Consulting and recruiting
AI for staffing firms: what your ATS history is good for, and what to keep out
By SourceX Editorial · Reviewed by Noah Loul ·
Short answer
AI for staffing firms is most useful on linked ATS history: job orders, submittals, client interview feedback, placements and lost-order reasons. Those records support rediscovery, matching, scheduling and fill forecasting. Keep out protected traits, screening results, I-9 and payroll details, health notes and unreviewed recruiter notes, and check client terms before job order details are reused.
Key takeaways
- The ATS history worth using links a job order to submittals, client feedback and a placement or a recorded loss reason.
- Most staffing AI uses need less candidate detail than teams assume; fill forecasting and job order analytics need none.
- EEO fields, background and drug screen results, I-9 documents, payroll and health notes stay out of AI projects by default.
- Recruiter notes are often the most valuable and the riskiest field, so they need human review before any reuse.
- Switching on a vendor's AI feature and licensing records to an AI developer are separate decisions with separate contract questions.
Where does AI help a staffing firm most?#
AI helps a staffing firm most in work that repeats on every requisition: resurfacing past candidates, ranking submittals, scheduling interviews, drafting outreach and predicting which job orders will fill. Each of those tasks runs on history already sitting in your applicant tracking system, so the useful question is which records each one needs.
The quality of that history decides the result. A firm whose recruiters logged submittal outcomes, client interview feedback and lost-order reasons gives any tool something to learn from. A firm whose ATS holds resumes and little else gets keyword search with a new label.
The table maps common uses to the ATS records they draw on, the personal data each one needs and the caution that comes with it. The uses that need the least candidate detail are the safest place to start.
| Use case | ATS records it draws on | Personal data it needs | Candidate-data caution |
|---|---|---|---|
| Candidate rediscovery | Past submittals, skills tags, placement end dates, availability updates | Name and contact details, because a recruiter must reach the person | Apply opt-outs, do-not-contact flags and deletion requests before resurfacing anyone |
| Job order matching and ranking | Job orders, required skills, submittal and interview outcomes | Skills and work history; identity is not needed to score a fit | Strip protected traits and proxies such as photos, birth dates and graduation years; automated ranking may fall under local hiring-tool rules |
| Interview scheduling | Interview events, recruiter calendars, client contacts | Contact details only | Low risk if contact details stay inside the scheduling tool |
| Outreach and job description drafting | Templates, past job descriptions, reply rates | Usually none in the prompt | Keep message threads about personal circumstances out of prompts |
| Fill forecasting | Job order status changes, stage dates, lost-order reasons | None; stage-level records are enough | Client terms may still limit how job order details are reused |
| Redeployment at assignment end | Assignment end dates, extension history, client feedback | Identity of workers on current assignments | Comments on worker performance are sensitive and need review |
| Recruiter coaching and desk analytics | Calls, tasks, submittal-to-interview ratios by desk | Recruiter identity, which is employee data | Check employee privacy notices before analyzing individual recruiters |
Which ATS records carry the most value?#
The ATS records with the most value connect a job order to what happened next: who was submitted, who the client interviewed, who was placed, how long the assignment ran and why it ended. Isolated resumes say who applied. Linked records show how recruiters and clients actually made decisions.
In Bullhorn and similar staffing systems, that chain usually spans several linked records: the job order, the submittal or pipeline record, interview and placement records, and the notes and activities attached to each. Back-office systems hold timesheets and pay, while client vendor management systems keep their own copy of requisitions and submittals under the client's terms.
A quick linkage test tells you where you stand. Pull a few dozen closed job orders from different desks and years and trace each one to a placement or a recorded loss reason. Where most of them break, fix logging before buying tools.
- Job orders: title, skills, work arrangement, required certifications, status history and close reason.
- Submittals and pipeline stages: which candidates went to which order, with dated stage changes.
- Client interview feedback: accept and reject decisions, often with a short stated reason.
- Placements and assignments: start and end dates, extensions, conversions to permanent hire and end reasons.
- Lost and unfilled orders: why the firm did not fill a requisition and who filled it instead, when known.
- Recruiter activity: calls, tasks and notes that explain the steps between stages.
