Short answer
An AI onboarding pilot should improve one defined part of how new hires learn and question the work. Keep payroll, identity, access, policy acknowledgements, safety training, supervision, and employment decisions in their authoritative human and administrative processes. For one comparable cohort, verify role access and practical learning at real work moments, route questions to accountable owners, and review what changed before expanding the workflow.
Separate administration, learning, and listening
Onboarding is a system of different jobs. Treating them as one AI workflow makes ownership and evidence unclear.
| Onboarding job | Appropriate system | Evidence of completion |
|---|---|---|
| Identity, right-to-work, payroll, benefits, and policy records | Authoritative HR, payroll, identity, and compliance process | Valid record in the system that owns it |
| Accounts, equipment, workplace access, and required materials | IT, facilities, security, or operations workflow | Access tested in the actual work context |
| Safety information, training, and supervision | Competent safety and operational process | Required training plus demonstrated understanding or supervised practice as applicable |
| Role expectations and priorities | Manager and role owner | New hire can explain what matters now and where to ask when priorities conflict |
| Practical local know-how | Trainer, buddy, approved guidance, and supervised work | New hire can use the current source in a relevant example |
| Experience and unresolved questions | Human check-in, group review, questionnaire, or focused conversation | Context reaches an owner who can answer or change the process |
Acas's current induction checklist includes documentation, emergency procedures, equipment and access, introductions, objectives, policies, job-specific training, reviews, and development needs. Acas also warns that induction should not be reduced to ticking boxes. Its planning guidance recommends adapting the programme to the role and considering adjustments and delivery format.
AI may help retrieve approved information, organise questions, or summarise a bounded conversation. It should not become the system of record or an evaluator of the new hire.
Define early-work moments
A universal day 7, 30, and 90 schedule can miss the point at which a role first uses a process. Map moments to the work:
- Before arrival: essential information, access needs, and named contact.
- First workday: safe access, equipment, emergency route, and immediate role orientation.
- First supervised task: current instruction, demonstration, practice, and questions.
- First exception: where to find the approved escalation path and who owns the decision.
- First independent cycle: what the new hire can use, what remains unclear, and what supervision continues.
- Cohort review: repeated access or transmission gaps, corrections made, missing voices, and next test.
The timing differs for a call-centre adviser, machine operator, nurse, engineer, store assistant, and remote analyst. Anchor the review to role milestones rather than choosing dates because the software offers them.
For safety-critical work, the formal training and supervision process determines readiness. HSE's new-to-the-job guidance tells UK employers to plan induction, provide relevant information and training, ensure adequate supervision, explain how to raise concerns, and check understanding during the first days and weeks. OSHA's guidance for young and new workers likewise stresses understandable hazard training, a mentor or experienced worker, questions, and competent safe practice in the US context.
Copy this cohort pilot brief
Choose one role, one location or comparable setting, and one onboarding question.
| Pilot field | Decision before launch | Evidence after the cohort |
|---|---|---|
| Work moment | Task or exception the pilot examines | When it occurred for each eligible hire |
| Practical outcome | What a person should be able to find, explain, or do under the approved process | Demonstration or example reviewed by the responsible human |
| Authoritative sources | Current documents, people, and systems | Outdated, conflicting, missing, or inaccessible sources found |
| Employee question | A neutral question about the work context | Accounts, counterexamples, and unanswered questions |
| Cohort | Role, site, start window, eligible count | Participants and relevant missing groups |
| Access | Paid time, device or channel, language and accessibility route | Where access failed or support was needed |
| Human owner | Manager, trainer, HR, IT, safety, or process owner for each issue | Response, correction, action, or reason for no change |
| AI boundary | Permitted sources and output; prohibited inference or decision | Unsupported output, correction, and stop-rule log |
| Review | Date and criteria to continue, change, or stop | Evidence supporting the next decision |
The pilot is about the onboarding process. Keep individual performance, potential, retention risk, and probation decisions outside the AI output.
Ask questions tied to work evidence
“How satisfied are you with onboarding?” can be a useful measure when defined properly, but it does not locate a missing instruction. Add questions that a process owner can investigate:
- Which part of the role is clear enough to use now? Give a recent example.
- Which approved source did you use for that task?
- Where did the written process and observed practice differ?
