Is the AI in your ATS actually doing anything? A buyer's guide

Sometimes, and you can find out in a week. Most AI in an ATS today reads CVs, scores matches and drafts text, and those features are useful when someone acts on them. The test is simple: count what the AI wrote, sent or changed in your systems, with a log entry to prove it, and compare that with the suggestions nobody touched.

What "AI-powered ATS" usually means

When a vendor says its ATS is AI-powered, it can mean one of five different things. They are worth separating, because they do different amounts of work.

• Parsing and extraction. Reading CVs, application forms and documents into structured fields. One of the oldest AI features in ATS products, and often the most valuable, because few agencies still type a CV in by hand.
• Matching scores and search. Ranking candidates against a job, or searching the database in plain language. Good at finding people you forgot you had. It does nothing until a recruiter opens the list and acts on it.
• Generated text. Job ads, candidate summaries, outreach emails and call notes. Saves typing, but only if the draft is good enough to send with light edits.
• Copilots and chat. A chat box inside the ATS where a recruiter asks questions about records, priorities or a candidate. Helpful for one person at a time, and only when someone thinks to ask.
• Agents that act. Software that does a step of the work without being prompted: creates the job request from a client email, sends the availability check, books the shift, updates the record, and hands over to a person when it should.

The large ATS vendors now sell several of these. Bullhorn describes Amplify as "digital workers" with skills such as Enrich, Match, Screen, Outreach, Present and Extract, and says Amplify Chat can launch screens, start outreach sequences or update records directly. Recman lists CV parsing and AI matching, plus RecBot, which helps keep the CRM updated (Bullhorn and Recman, vendor pages checked October 2026). The labels are similar across vendors. What they do on a Tuesday in your agency is the part you have to measure.

Is ATS AI real, or a new label on old features?

Both exist. An ATS on its own is a database with workflows, and the classic keyword filter is a rule, not AI. CV parsing and semantic search have been around for years and work well. What is new since 2023 is language models that write text and, more recently, features that take actions.

The honest view: parsing and matching are good enough for most agencies, and you should not pay extra for a new name on them. The question that separates vendors is the last category. Does anything happen in your systems when nobody is prompting the AI?

How to test whether the AI in your ATS does anything: a one-week test

You do not need the vendor for this. Pick one desk, such as one temp desk or one recruitment team, and run it for five working days on your live data.

1. Before Monday. List every AI feature that is switched on and that you pay for. Note a baseline for the desk: job requests received, shortlists made, candidate messages sent and hours spent on admin last week.
2. Every day. Each recruiter spends five minutes filling one shared sheet, using the table below. Pull the ATS activity or audit log for the same day and count entries made by the AI rather than by a person.
3. Friday. Add it up per feature: how much it produced, how much was used unchanged or with light edits, how much it did without a prompt, and how often someone had to correct it.

Two numbers decide the verdict. The first is the acted-on share: outputs used divided by outputs produced. As a rule of thumb, below about one in three, the feature is mostly noise for that desk. The second is unprompted work: records written and messages sent by the AI with a log entry, with nobody asking for it. If that number is zero, you have assistants, not automation, which is fine as long as you are paying for assistants.

Watch for one trap. A feature that creates a task for a recruiter ("follow up with these candidates") has moved work, not done it. Count it under suggestions.

Questions to ask the vendor in a demo

Demos run on clean data and chosen examples. These questions move the conversation to your agency.

• What does it do when nobody is prompting it? Ask for one example from a live customer, not the roadmap.
• Where does it write? Which records can it create or change, in which systems, and with which permissions?
• Can I see the log? Ask to see yesterday's actions for a real account: what it touched, when, and why. It is the most useful of the three boring questions to ask before buying AI.
• What happens when it is wrong? Who is told, how a person takes over, and how a correction feeds back.
• Does it reach beyond the ATS? The job request inbox, client VMS portals and SMS are where much of an agency's volume sits.
• Can we reach our own data? Is there an open API, and can tools such as ChatGPT or Claude connect to it, for example through an MCP server, with the same permissions and logging as everything else?
• What is included, and what is metered? Which AI features are in your current plan, which are add-ons, and which are charged per use.

