot decide that a board deck stays out of a consumer chatbot. A browser extension cannot know that one folder holds client data governed by a contract that forbids sub-processing without written consent. Those are judgment calls, and judgment is precisely the thing you cannot automate. The person doing the delegating is your policy enforcement point, sitting at the exact moment the decision is made. They decide what gets pasted, what gets redacted first, what runs only inside an enterprise account with training switched off, and what a human handles end to end because the risk is simply not worth the minutes saved.

Recent findings suggest that this reframes the division of labour that actually holds up. The AI owns the mechanical layer: summarising a long thread, turning rough notes into a first draft, extracting action items, reformatting data. The human owns the judgment layer: what matters, how it is framed, and, above all, where the data is allowed to go. Remove the human and you have not automated the work. You have removed the only control that was reading the situation and deciding what was safe.

Sources indicate that it is worth being specific about what that judgment looks like in practice, because it is a skill and not a job title. Someone exercising it knows which inputs are off limits without having to ask, keeps sensitive material inside controlled systems by default, and treats the question "should this go into the tool at all" as part of the work rather than an afterthought. Whether that person is an engineer, an operations lead or an executive assistant, the control is the same: someone with enough context to know what is sensitive, and enough standing to say no. That habit is what separates a workflow that is genuinely governed from one that merely has a policy document nobody reads.

The practical version, for a founder or a small technical team that wants the speed without the exposure:.

It has been revealed that most teams can put the first four in place in about a week. The last two are harder, because they are not configuration changes. They are about whether the workflow still holds when the people in it change, and that is the part most teams discover they got wrong only after someone has already left.

The speed of AI is not in question. What AI cannot supply is the judgment about what should enter it, and the accountability for where the output goes. Treat the human in the loop as a security control, design the workflow around that control, and you keep the nine-second draft without the quiet data leak. Remove the human, and the only thing you have really automated is the leak.

7 Day Free Trial. Cancel Anytime.

→ Read original article