Conversations and speech interaction¶
HiveAI conversations bring a question, its available supporting information and the progress of related work into one place. They can help a person understand an accessible record, find approved guidance or prepare text for review. A saved question, a delivered answer, an approved proposal and a completed change are different outcomes. The conversation should make those differences clear.
Availability: Conversation services and saved-request recovery are implemented in the backend. The account product chooser still shows HiveAI as Not ready yet, and the inspected frontend has no standalone HiveAI application. The user guide describes the intended 06 Final Screens experience for a compatible client when delivered. Provider provisioning, browser operation and live delivery remain unverified. This is a draft for the initial reviewer, not a launch or acceptance record.
What a conversation contains¶
| Part | What it means |
|---|---|
| Your question | The instruction you choose to submit in the current organization |
| Saved request | A record that the instruction was accepted; work may still be waiting or running |
| Answer | Generated content that needs human review, with accessible supporting sources where deliverable |
| Progress | What has been saved about the request, including completed steps and outstanding work |
| Clarification | A specific question or current set of choices that must be answered before continuing |
| Proposed change | A separate decision requiring review and the necessary authority |
| Required step | A protected action outside the conversation that must be completed before the same request continues |
| Costs for a part | Usage associated with that part of the request; an unknown amount remains unknown |
Read the detailed conversation guide for entry, estimates, sending, history, choices, cancellation and recovery. HiveAI access explains the account and organization prerequisites.
Value for people and organizations¶
An employee can ask a focused question about information they can already access. A manager can review the evidence and the state of a request before relying on its result. An executive can use the same distinctions when evaluating adoption: how often work is accepted, how often useful answers are delivered, what needs human review, and what remains unresolved. This feature does not establish measured time savings, accuracy or a financial return.
Governance remains with the organization. HiveAI access does not grant access to another person's conversations or to every source product. A generated draft is not a sent message or an approved business record. Preparing a proposal does not execute it. Approving a proposal does not itself prove that the source system completed the change.
Access, context and history¶
Conversations are scoped to their owner and organization. The intended sidebar groups recent conversations; on a phone, the same navigation is behind Menu. The current service returns a bounded recent list, not an unlimited searchable history archive. Available messages can change when source access or retention changes.
A conversation opened from a record carries that context, which is checked again for later questions. It does not bypass the record's access rules. The model also receives a bounded recent history, so a long conversation is not a promise that every earlier detail will be remembered. Include the essential context in a focused question and verify the answer against its sources.
The current governed delivery path requires usable citations. A final-screen example of a completed answer without sources is still a design reconciliation item; it must not be read as a promise of unrestricted, source-free chat.
Costs and decisions¶
The optional estimate is approximate and nonbinding. It does not reserve credits, set the final price or act as the spending limit. Conversation history, supporting information and the work actually performed can affect usage. Editing a question makes its earlier estimate stale.
Work uses the organization's governed credit controls. Each deliberate continuation can start another separately accounted part. Costs for part 2 is not the total for the entire conversation. Released credits are unused reserved credit returned to availability; the label does not promise a cash refund.
An interrupted response is not necessarily free. If a provider result cannot be recovered, the implemented recovery policy can settle the approved reserved maximum once and withhold the unavailable answer, without repeating that provider work. Check the saved result and ask support about an unresolved charge before submitting the same work again.
Speech belongs to the same governed request¶
The final designs include microphone input, transcript review, spoken clarification and spoken answers. Speech does not provide extra access or remove the need to review a proposed action. The speech guide explains recording, transcript correction and deliberate submission, then separately covers spoken answers, playback, charges and recovery. A working microphone/player journey remains unverified.
Data, retention and exports¶
The data lifecycle guide explains what remains after audio expires, what an authorized export includes, how member removal differs from hold minimization, and why cleanup and charging can finish separately. Organization deletion requires coordinated evidence across the responsible services; a HiveAI result covers its own contribution.
Limits to account for¶
- The intended client journey is ahead of the current frontend implementation.
- Saved confirms acceptance; Completed describes the saved request outcome. Neither label alone proves an unrelated business action or a current source is accessible.
- Partial answers and partial completion need explicit review. Stopping remaining work does not undo completed changes.
- Reconnecting reads the original saved request. Creating another question can create new work and new usage.
- Changed access, expired choices, missing history and uncertain costs need different recovery paths.
- Current-source refresh, cross-product follow-up, live verification and audience walkthroughs remain open.
The current assistant implementation does not yet provide every connected read or preserve every detail returned by a source. A list can be unavailable or incomplete even when records exist; some workflow lists that exceed one returned page cannot yet be delivered through the assistant. An unavailable answer does not mean there are no matching records.
Before relying on a total, due date, form field or available action, check the current source page you are authorized to use. A label or link in an answer does not prove that all supporting details arrived or that a protected action completed. Keep the original saved request when asking for help, and check an uncertain outcome before repeating work that may incur usage.
Engineering, design and operational references are delivered through the private GitHub documentation repository.
Check the evidence behind an answer¶
Use the source-inspection guide to understand citation counts, freshness, incomplete evidence and unavailable sources. The published-knowledge reading guide distinguishes the original cited version from the current authorized document and explains viewing limits.
Preferences across conversations¶
The personal-memory guide explains how the intended Save memory flow differs from ordinary conversation text. Active confirmed preferences can be selected within limits for later requests; pausing or deleting them does not rewrite earlier conversations or recall an in-flight request.
Questions about today's work¶
Use the My Day task guide for the intended question-and-response journey. These checks read current work within limits; they do not accept or complete tasks, send correspondence or change a record. Verify the owning WorkOS item before acting and recover an uncertain prior action before repeating it.
Approve and follow a proposed change¶
Use the governed-action guide to review the exact target and effect, submit a typed or spoken decision, and recover the original operation when its result is uncertain. A task can change before the conversation finishes updating. A restoration is a separate proposal that needs its own review and confirmed outcome.
Platform operators use the separate staff voice and spoken-response capability. Staff authority, platform internal costs and staff continuation rules differ from this tenant experience.