A workforce conversation can do more than collect a form. It can help someone explain their experience, clarify a requirement or find the next step in an unfamiliar process. Large language models make that exchange more flexible. The useful result is an answer, record or action that the team can understand and use.
What the LLM adds to EVA Bot
EVA's April 2023 announcement introduced LLM-assisted question answering alongside its existing talent conversations. The announcement covered recruitment questions, employee information and internal opportunities. The underlying aim was practical: let people ask questions in their own words and respond in the context of the organisation.
A language model can interpret differently phrased requests, produce a readable explanation and ask a follow-up question. Those abilities help when a participant describes work in a way that does not exactly match a form label. They do not, by themselves, establish whether an answer is current, whether a reported skill is verified or whether the requested action is permitted. Those questions belong in the surrounding process.
Connect an answer to the right company information
Consider an employee asking how to apply for a temporary assignment. A useful answer depends on the current process, the employee's location and the assignment involved. General model knowledge cannot settle those details. Agree which documents and records the conversation may use, who maintains them and how the answer should identify its source.
Retrieval can bring relevant passages from an approved knowledge collection into the context used to prepare an answer. If the material is missing, conflicting or out of date, the process needs a way to say so and route the question to an owner. A fluent explanation should never silently resolve an ambiguity that the source material leaves open.
Turn the conversation into evidence people can review
A participant might say, ‘I coordinated a regional rollout and can take on another project in November.’ That sentence contains several different pieces of information: reported experience, an availability date and an expression of interest. Keep their meanings separate when updating a record. Ask which part of the rollout the person owned and what capacity they can offer.
Store the original response, its date and the relevant follow-up alongside the structured fields. An employee's own assessment and a manager's assessment may differ without either being discarded. Clients can configure who may view individual assessments, summaries and exports according to their organisation and use case. A complete-looking profile should not hide missing or unreviewed evidence.
Give participant conversations and SuperBot different jobs
Participant-facing EVA Bot conversations help employees, candidates and other invited people provide information or navigate an agreed process. EVA SuperBot supports authorised teams working with configured administrative operations. A request made in a participant conversation does not grant the administrative permissions needed to change campaigns, access another person's assessment or run a workflow.
Customer Success and Business Analysis colleagues work alongside client teams to define questions, records, permissions and the next action. For multinational programmes, that includes testing how shared definitions work across local terminology and languages. Start with one real task, examine the records it produces and agree how exceptions reach the people responsible for resolving them.
Connect engagement to the decision that follows
The conversation is an entry point into EVA's five modules. Competency Mapping gives teams a consistent structure for requirements and evidence. Team & Leadership Analytics helps them examine patterns and differences in assessment. Workforce Planning combines demand, capability and usable capacity. Talent Matching compares people with roles, projects or assignments. The Full Automated & Approvals Platform carries out configured workflows, with approval steps where the client requires them.
Choose the relevant starting point for the work. A matching or planning discussion does not have to begin with a new competency-mapping project. Define what the conversation should produce, which module will use it and what completion means. An answer about how to apply is different from a submitted application; a recorded preference is different from an approved assignment.
What should happen after this message?
Illustrative workforce conversations. Choose a message to see the source, permission and next step it needs.
‘How do I apply for a cross-border assignment?’
How to read it
The answer needs the client's current assignment process and the participant's context. An explanation does not submit an application.
The next useful action
Retrieve the approved process, identify its source and explain the next step. Route unclear eligibility or policy questions to the named owner.
Compare all 3 cases
| Case | Known information | Next action |
|---|---|---|
| Answer a question | ‘How do I apply for a cross-border assignment?’ | Retrieve the approved process, identify its source and explain the next step. Route unclear eligibility or policy questions to the named owner. |
| Update a profile | ‘I led the rollout in France and Belgium last year.’ | Ask what the person owned, retain the dated response and propose the relevant fields for the configured review process. |
| Start an action | ‘Put me forward for the November assignment.’ | Confirm the assignment and required information. Execute only the configured, permitted action and report whether it was submitted, queued for approval or needs further input. |
Illustrative example. These controls explain the method; they do not access client records or execute workforce actions.
Choose one conversation your team handles repeatedly. Define the source, the record it should leave behind and the action it is allowed to start.
Explore the relevant EVA capabilitiesSources & further reading
Historical context comes from EVA's 4 April 2023 LLM announcement. The practical guidance and examples were refreshed on 15 September 2026. The links below describe current product scope and operating principles.