How to Connect Your ATS to an AI Without Leaking Candidate Data

How to Connect Your ATS to an AI Without Leaking Candidate Data

Javier Caldera is Head of Product B2B at Nova, where he decides what the product is allowed to touch. The step people skip takes five minutes, and skipping it is how a convenience becomes a reportable breach.

Updated 6 min read

Connecting your ATS to an AI agent is the step that turns a clever assistant into something genuinely useful. It’s also the step where recruiters hand over a key to every candidate record they hold, usually without noticing that’s what happened.

This is the step that turns the setup in what an AI recruiter actually does from a clever assistant into one that knows who you have already spoken to.

The connection itself is easy. Most ATS platforms expose an API, and an agent can read it with the key and a documentation link. Some ship a direct connector now. None of that is the hard part.

The hard part takes five minutes and almost everyone skips it.

Your API key is not a login

When you generate an API key in your ATS, you’re not creating a password for yourself. You’re creating a credential that reads the system.

What sits behind it: candidate names and contact details, CVs, interview notes and scorecards, stage history, offers, salary expectations. The notes your hiring managers wrote assuming nobody outside the room would read them.

And API keys usually don’t inherit the permission limits your user account has. The key often reads more than you can see in the interface.

Do this: before generating anything, check whether your ATS offers scoped or read-only keys. Many do, almost nobody uses them, and a read-only key removes an entire category of accident.

The risk isn’t the AI. It’s where the key sleeps

Here’s the part worth sitting with. The agent reading your ATS is not the threat. The threat is a plain text file called something like secrets.txt sitting in a folder on your laptop.

Think about everywhere that file quietly goes. Into every backup, automatic ones included. Into the cloud folder that syncs to your phone. Into the zip you sent a colleague because it was easier than explaining the setup. Onto the laptop you leave in a bag at a conference.

Each of those is a copy of the key, and the key is a copy of your candidate database.

This isn’t theoretical and it isn’t an IT problem you can hand off. The key was generated by you, under your name, for your convenience.

Do this: put the key in a password manager, and use one with a command-line companion so your agent can request it when a task needs it rather than reading it from a file. Your secrets.txt ends up holding nothing sensitive at all. Five minutes, once.

What a leak actually costs you

If candidate records get out, this stops being an internal embarrassment.

Under GDPR, where a breach is likely to risk people’s rights and freedoms, the controller has to notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it. Where the risk is high, the affected people have to be told too.

Read that timing carefully, because it’s the part that catches teams out: the clock starts when you become aware, not when you finish working out what happened. You’ll be assembling the facts and notifying at the same time.

And the people in that database applied to you in confidence, often while employed somewhere else. A leak of interview notes and salary expectations is not the same category of problem as a leaked marketing list.

Do this: write down today who your controller contact is and who gets called first. Deciding that during an incident is how 72 hours turns into five.

Prove the connection works, don’t assume it

Once it’s connected, test it with a question that can’t be answered by guessing.

The weak test is asking it to list your candidates. An agent can produce a plausible-looking list from a single collection, or from nothing at all, and you’d never know.

The strong test is asking for the full profile of someone currently sitting at phone screen. That forces the agent to navigate the relationship between a candidate and a stage, which is where a half-working connection falls over. Then open your ATS and check the answer against the record.

A confident wrong answer is worse than an error message, because an error you’ll fix and a wrong answer you’ll act on.

Do this: run your test against a candidate whose record you know by heart. You’re checking the plumbing, not the candidate.

Where the law puts the responsibility

Two rules matter here, and the second surprises people.

The first is the one above: a candidate data leak is a notifiable breach.

The second is that under the EU AI Act, candidate evaluation systems are classed as high risk, and the obligations sit with the employer deploying the system, not with the vendor who built it. Buying a tool doesn’t transfer responsibility; it gives you a supplier.

The timing is worth getting right, because most coverage still has it wrong. Those obligations were originally due in August 2026, and the Digital Omnibus moved them to 2 December 2027. So there’s more runway than the headlines suggest, and what the runway is for is building the evidence trail, since human oversight and record keeping are not things you can produce retroactively. Penalties, once the obligations bite, reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher for a large company.

The practical line is cleaner than the regulation makes it sound. An agent that drafts, summarizes, prepares and organizes is support. An agent that decides who gets rejected is a system you have to be able to account for, including showing where a person with the authority to change the decision actually looked.