What candidate data should stay out?#
Candidate data that identifies a person beyond the job, or records a sensitive judgment about them, should stay out of AI projects by default. That covers protected-trait fields, screening results, identity and immigration documents, pay and banking details, and health or accommodation notes, wherever they sit in the ATS or back office.
Some fields are not obviously sensitive but act as proxies. Graduation years and birth dates point to age, photos point to race and gender, and home addresses can stand in for neighborhood demographics. A matching model trained on those fields can learn patterns your firm would never defend to a client or a regulator.
Which laws may apply depends on where candidates live and on the purpose of each use. Federal and state anti-discrimination law, the FCRA for background reports, state privacy laws such as the CCPA, salary history rules, local rules on automated hiring tools and, for candidates in the EU, the GDPR may all be relevant. They are assessed project by project with counsel.
| Record or field | Default | Why |
|---|---|---|
| EEO self-identification and demographic fields | Exclude | Protected traits that are meant to stay apart from hiring decisions |
| Background check and drug screen results | Exclude | Screening laws such as the FCRA and the screening vendor's terms may limit other uses |
| Form I-9, work authorization and ID documents | Exclude | Federal law limits what I-9 information may be used for, and it adds nothing to workflow AI |
| SSNs, bank details, tax forms and payroll | Exclude | Financial identifiers that belong in the back office only |
| Medical, accommodation and workers' compensation notes | Exclude | Health information carries its own confidentiality obligations |
| Salary history and current pay | Review | Some states and cities restrict asking about or relying on pay history |
| Resumes and work histories | Review | Useful for internal matching, but full of contact details and identifiers |
| Recruiter notes and client feedback | Review | High value, yet often mixed with opinions, health and family details |
| Client names, job orders and bill rates | Review | Client agreements may limit use to delivering the service |
What to ask before switching on AI features in your ATS#
Before switching on AI features in your ATS or a connected add-on, get written answers on how the vendor will use your candidate data. Matching, summarizing and outreach features process the same records described above, and the rules that govern them sit in the subscription agreement and the data processing addendum, not in the product settings screen.
Licensing records to an outside AI developer is a separate decision. An internal tool processes your records for your own recruiting. A license gives a developer permission to use a prepared copy of selected records for a defined purpose, so the records serve a new purpose and need their own rights review. The firm keeps ownership and licenses, rather than sells, a de-identified package built around the workflow, not the people: how job orders move, which submittals clients accept and why orders are lost.
| Question for the vendor | Why it matters |
|---|---|
| Is our data used to train or improve models shared with other customers? | Shared-model training puts your candidates' records to work in a product your competitors also use. |
| Can we opt out, and does the opt-out cover data already processed? | An opt-out that applies only going forward leaves past records in scope. |
| Which subprocessors, including AI model providers, receive our data, and where? | Each subprocessor is another party holding candidate data under its own terms. |
| Are prompts, outputs and match scores retained, and for how long? | Derived data can outlive the records it came from, including after a deletion request. |
| Do marketplace add-ons, texting tools and resume parsers follow the same terms? | Integrations often carry their own agreements and data flows. |
| If the feature ranks or screens candidates, who handles bias audits and candidate notices? | Rules on automated hiring tools may place duties on the firm using the tool, not only on the vendor. |
| What happens to our data and any derived data when we cancel? | Export formats, deletion confirmations and timelines are easier to agree before signing than after. |
How to audit your ATS history before any AI project#
An ATS audit sorts the field list, and the paper trail behind it, into keep, review and exclude. The COO or operations lead usually runs it with the ATS administrator, and counsel reviews the notices and contracts it surfaces.
Free-text notes deserve the most attention. Open-source tools such as Presidio can flag many personal identifiers in text, but Presidio's own documentation warns that automated detection gives no guarantee of finding all sensitive information and that additional systems and protections should be used. Plan for human review of notes, not a single automated pass.
- Export the full field list, including custom fields, and mark each one keep, review or exclude.
- Collect every version of the candidate privacy notice and application consent language, with the dates each was live.
- Pull client MSAs and VMS program terms and flag confidentiality and data-use clauses.
- List integrations that already copy data out, such as resume parsers, texting tools and VMS connectors.
- Sample recruiter notes across desks and years, looking for health, family, immigration and opinion content.
