AI workplace directions: confirm the right site before a shift
A practical workflow for checking identity, the live assignment and the approved entrance before sending safe directions to a temporary worker.
Updated:

The short answer
An AI worker hotline can verify the caller, read the current assignment, check the approved workplace record and send the correct entrance instructions. It can also ask the worker whether they can follow those directions. The system should record these as separate facts because a roster entry, the worker's reply and a decision to dispatch are not the same thing.
It should not invent a route, reveal access codes without the required check, tell someone to cross a vehicle yard or decide that a late arrival is absence or misconduct. If the location record conflicts with the worker's assignment, the gate is closed or the route looks unsafe, the automated flow stops and a named human owner takes over.
Keep one approved arrival record for each workplace
A postal address is rarely enough for a warehouse, factory, hospital campus or construction site. The useful record also contains the worker entrance, gate or building, pedestrian approach, public transport stop, parking rule, reception hours, required identification and an after-hours contact. Each field needs an owner, review date and version.
EU workplace rules require traffic routes and access to be arranged so pedestrians and vehicles can move safely. EU-OSHA guidance also stresses separation between pedestrian and vehicle routes. The hotline should therefore use instructions approved by the site, not a route guessed from a consumer map or copied from an old message.
Verify the assignment before sharing site details
Start with the agency's proportionate identity check and match the caller to a current assignment. Then confirm the workplace name, shift date and start time in plain language. This prevents a map link for Client A from being sent to a worker scheduled at Client B and avoids disclosing private site instructions to an unverified caller.
The minimum useful lookup is usually worker, active assignment, site and shift. GDPR principles require personal data to be adequate, relevant, limited to what is needed and kept accurate. The directions flow does not need payroll notes, performance history, medical details or the names and locations of other workers.
Keep four checks separate before dispatch
A safe workflow does not turn identity, assignment, the worker's answer and the arrival record into one yes or no result. It reads each source separately and keeps its version or timestamp. "Scheduled" means the live roster contains an assignment. "Can travel" is what the worker says during this contact. Neither result gives AI authority to allocate work or change attendance.
Use the table as the contract between telephony, the ATS or roster, the site directory and the human queue.
| Check | Approved source | What AI may confirm | Stop condition |
|---|---|---|---|
| Caller | Proportionate identity check | The caller matches the worker record | Identity cannot be confirmed |
| Current assignment | Live roster or ATS with version and timestamp | Date, start time, client and workplace | The record is missing, stale or conflicts with a message |
| Worker response | Statement made during this contact | Can travel, cannot travel or unsure, in the worker's own words | The answer raises a safety, health or employment issue |
| Arrival record | Site directory approved by the agency or client | Address, pedestrian entrance and approved map link | The gate, access rule or route is unclear, closed or unsafe |
Send a safe route, not only a map pin
A good response states the exact site and entrance, then sends the approved link through the worker's agreed channel. It describes the final approach with visible, non-sensitive landmarks and says which pedestrian gate or reception desk to use. Dates, gate numbers and local names should be tested in every supported language.
Instructions must not encourage shortcuts through loading bays, vehicle gates or restricted areas. If safe access depends on personal protective equipment, a shuttle, a security escort or a site briefing, the record should say what the worker must do and who is authorised to help. AI should never improvise an exception to a safety or access rule.
Confirm what happened after the message
Sending a link is an action, not an outcome. Ask whether the worker can see the entrance and whether the instructions solve the problem. If yes, record the time, workplace record version and confirmation. If the shift is close, an optional arrival check can show that help worked without tracking the worker continuously.
If there is no confirmation, use the branch's agreed follow-up rule. A failed SMS, unanswered return call or closed app does not prove that the worker ignored instructions. It creates an unresolved case with a clear next owner.
Model scenario: an old message points to the wrong gate
At 05:40 a worker calls about a 06:00 shift. Yesterday's message says Gate 2. The live assignment still names the same site, while site-directory version 14 now says Gate 4. After the identity check, the assistant states the site and shift, sends the approved Gate 4 link, asks whether the worker can follow it and records the reply with the source versions.