What it costs to find out

The one-week test costs about five minutes per recruiter per day, plus an hour on Friday to add up the sheet and pull the audit log. Most agencies can run it without asking the vendor for anything except, sometimes, an export of the activity log.

Not knowing costs more. AI features in an ATS can be part of the plan, an add-on, or charged by usage, and a feature that nobody acts on is still paid for every month. The reverse also happens: a desk turns off a feature that works because nobody measured it. A week of logging settles both.

If the test shows the AI mostly produces suggestions, that is not a failure. It tells you what you have bought. Then decide whether the next step is better use of what you have, or work done for you in the workflows where volume decides whether you win the job.

Red flags

• No log, or a log you cannot see. If the vendor cannot show what the AI did yesterday, you cannot check it, correct it or defend it to a client.
• "Agent" means a chat window. If a person still copies the answer into the ATS and presses send, the workflow is still manual.
• Results only from other customers. Vendor case figures are a reason to test, not a substitute for testing on your own job requests and data.
• Read-only integrations. AI that can see your ATS but cannot write job requests, notes or bookings back will add work, not remove it.
• Scores without reasons. A match score you cannot explain to a client or candidate is hard to use and, for screening and ranking, hard to defend under the EU AI Act.
• Silent failures. Ask what happens when a flow stops at 06:00. If the answer is that nobody is told, plan for missed job requests.

What this means for a staffing agency

For most agencies the one-week test ends the same way. Parsing and matching are doing useful work inside the ATS. The big volume, client job requests arriving in the inbox and portals, availability checks, shift offers and candidate replies, is still handled by people, because the AI suggests and a recruiter does the rest. Moving that work takes AI agents that act across the inbox, ATS, VMS and candidate channels, under rules you set, with a log of what they did and a named person in charge.

Globus.ai is one example: a full AI Agent ATS for staffing and recruitment agencies, with AI agents built in from job request intake to matching, screening, candidate communication, booking and compliance, and all agent activity visible to recruiters in a real-time Control Room. An agency can also keep its current ATS, such as Bullhorn, Recman or Carerix, and connect the AI agents to it. Two published results, both counted as work done rather than suggestions: Helsebemanning, Nordic healthcare staffing, with 32,000+ jobs auto-handled in 1 year, and TXM Healthcare, UK healthcare staffing, with 8,000 emails processed per week.

Whichever vendor you look at, run the same week of logging on it before you sign.

Updated October 2026. Related: What is an AI agent? Explained for staffing and recruitment and AI for staffing agencies: what it actually does today.

Frequently asked questions

Are ATS systems AI?

Not by themselves. An applicant tracking system is a database with workflows for candidates, jobs and clients. Most ATS products now include AI features such as CV parsing, match scores, text drafting or a chat assistant, and some add automations that act on records. The classic keyword filter is a rule, not AI.

What is an AI-native ATS?

The term is used for an ATS designed around AI from the start, rather than one with AI features added later. In practice, judge it by what it does, not the label: which steps of the work it carries out without a prompt, where it writes, and whether every action shows in a log you can read.

Is the AI matching in an ATS good enough?

Often, yes. Plain-language search and match scores are good at finding candidates already in your database, including people nobody has contacted for years. They are weaker where the real rules are not written down, and a person should decide who is put forward. Test it on your own jobs for a week before paying extra for it.

Which ATS connects to ChatGPT or Claude?

Check how, not only whether. Ask if the ATS has an open API or an MCP server, what a connected tool can read and write, whether it uses each user's permissions, and whether its actions appear in the same audit log. A connection that can read but not write helps one recruiter; it does not run a workflow.

Is AI screening in an ATS legal under the EU AI Act?

Yes, with duties. AI used to analyse and filter applications or evaluate candidates is high-risk under Annex III of the EU AI Act, with those rules applying from 2 December 2027. That means human oversight, logging and information to candidates. Reading client job requests or drafting emails is usually not high-risk.

Want to see AI agents handle a real job request from inbox to booked shift, with the log to show what they did? Book a meeting with Globus.ai.