How to Integrate AI Into Your ATS

Why your ATS isn’t enough (and what your two options are)

Your ATS is the beating heart of your recruitment operation. All candidate data, vacancies, pipelines and communication history lives there. Without it, you can’t work.

But honestly: most of them aren’t smart. They store data. They structure workflows. They do exactly what you tell them. Nothing more. They don’t write summaries. They don’t fill in fields after a conversation. They don’t format CVs. They don’t tell you which of your eight hundred candidates fits the vacancy that just came in, or why.

That’s where AI comes in, and there are exactly two honest ways to get it there.

Route one: a layer on top. You keep your ATS and add a product that records conversations, summarises them and pushes structured data across a connector. This is the cheaper, faster move, and for a lot of teams it is the right one — particularly if your ATS genuinely fits your business and the only thing you want gone is the write-up after every call.

Route two: an ATS where the AI is the foundation. The records, the recording, the extraction, the matching and the reporting are one system. Nothing crosses a connector because there is no gap to cross.

The reason to state both is that route one has a running cost most vendors leave out of the pitch. Every field the AI wants to fill has to be mapped to a field in a system it doesn’t own, in a format that system expects, under permissions it can’t see. That mapping is a project, and then it is maintenance, and it breaks quietly when either side changes. The question is not which route is better in the abstract. It is where your friction actually is.

The three levels of AI-ATS integration

Level 1: Data export (minimal)

The simplest form: AI processes a conversation, generates a summary, and you export that summary to your ATS. Copy, paste, or a CSV upload.

Better than nothing. But it doesn’t solve the core problem. You save time writing the summary, and you’re still moving data by hand. You’re still switching between tools. The admin has shifted, not disappeared.

Level 2: Field-to-field integration

Here the AI talks to your ATS over a live connection. The summary lands as a note on the right candidate. Data fields get filled — availability in the availability field, salary in the salary field, in the right format. Dropdowns get selected. The CV gets formatted and attached.

This is the level most “AI for your ATS” products are selling, and when it works it works well. What it costs you is the mapping: somebody has to teach the AI that Bullhorn calls it Current Title, that your date fields want a particular format, that this dropdown has eleven options and only four are in use. That knowledge lives in a configuration file between two companies, and it needs an owner.

Level 3: No integration, because there is nothing to integrate

The third level isn’t a deeper connector. It’s the absence of one. When the AI and the record live in the same system, the extracted value goes into the field because the field is right there. There is no mapping layer, no format translation, no sync delay and no reconciliation job.

This is what an AI-native ATS means in practice, and it is worth being precise about the trade. You give up the ability to keep your current ATS. In exchange you stop paying the integration tax on every field, forever. Contextual recruitment — the system knowing that the person you just spoke to fits the vacancy that has been open three weeks — is only really available at this level, because it needs the conversation and the vacancy to be in the same place.

Which ATS systems work with AI?

By 2026, nearly all of them claim to. The useful distinction isn’t whether a vendor has AI features; it’s whether AI was the foundation or an addition, because that determines how deep it can reach.

Established systems with AI added

Bullhorn is the global standard in staffing: deep, heavily configurable, with a large partner ecosystem. Mysolution and Byner are both built on Salesforce and long established in the Dutch flex market, with mature contracting and mid-office flows. Carerix is strong in Dutch staffing with mature back-office; its AI arrived as an addition rather than as the foundation. OTYS has been around a long time with broad functionality and a large partner network. Recruitee is well designed for in-house hiring teams and career sites.

All of them have AI features now — tagging, matching suggestions, parsing improvements, mail drafting. The pattern is consistent: AI as a supporting layer over a data model that was designed before any of it existed. For a team already running on one of these and happy with it, that can be enough.

Salesforce as a platform

Salesforce isn’t an ATS; it’s a platform, and it is only as good as the recruitment layer somebody builds on it. That is exactly why several Dutch ATS vendors are built on it. The question for a buyer is usually whether they want to own that layer, and what happens to it when the person who built it leaves.

AI-native systems

The newer category, where the model, the record and the reasoning were designed together. Simply is in this one. That means something specific and slightly awkward to say on our own blog: Simply does not integrate with the systems above. It is an alternative to them. If you are looking for a product to bolt onto Bullhorn or Carerix, this is not it, and the comparison pages are the honest place to start instead.

The implementation process step by step

If you take route one, the shape is: connect, map fields, test on real conversations, expand. Weeks two and three are the mapping, and that is where the effort actually lands.

If you take route two, the shape is a migration. Here it is, week by week.

Week 1: What you have, and what carries over

• We count what is actually in your current system, per object, and put it next to what you think you have. Duplicates and empty fields surface here.
• You decide what moves. Not everything should.
• The uncomfortable part comes first: a migration is only as good as your export. What your current vendor will not hand over, nobody can read in.
• What you need to supply: half a day from someone who genuinely knows your current setup.

