How to implement AI in a staffing agency: a phased rollout plan
How to implement AI in a staffing agency without a big bang: start read-only on one desk, measure a baseline, and widen AI permissions step by step.
Start small and start read-only. Pick one desk and one workflow with real volume, such as client job requests arriving in the inbox or weekly availability checks, measure how it runs today, and let the AI read and suggest while a person approves everything before it is written or sent. Widen what the AI may do one step at a time, only when the numbers hold, with one named person who owns the result and can pause it.
Why roll out AI one desk at a time
When every team starts at once, nobody can tell whether a bad result came from the tool, the data or the process. We have written about why that stalls projects in why big-bang AI programmes fail in staffing agencies. One desk and one workflow gives you a clean comparison, and a mistake stays small enough to fix in a day.
Before you start: one owner, one desk, one workflow
Three decisions come before any software is switched on.
- The owner. One named person who answers for the result, usually the team lead or operations manager for the desk, not IT and not the vendor. The owner decides when to move to the next phase and has the authority to pause.
- The desk. A team with steady volume and at least one recruiter who is curious about the change. Avoid a desk in peak season or mid-migration.
- The workflow. One piece of work that repeats many times a week, has a clear point where it is done, and causes little harm if a draft is wrong because a person checks it first.
Good first workflows for an agency:
- Job request intake. Reading client job requests from email, PDFs and VMS portals into structured jobs. High volume, easy to check against what the recruiter would have typed.
- Availability checks. Asking the candidate pool who can work next week and recording the answers. Repetitive, and every answer improves the data you match on later.
- Candidate follow-ups. Reminders, confirmations and “are you still looking?” messages, drafted by the AI and approved by a recruiter.
Poor first workflows: anything that rejects or ranks candidates out of a process, anything that sends to clients without review, and anything nobody can measure today. Screening and ranking can come later, with the controls described in the compliance step below.
Measure a baseline before you switch anything on
Without a baseline, nobody can tell later whether the rollout worked. Take two normal weeks on the chosen desk and write down:
- job requests received, and the time from arrival to the first candidate contacted;
- share of job requests filled, and how many were missed or answered too late;
- hours per recruiter per week spent on the workflow, from a short self-reported log;
- errors found later: wrong dates, wrong client, duplicate records, messages to the wrong person.
Most of this comes from timestamps in the inbox and the ATS. If the ATS cannot tell you when a job request arrived, that is your first finding.
The rollout plan, phase by phase
Each phase has one rule: the AI gets one new permission, and only when the last phase met the bar you agreed in advance. Give each phase at least two full weeks of live volume, so you see a Monday peak and a quiet Friday more than once.
| Phase | What the AI may do | What people do | What to measure | Move on when | Stop or step back if |
|---|---|---|---|---|---|
| 0. Prepare | Nothing yet. Access is set up and tested. | Owner records the baseline, checks the data the workflow depends on and tells the desk what will change and why. | The baseline numbers above. | Baseline recorded, owner named, desk briefed. | Nobody can or will own it. Do not start. |
| 1. Read-only | Reads the inbox, ATS and portals for the chosen workflow and shows what it would do. Writes and sends nothing. | Works as before. Compares the AI's output with what they did, and flags differences. | How often the AI's version matches the recruiter's, and the kinds of misses. | The desk agrees the output is right often enough to be worth checking, and misses have a known cause. | Misses trace back to bad data. Fix the data first. |
| 2. Suggest and approve | Drafts records and messages. Nothing is saved or sent until a person approves it. | Approves, edits or rejects every item. Logs what they changed. | Share approved unchanged, share edited, time per item against the baseline. | Most items are approved with no or light edits, and time per item is down. | Edits stay high, or the desk is slower than at baseline. |
| 3. Act within rules | Writes and sends for cases the owner has defined, for example job requests from known clients or standard shift offers. Everything else goes to a person. | Handles the exceptions. Owner reviews the activity log daily, then weekly. | Baseline numbers again, plus errors that reached a candidate or client. | Results beat the baseline for a full cycle, with no error that a client or candidate would notice. | An error reaches a client or candidate. Step back to phase 2 for that case. |
| 4. Widen | Same permissions, one more desk or one more workflow. Not both at once. | The first desk's recruiter shows the next desk how they work with it. | Same numbers, per desk. | Repeat the phases for the new scope. They should go faster. | The second desk does not match the first. Find out why before a third. |
Write the “move on” and “stop” lines for your own desk before phase 1, with numbers from your baseline, so the rollout is judged on the bar and not on the one mistake everyone remembers.
How to widen AI permissions step by step
Think of permissions as a ladder, set per workflow and per desk:
- Read the inbox, ATS and portals.
- Draft records and messages for a person to approve.
- Write to records: create the job, update availability, add notes.
- Send to candidates: shift offers, reminders, confirmations.
- Act toward clients: confirm a booking, reply on a job request.
Climb one rung at a time. If a tool cannot show you a log of what it did yesterday, and why, you cannot run phase 3 safely. The questions to ask a vendor about this are in Is the AI in your ATS actually doing anything?
AI training for recruiters: train on the workflow
General AI courses rarely change how a desk works on Monday morning. The training that matters is short and specific to the workflow:
- what the AI does in this workflow, and what it never does;
- how to read its output and the activity log;
- how to correct it, and where corrections go;
- when to override it, and who to tell when something looks wrong.
Run it on the desk's own job requests, not on vendor examples. If the training is not tied to a change in how the work flows, people go back to the old way within weeks. We have argued that in AI adoption in staffing is a fit problem, not a training problem.