- What happened the first time the normal route did not fit?
- Who could answer, and how long did the work remain blocked?
- Which demonstration, example, or supervised practice helped?
- What is still unclear, and who should own the answer?
- What did we ask that is not relevant to this role?
Do not ask the AI to infer whether a person is capable, committed, culturally aligned, or likely to leave. A question can expose a gap in the process without turning the new hire into the object being scored.
Worked example: access completed, learning incomplete
This fictional service organisation pilots onboarding with 12 new support coordinators. The administrative system shows that all 12 accounts were created. That proves account creation, not that the coordinators can use the right knowledge in an early customer exception.
Ten employees take part in a focused conversation; two do not, so the findings cannot describe the whole cohort. Four participants say they struggled to locate the approved escalation path. The training owner checks the source rather than accepting the theme as fact and finds that the welcome page links to two documents, one superseded. Other participants describe a live trainer demonstration that worked well.
The process owner removes the outdated link through the normal document-control route and adds a supervised fictional exception to the next cohort's training. No customer case or personal performance record is used.
For the next cohort, the review records:
- 11 eligible hires and who had access to the exercise;
- whether each person could locate the current source with the permitted support;
- questions the source did not answer;
- whether the trainer gave the planned demonstration;
- any conflicting source that reappeared; and
- what participants said was still missing.
Those checks show whether the revised process was implemented and usable in the exercise. They do not prove faster productivity, better retention, or individual competence across the full role.
Test the AI-supported step
Use fictional or appropriately controlled content before involving new hires.
| Test | Expected behaviour |
|---|---|
| Approved source contains the answer | Output stays within the source and points to it |
| Two sources conflict | Output exposes the conflict and routes it to the named owner |
| Source is missing | Output states the gap rather than inventing a policy |
| Question concerns a safety procedure | Output directs the user to the approved current procedure and human route |
| Prompt asks for a hire's potential or departure risk | Output does not produce the inference or score |
| Summary changes the employee's meaning | Reviewer can correct it before the brief is used |
| Employee raises an access or adjustment need | Workflow uses the appropriate confidential human process |
| Product, prompt, model, or source changes | Owner retests before relying on the changed workflow |
The NIST AI Risk Management Framework Core is voluntary and not an HR certification. Its Govern, Map, Measure, and Manage functions are useful for documenting purpose, roles, affected groups, tests, limits, monitoring, and human oversight.
Review the cohort without rating people
Use process measures with explicit denominators:
- eligible, invited, and participating new hires;
- access provided and access failures;
- planned and completed human reviews;
- source conflicts found and corrected;
- practical exercises completed under the responsible training process;
- questions answered, still open, or routed elsewhere;
- actions implemented by the promised date; and
- roles, shifts, locations, or access needs missing from the evidence.
Compare cohorts only when role, work moment, source version, and opportunity are sufficiently similar. A higher participation rate does not prove trust. A completed checklist does not prove practical understanding. Retention is a later outcome with many possible causes.
The employee check-in software guide covers recurring check-in workflows. This guide owns the narrower cohort-pilot decision: whether an AI-supported step helps new hires access and question real work knowledge.
Where Lontra can fit
Lontra can support a focused employee conversation at a chosen onboarding moment and prepare an action-oriented manager brief. Managers receive the brief, not employees' raw responses. Human owners check the context, correct summaries, answer questions, and change the onboarding process.
It does not replace the HRIS, payroll, identity, IT access, required training, safety procedures, supervision, adjustments process, or probation and performance decisions. Do not use conversation output to score a new hire.
Explore Lontra's employee-conversation workflow and the onboarding use case, then bring one completed cohort pilot brief to the review.
Frequently asked questions
What is AI onboarding?
AI onboarding uses an AI-supported workflow for a defined part of induction, such as answering from approved sources or collecting contextual questions. Human owners remain responsible for training, support, corrections and employment decisions.
How is onboarding automation different from practical onboarding?
Automation can create accounts, route forms and record completion. Practical onboarding checks whether a new hire can access current information, perform or explain a real work task, ask for help and handle a relevant exception under human supervision.
What should an onboarding cohort pilot measure?
Measure eligible and participating hires, access, task evidence, unanswered questions, owner response, corrections, implementation and missing groups. Do not turn these measures into an individual potential or performance score.