This is a test scenario, not a customer case. If the roster names another site, the directory has no approved entrance or the worker says the route is unsafe, the assistant stops. A person decides whether to dispatch, replace, change the shift or record attendance.
Escalate conflicts, closed gates and unsafe situations
Hand the case to a person when the schedule and location records disagree, the worker is at a different site, the entrance is locked, security refuses access, the map link fails or the worker reports poor lighting, moving vehicles or another hazard. The handoff should include confirmed facts, what was already sent and a callback number.
A coordinator or site contact decides whether the worker should wait, move to another entrance, use alternative transport or return home. The same people decide attendance status, replacement, pay, reimbursement, fault and any employment consequence. The hotline records and routes the problem; it does not turn navigation uncertainty into a personnel decision.
Connect telephony to the live schedule and site directory
The call platform needs a controlled read from the current schedule or ATS and a separate site directory with approved arrival instructions. The case written back to the CRM or dashboard should include a stable event ID, the site-record version, delivery status and the human owner if escalation was required.
When an integration is unavailable, the assistant may collect the worker's account and create a pending case. It must not claim that an address or shift was verified. Retried lookups should check the newest record so a delayed event cannot overwrite a later correction. The caller should also be told that the service uses AI, in line with the configured transparency notice.
Run seven acceptance tests before launch
Use controlled records and the real message-delivery path. Run each case in every supported language, including local site names and gate numbers.
| Test | Evidence to capture | Pass condition |
|---|---|---|
| Current assignment and entrance | Roster and site-directory versions | The answer matches the current client, date, time, site and approved entrance |
| Stale message | Old link plus current records | The old link is not reused; the current record is sent or the conflict is handed over |
| Similar site names | Two sites with near-identical names | The assistant states the full client and address and never chooses from a partial match |
| Worker cannot travel | Call transcript or structured reply | The statement is recorded and routed without marking absence, refusal or fault |
| Closed or unsafe entrance | Test report of a locked gate or moving vehicles | The assistant stops, names the human owner and gives no improvised shortcut |
| Delivery failure | Failed SMS or messaging receipt | The case stays unresolved and follows the approved verbal or human fallback |
| ATS or directory unavailable | Simulated read failure | The result is called unverified, a pending case is created and no unknown old value is used |
Measure corrections and keep decisions with people
Measure verified lookups, delivered links, worker-confirmed resolutions, time to human pickup, wrong-location corrections and staff repair time. Review repeated problems by site because several direction calls may reveal stale instructions or unsafe signage.
AI can retrieve, explain, confirm and document approved directions. People own safety exceptions, access overrides, dispatch, schedule changes, attendance and employment decisions.
FAQ
Can an AI voice agent confirm a worker's location and availability before dispatch?
It can confirm the current assignment and workplace from approved systems, then record whether the worker says they can travel. It should not infer availability from silence or decide dispatch, allocation, absence, replacement or pay.
Should AI tell a worker to use a different gate when the main entrance is closed?
Only when that alternative is already approved for the worker's site and shift. It should not invent a shortcut, disclose an unapproved access code or direct someone through a vehicle or restricted area. Otherwise it escalates to the named site contact.
Can the system mark the worker late or absent if they cannot find the workplace?
No. It can record call time, confirmed assignment, instructions sent and the worker's status. Attendance, absence, fault, pay and disciplinary consequences remain decisions for authorised people under agency and client policy.
Sources and further reading
- EUR-Lex: Directive 89/654/EEC on minimum workplace safety requirements
- EUR-Lex: Framework Directive 89/391/EEC on safety and health at work
- EU-OSHA: Workplace transport safety
- EU-OSHA: EU e-guide on work-related vehicle safety risks
- EUR-Lex: General Data Protection Regulation
- EDPB: Guidelines 4/2019 on Data Protection by Design and by Default
- European Commission: Guidelines on AI Act Article 50 transparency obligations
- EUR-Lex: Regulation (EU) 2024/1689, including human oversight requirements