Week 2: Your model, and the channels

• Your objects, fields and relations are configured to match how your desk actually works — including the ones the standard template never had.
• Mailboxes and calendars are connected, so from this point communication lands on the candidate rather than in somebody’s inbox.
• Historic email does not come out of an ATS export. It lives at your mail provider, and connecting the mailbox is what fixes that going forward.

Week 3: Running in both, on purpose

• A few recruiters work in it for real for a week, alongside your current system.
• Anything that grates gets changed while changing it is still cheap.
• Yes, that is a week of double work. It is annoying, and it is the reason week four is boring.

Week 4: Final sync, old system read-only

• A fresh export is loaded so the delta from week three comes across.
• Your old system goes read-only. No big-bang weekend.
• From here it is one-way: there is no connection back, which is the point.

Common mistakes in ATS integration

Mistake 1: Buying on the length of the integration list

A connector gallery with three hundred logos mostly measures how many partnerships were announced. A dozen are maintained and the rest are a form that emails somebody. Ask which ones a named customer is using in production today, and ask what happens when a token expires at three in the morning.

Mistake 2: Treating field mapping as setup rather than as maintenance

Mapping is not a one-time task. Your ATS changes, the AI vendor changes, someone adds a dropdown option. If you take route one, name the person who owns that mapping before you sign, because the alternative is discovering they don’t exist in month five.

Mistake 3: Accepting a score you can’t recalculate

Under the EU AI Act, recruitment AI that ranks or supports decisions sits in high-risk territory. “The model said so” is not an answer you can give a candidate. Ask any vendor to show the arithmetic behind a match, and treat an inability to do so as the answer.

Mistake 4: Not asking where each part of the processing happens

“Everything is EU-based” is one sentence covering two different facts, and they are usually not both true. Ask separately where the platform and the candidate data are hosted, and where model inference runs. Then ask for both in writing. Enterprise security is the minimum here, and so is a straight answer about what a vendor has not certified.

The impact on your organization

What does getting this right actually deliver? Honestly stated, because a number measured in someone else’s workflow is not a number from yours:

• Less time on admin per conversation. Teams commonly report the write-up going from around twenty minutes to a couple, which across a day of interviews is the largest single line item.
• Better data quality. More filled fields, in a consistent format, with fewer of the half-empty profiles that make search useless later.
• Faster time-to-submit. Candidates reach clients sooner, which is a direct competitive position rather than an efficiency metric.
• Higher recruiter satisfaction. Less admin, more of the work people took the job for.
• Scalability. Growing the team without adding admin headcount at the same rate.

Do the arithmetic on your own numbers rather than on ours: hours of admin per recruiter per week, times the hourly rate, against the total cost of what you would run. If you are comparing route one and route two, count the integration ownership on route one’s side of the ledger. It is the line everybody leaves out.

How Simply handles the systems around it

Simply is the system of record, so the integration story is deliberately short and complete.

• Six connectors, and we name all six. Gmail, Outlook, Google Calendar, Zoom, Slack and HubSpot, connected once per workspace through OAuth, with tokens held and rotated outside the app. No logo wall with “and 300 more” underneath it.
• No sync with another ATS. Not Bullhorn, not Carerix, not anything else in that category. It is a decision, not a gap.
• A migration instead. Four weeks, one-way, temporarily free of charge with our migration team, with a mapping you review.
• A public API for the rest. REST endpoints, workspace API keys, HMAC-signed webhooks and an MCP server, so what the six connectors don’t cover is something you can build rather than something you wait for.
• Omnichannel recording: online meetings captured natively through Google and Microsoft with no bot joining as a guest, phone calls through the desktop app, in-person conversations through the mobile app. All of it lands on the candidate.
• Summaries shaped by conversation type, and transparency so every AI sentence is clickable back to the moment it came from.
• CV parsing and formatting into your own house style, with per-row confidence and provenance.
• Proposals rather than silent writes: every AI change names the field, the value and the source, and waits for a human.

Want to see what your current setup looks like on the other side? Bring your export to the call. Not a generic pitch — your objects, your fields, and a straight answer about what would not survive the move.

Or read more first: how to get started with AI in recruiting, or why AI should fill your data automatically.

The first two weeks: what to expect

After go-live you’ll notice the difference in the first week. Conversations get summarised, and the data points that used to need typing arrive as proposals on the record instead. Most teams need two or three days to adjust to reading and approving rather than typing. After that, the time saving becomes obvious. Recruiters who previously spent twenty minutes per conversation on notes and data entry spend a fraction of that on checking what was proposed. The second week is usually when someone asks for a field that does not exist yet — and finds out that adding one is a setting rather than a support ticket.