The EU AI Act points the same way. Since the AI Omnibus, Article 4 requires providers and deployers to take measures to support the AI literacy of their staff, but does not require them to guarantee any specific level for any individual. For high-risk uses such as screening and ranking candidates, Article 26(2) requires the people overseeing the system to have the necessary competence, training and authority, from 2 December 2027 (checked on the European Commission's AI Act Service Desk, October 2026).
What ChatGPT and Copilot are good enough for during the rollout
Many agencies already pay for ChatGPT or Microsoft Copilot (formerly Microsoft 365 Copilot), and for individual work they are often good enough:
- drafting and rewriting job ads and candidate messages;
- summarising long CVs, meeting notes and client briefs;
- preparing for client meetings and writing proposals.
Microsoft says Copilot only surfaces organisational data that the individual user has at least view permission for, and that prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs (Microsoft Learn, checked October 2026). For any tool, use a business plan with a data processing agreement before candidate data goes in, never a personal account.
Out of the box, they do not run a staffing workflow end to end: watch the job request inbox at 06:00, create the job in the ATS, ask the right candidates, record who said yes and log every step. You can build parts of that on top of them, but then you own the integrations, the permissions and the log, and you still need a system of record.
The compliance step before the AI acts
Do this between phase 2 and phase 3, before the AI writes or sends anything on its own:
- Check which uses are high-risk. Reading job requests into structured jobs is unlikely to be. Analysing and filtering applications or evaluating candidates is, and choosing which candidates get a shift offer can be, since Annex III also covers targeted job advertisements and allocating work. The high-risk rules apply from 2 December 2027 (Commission timeline).
- Tell candidates and workers when AI is used in decisions about them.
- Keep a person with real authority on any decision that shuts someone out, such as rejecting an applicant. GDPR Article 22 already gives people the right not to be subject to a decision based solely on automated processing that significantly affects them.
- Sign a data processing agreement, keep the logs, and run a data protection impact assessment where screening is involved.
The full picture, including Norway and the UK, is in The EU AI Act and recruitment: what staffing agencies need to know.
Why AI rollouts fail in agencies, and how to avoid it
- Bad data in the ATS. Duplicate candidates, availability last updated a year ago, skills in free text. Fix the slice the first workflow needs, often with an availability campaign, before you judge the AI on it.
- Too many tools. A tool that adds a screen without removing a step is one more system to check. Name the step the AI takes over and stop doing it by hand.
- Nobody accountable. A steering group does not review the log on a Tuesday. One owner per desk, with time set aside.
- Training without a workflow change. Change the workflow and train on it in the same week.
- Widening too fast. A second desk and a second workflow in the same month hide what caused a problem.
- Silent failures. A flow stops at night and nobody is told. Ask who is alerted when the AI stops or is unsure, and test it during phase 1.
When to stop, and when to expand
Expand when the desk beats its baseline for a full cycle, recruiters would object if the AI were switched off, and the owner can explain every error from the log.
Stop or step back when errors reach clients or candidates, when the desk is slower than at baseline after the suggest-and-approve phase, or when misses keep tracing back to data you cannot fix yet. Stopping one workflow still tells you where the work is.
What this means for a staffing agency
A phased rollout needs three things from the software: AI agents that work across the inbox, ATS, VMS portals and candidate channels; a person in the loop on the steps you choose; and a log of every action.
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, Carerix or Access, and connect the AI agents to it.
Two published customer stories show what a structured start can look like. Ecura Care, Norwegian healthcare staffing, first rolled Globus.ai out to parts of the organisation in autumn 2023, then set up a dedicated project group in September 2025 to map workflows and scale it across the organisation, naming one clear point of contact who knows its processes as a key success factor; the story reports 15k+ jobs handled with Globus and a candidate acceptance rate above 70%. TXM Healthcare, UK healthcare staffing, started with setup, integration with Access, AI training on its own request patterns and guided user training; in the first user training the team found real customer requests that had been hidden in the inbox, and the story reports 8,000 emails processed per week.
Whichever vendor you choose, run the same phases and hold it to the same baseline. Other routes are real too: AI features in your current ATS, automation tools for simple fixed flows, or another agent vendor. Ask each for the same log and the same human approval steps.
Updated October 2026. Related: AI for staffing agencies: what it actually does today and Is the AI in your ATS actually doing anything? A buyer's guide.
Frequently asked questions
How do you implement AI in recruitment?
Start with one desk and one workflow that repeats many times a week, such as job request intake or availability checks. Measure a baseline for two normal weeks, let the AI run read-only, then suggest-and-approve, and only then act within rules you define. Name one owner who decides when to widen permissions and can pause.
What should a staffing agency automate first with AI?
Pick work with high volume, a clear point where it is done and little harm if a draft is wrong: reading client job requests into structured jobs, weekly availability checks, or candidate reminders and confirmations. Leave screening and ranking candidates until the basics run well and the compliance checks are in place.
What AI training do recruiters need?
Short, workflow-specific training on the desk's own job requests: what the AI does and never does, how to read its output and the activity log, how to correct it, and when to override it. A general AI course helps with literacy, but usage fades unless the workflow itself changes in the same week.
How long does it take to roll out AI in a staffing agency?
It depends on volume and data quality more than on the software. Give each phase at least two full weeks of live volume, so you see more than one peak, and plan for a few months of steady effort on the first desk. Later desks should go faster because the rules and training already exist.
Is ChatGPT or Copilot enough to roll out AI in an agency?
For individual work, often yes: drafting job ads and messages, summarising CVs and notes, and helping recruiters get used to AI. Out of the box they do not run a workflow across the inbox, ATS and VMS portals or write records with a log. Building that on top of them means owning the integrations, permissions and logging yourself, plus a system of record.
Want to see how AI agents handle your own job requests before you plan the first desk? Book a meeting with Globus.ai.