Do this: automate the yes, never the no. Moving someone to a conversation is reversible and low stakes. A rejection is neither, and it’s the one you’d have to defend.

The order that makes this safe

Nothing here is hard. It’s five steps, and only one of them is about the AI.

  1. Generate a scoped or read-only key if your ATS offers one.
  2. Put it in a password manager, not in a file.
  3. Point your agent at the password manager, so the key is requested rather than stored.
  4. Test with a lookup that forces a real relationship, and check the answer against the record.
  5. Decide, in writing, what the agent drafts and what a person decides.

Do those once and connecting an ATS stops being the risky part of the setup. Skip the second one and everything else you build sits on top of a copy of your candidate database in a folder.

If you want the whole workflow walked through with the prompts, that’s what Become an AI Recruiter covers across sixteen lessons.

For in-house teams and agencies running their own hiring data, the key handling above is worth doing this week, whether or not you ever connect an agent. If your ATS is administered by someone else and you’ve never generated a key, the person who holds them is who this conversation belongs to.

Frequently asked questions

Is it safe to connect my ATS to an AI assistant?

It can be, and the risk is almost never the model. It's the key. An ATS API key usually carries full read access to candidate records, so the question is where that key lives and who can reach it. Stored in a password manager and pulled in only when a task needs it, this is a normal integration. Pasted into a plain text file on your desktop, it's a copy of your candidate database sitting in every backup you own.

What does an ATS API key actually give access to?

Typically everything in the system: candidate names and contact details, CVs, interview notes and scorecards, stage history, offers and salary expectations. API keys generally don't carry the permission limits a user account has, so the key often reads more than the person who created it can see in the interface. Check your ATS for scoped or read-only keys before you generate one.

Do I need to know how to code to connect an ATS?

No. Most ATS platforms expose an API that an agent can read with the key and the documentation link, and a growing number ship a direct connector. What does take a deliberate decision is where the key is stored, and that part is a password manager with a command-line companion, not a coding exercise.

What happens if candidate data leaks from my laptop?

Under GDPR, if the breach is likely to risk people's rights and freedoms, the controller has to notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it. The clock starts at awareness, not when the investigation finishes. Where the risk is high, the affected candidates have to be told as well. In practice this is a security incident involving the personal data of people who applied to you in confidence.

Does the EU AI Act apply to a recruiter using AI on their own laptop?

The obligations for AI used in candidate evaluation sit with the employer deploying the system, not only with the vendor who built it, and they include human oversight of decisions and record keeping. On timing: they were due in August 2026 and the Digital Omnibus pushed them to 2 December 2027, so a lot of what is published about this deadline is now out of date. Penalties, once they apply, reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher for a large company. The practical reading has not changed: an agent that drafts and summarizes is support, and an agent that decides who gets rejected is a system you have to be able to account for.

Does Nova Recruiter connect to my ATS?

Yes, with 60+ ATS platforms, plus Gmail and Outlook. Nova Recruiter also works directly with Claude, ChatGPT and Cursor on paid plans, with human approval required before anything gets created. Worth checking on any vendor, not just this one: whether the sync runs both ways, and whether the integration is on your plan or only on the top tier.

Sources and method

GDPR breach notification timing is Article 33 of Regulation (EU) 2016/679. The penalty tiers and the allocation of deployer obligations are Articles 26 and 99 of the EU AI Act, https://artificialintelligenceact.eu/article/99/. The December 2027 application date for Annex III high-risk systems, which cover recruitment, comes from the Digital Omnibus amendment, Regulation (EU) 2026/1744, in force since 27 July 2026. Nova Recruiter figures (60+ ATS integrations, MCP on paid tiers with human approval) are Nova's own, as published on the product page.

Javier Caldera
Javier Caldera

Head of Product B2B

Javier Caldera is Head of Product B2B at Nova, where he connects the product with what clients actually need. Javi joined Nova in 2021 and worked across Talent, Sales and Operations before moving into Product, building Nova's Spanish sales team along the way. Before Nova, he was a consultant at Management Solutions.

Connect on LinkedIn →
AI in recruiting AI Recruiter: What the Job Looks Like When You Actually Use One AI in recruiting Recruitment Automation: From 30 Hours to 30 Minutes