- Confirm that deletion requests and retention purges have run in the ATS and in every connected tool.
- Name one owner for the project and one signer for any outside agreement.
Illustrative: an office staffing firm picks its first AI uses#
Illustrative: a fictional office and administrative staffing firm with four branches runs Bullhorn as its ATS, a texting add-on for outreach and a separate back-office system for payroll. Two of its largest clients release orders through vendor management systems. The CEO wants recruiters to stop rebuilding searches from scratch, and a board member has asked whether the firm's records could be licensed.
The audit finds EEO answers in a separate module, older recruiter notes that mention candidates' health and childcare, and two client MSAs that limit job order details to service delivery. Background screening results never entered the ATS; they live in the screening vendor's portal. Placements link cleanly to job orders, but only two branches logged lost-order reasons.
The firm starts with rediscovery and scheduling, using skills tags, submittal outcomes and placement end dates, with do-not-contact flags applied and notes held back until reviewed. It sends its ATS vendor the questions above and postpones automated ranking until counsel confirms which hiring-tool rules apply. For licensing, it narrows scope to de-identified workflow records from the two branches with complete loss reasons, with client names removed and the two restricted clients excluded. Payroll, screening and EEO data never leave their own systems.
Mistakes staffing owners make with ATS data and AI#
The most common mistake with ATS data is treating the database as one asset. A candidate pool, a placement history and a set of client job orders carry different rights and different risks, and each needs its own decision.
The other mistakes are smaller but just as avoidable, and each is cheaper to fix before data moves than after.
- Letting recruiters paste resumes or notes into general-purpose chatbots before the firm has an AI use policy.
- Switching on vendor AI features without reading the data terms or settling who handles bias audits.
- Training a ranking model on past placements without removing proxies for age, race or gender.
- Assuming records from an acquired agency carry the same candidate notices and client terms as your own.
- Treating job orders received through a client's VMS as the firm's own, when the client's program terms govern them.
How SourceX approaches staffing records#
SourceX treats staffing records as workflow history rather than a candidate list, and runs any license through the SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery. During the fit check, a staffing firm describes its ATS, how many years of placements remain accessible and which record families exist; no candidate files change hands at that stage.
Rights review covers candidate notices, client MSAs and VMS terms. Preparation removes personal and confidential details, including candidate identities, contact details and protected traits, and the firm approves each step. Every package that proceeds gets a SourceX Evidence Packet recording provenance, licensing rights, permitted use, the privacy record and release authorization. SourceX's dataset rights are set out in the signed supplier agreement.
Frequently asked questions
Can recruiters paste candidate data into general-purpose AI chatbots?
Only under a written policy. Depending on the plan and settings, a chatbot provider may retain prompts or use them to improve its models, and pasting resumes or notes sends candidate data to a party the firm has not vetted. A workable policy names approved business tools, bans pasting screening, health and pay details, and lists permitted uses such as drafting job descriptions.
Should candidates be told that we use AI in recruiting?
Transparency is usually the safer course, and some places have notice rules for automated hiring tools. Many firms update the candidate privacy notice to describe AI-assisted sourcing and scheduling and keep dated copies of each version. Whether a specific notice is required depends on where candidates live and how the tool is used, so confirm with counsel.
Can we use AI on the database of an agency we acquired?
Treat the acquired database as its own question. Those candidates received that agency's notices, not yours, and its client contracts may carry different restrictions. Check both before merging records into your ATS or using them in any AI tool, and keep the source flagged so it can be scoped separately later.
Do opted-out or deleted candidates need to be removed first?
Yes, as a working rule. Before any reuse, confirm that deletion requests and retention purges have actually run in the ATS and in connected tools such as texting and resume parsing add-ons. Copies that survive in old exports, spreadsheets or backups are a common gap, so check those too.
What if our recruiters never logged outcomes consistently?
Partial history still helps. Internal tools can start with the desks and years where submittal outcomes and close reasons were recorded, and recruiters can begin logging them now. For licensing, gaps narrow what can be packaged, and a metadata fit check shows which record families and periods are complete enough to consider.
Sources
- Presidio's own documentation warns that "because it is using automated detection mechanisms, there is no guarantee that Presidio will find all sensitive information. Consequently, additional systems and protections should be employed." Source
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.