Short answer
Conversational AI in HR support uses adaptive dialogue to collect, clarify or retrieve information through text or speech. It can support HR service questions, employee listening and conversation preparation, but it should not make consequential employment decisions or be treated as an objective account of an employee. Use it only when follow-up dialogue improves the task, with a defined purpose, limited data, transparent access rules and accountable human review. The most valuable application is ensuring input quality: a system that refuses vague answers and cross-references self-assessments against manager ratings creates an audit trail while forcing data quality at source.
Why does performance data remain unreliable?
Managers avoid giving honest feedback for predictable reasons. A director of people operations told us in May that managers consistently rate employees higher than warranted because they want to avoid conflict or preserve team morale. When a business needs to reduce headcount by twenty percent, HR teams discover they cannot identify top talent because the previous manager lacked the courage to document poor performance. The file shows three years of satisfactory ratings for someone the manager now describes as unbearable.
This is not a training problem. You can train managers as much as you want, but they will never consistently do the work. A senior HR leader in retail told us in December that the biggest piece of feedback from employees is that there is not enough development, and this is linked to a lack of communication. When you have great managers who are great people, you have happy teams. When you have great managers who are good at putting money in the till but not much at managing people, you have unhappy teams and higher turnover.
How does conversational AI address input quality?
Conversational AI can refuse to accept vague answers. In a demonstration we observed in May, the system asked a manager to specify when and how a task would be completed. When the manager replied with a general commitment, the system asked for a date, hours and method to verify. When the manager tried to move on, the system insisted on a concrete plan before proceeding.
This is not about replacing the manager. It is about having a version of the HR business partner that makes the manager do the form properly. A former retail operations director told us that if you can take a thousand versions of a good HR person and create an AI that makes managers give input at the quality required, otherwise they are not going to leave the room, then you have one hundred percent of what you need.
The system can also cross-reference self-assessments against manager ratings. In the same demonstration, the system showed an alignment score of fifty-two percent for one employee, indicating that the employee seemed to overestimate themselves. The employee scored almost five everywhere when talking, and the manager was a little bit lower. This discrepancy becomes visible to the HR team, who can then understand why it was like that and prepare the conversation.
What does an audit trail protect?
An audit trail protects the business when employment decisions are challenged. A people director told us in December that with new employment rights legislation, businesses are not going to be able to dismiss anyone without the right data. It is just getting harder and harder for businesses. The same director noted that in France, when a manager wants to dismiss someone, HR asks what is in the file. If the manager gave good ratings because they did not want to give a bad mood in the team, HR has nothing to support the decision.
Conversational AI can create this record by attributing everything: this being a good grade, this being a bad grade, with the justification and the exact quote logged somewhere. The director of people operations told us that everything is recorded, you cannot access it if you do not have all the admin rights, but it pre-attributes everything. This is not about secretly monitoring employees. It is about ensuring that when a manager makes a claim, there is authorized evidence to support or challenge it.
Where should conversational AI not be used?
Do not use conversational AI to make a final hiring, promotion, compensation, disciplinary or termination decision. Do not use it to infer health, disability, union activity or another sensitive characteristic. Do not use it to label an employee as likely to leave. Do not produce an opaque engagement or performance score. Do not expose raw employee conversations to people without a defined need. Do not collect information without a clear purpose, retention rule and access model. Do not present generated themes as objective truth.
The U.S. Equal Employment Opportunity Commission makes clear that federal employment discrimination laws apply when AI and automated systems are used in employment. The practical implication is simple: adding a model does not remove the employer's responsibility. The U.S. Department of Labor's workplace AI principles also emphasize worker input, transparency, protection of labor rights and meaningful human oversight.
For EU deployments, the EU AI Act prohibits AI used to infer emotions in workplaces except for medical or safety reasons. It also classifies specified recruitment, selection, task-allocation and performance-evaluation uses as high-risk, depending on the system's intended purpose. The GDPR separately requires purpose limitation, data minimization, accuracy, storage limitation, security and accountability. Employee consent is not an easy substitute for this work: the European Data Protection Board warns that power imbalance often prevents employee consent from being freely given.
What should HR teams ask before deployment?
Before deploying conversational AI for HR support, ask these questions:
- What exact decision will this conversation inform, and who owns it? Ask for one sentence without broad AI language.
- What data enters the system, where is inference performed and which subprocessors receive it? Distinguish approved company documents, employee input, public model knowledge and inferred data.
- What is retained, for how long and for which prohibited secondary uses? Employees may disclose information that the original conversation did not request.
- Can employees understand participation, confidentiality, access and the consequences of declining? A generic privacy notice is not enough.
- Can reviewers trace an output to authorized evidence and see uncertainty? The safest response to missing evidence is not confident improvisation.
- How are leading questions, hallucinations, accessibility failures and subgroup disparities tested? Ask about realistic failure cases, not only average accuracy.
- Can a person correct, contest or escalate an output before it affects them? Correction and escalation need an owner.
- What would make the pilot stop? Define the incidents and evidence thresholds before launch.
The NIST AI Risk Management Framework offers a vendor-neutral way to think about governance, mapping context, measuring risk and managing it over time. For organizations operating under UK data protection law, the ICO guidance on AI and data protection adds practical guidance on lawfulness, fairness, transparency, security and individual rights.
How should a pilot be structured?
A retail HR director told us in December that the ideal launch would be in a smaller country with less infrastructure, where the system can be tested without disrupting ongoing system implementations or training. Italy was mentioned as an example: a country with growth, limited HR support and basic performance management on paper, where employees are desperate for something.
The pilot should focus on one narrow question. Do not begin with "transform HR." Begin with a bounded need, such as reducing unresolved onboarding questions in one population or ensuring that performance ratings in one region are comparable to another. Define the action before collecting input. Name who will review the output, what they can change and how employees will hear what happened next. Minimize the data. Use only the attributes needed for the purpose. Exclude sensitive fields and free-text history that the pilot does not require.
Test failure cases. Include missing policies, ambiguous employee answers, conflicting evidence, sensitive disclosures and requests that must reach a human. Useful pilot measures include answer escalation rate, unsupported-inference rate, accessibility failures, subgroup failure rates, completion of agreed follow-up, time to resolve a request, employee trust and adverse incidents. None of these proves that the system understands a person.
What remains the manager's responsibility?
The manager still has to listen, test assumptions, explain the decision and own the consequences. A former operations director told us in May that conversational AI is preparation, not evaluation. The system can organize approved notes, list unresolved actions and draft questions for the next one-to-one. It should separate facts, interpretations and missing context. But a manager cannot delegate the conversation itself.
The same director noted that employees still need to have the face-to-face. The system can prompt managers, give them ideas, help them structure the conversation, but they must still have the face-to-face. This is especially true for more junior employees, where the sophistication of what you do in head office and management will be different from what you do with employees, which will probably be a bit more basic.
Where does Lontra fit?
A generic assistant can answer a question or prepare a single conversation. Lontra is relevant when an organization needs continuity: structured conversations, useful context preserved over time, clear provenance, recurring patterns to investigate and preparation for the next authorized human action. Lontra can prepare, structure, remember, compare, flag and recommend. Humans retain listening, judgment, decisions, explanations and responsibility.