Assistant

Ask a question you can verify

Choose a task first. Use an account with access to that feature in the active organisation; each linked procedure lists its prerequisites.

1. Start with a customer question
2. Add the relevant sources
3. Review proposed changes

Check the result: Read the answer, open its cited records, and review any proposed change separately. An answer is not proof that a record changed.

If you get stuck: If the answer lacks evidence, name the customer, add the source and ask for the missing detail. Check the proposal state before deciding whether to confirm it. AI requests may use billable resources.

Start a conversation and ask a useful question

Screen: /assistant

Use case: Ask for a customer summary, priorities, a plan or a draft.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Choose AI Assistant → New chat. Select Create a customer from notes, Rank key accounts, Plan today’s follow-ups, Find a customer, Draft a follow-up task or Draft a greeting, or type your own request.
2. Name the customer and desired outcome. Add dates, constraints and source material; answer clarification questions instead of assuming the assistant knows missing facts.
3. Press Enter or Send to submit; Shift+Enter inserts a line break. While a response is streaming, Stop ends generation.
4. Read the response, citations and any proposed action. Open referenced records to check the underlying evidence before relying on a conclusion.
5. From any other page, click the round Assistant launcher fixed in the bottom-right corner instead. It opens the Assistant window in place, with quick actions Today, Log a contact and Find a customer, so you never lose the page you were on.

Expected result: You receive an answer or a request for missing information; proposed changes still require their own review.

While a modal dialog or side sheet is open, the launcher window stands down entirely. It comes back with the same conversation and the same half-typed message once the dialog or sheet closes.

Carry the right context into work

Start a new chat for a separate customer or task. For continued work, reopen the existing conversation and check its current customer, attachment and pending proposal before sending another request. A half-typed composer can survive page navigation and a switch to the launcher; review it before pressing Send. Assistant answers and tool results may point to records or create prepared work. Open the linked record or Workbench item to verify the result rather than treating a suggested action as completed.

Create a personal assistant on desktop

Internal desktop · Staging

Use a signed-in internal desktop build connected to staging. Personal assistants appear under My assistants; the web AI-team hiring workflow is described separately below.

1. Open Assistants and select + beside My assistants.
2. Enter a name, choose an avatar and select Create.
3. If the name does not identify one clear job, enter the short job description requested and select Create again. A clear job name skips this question.
4. Review the proposed name and job focus, then select Confirm creation.
5. Wait for the saved assistant to appear in My assistants. Selecting it opens its own conversation.

Atlas selects existing enabled organization skills, connected plugin tools, built-in tools/functions and authorized knowledge for each task. Published organization guidance and your active user memory resolve under your existing access. You do not manually assign capabilities, connect providers or grant new permissions when creating an assistant.

If the review result is uncertain, keep the draft and use Recover review to check it before continuing.

Inspect assistant properties on desktop

Developer view · Internal desktop on staging

Properties is available to signed-in developers whose verified email domain is exactly forth.ai, forthai.work or forthai.cn. It covers Head and your own named assistants in the current organization.

1. Right-click an assistant in the list, or focus its button and press Shift+F10.
2. Select Properties to view five read-only sections: Model, Instructions / system prompt, Skills, Tools / functions and Knowledge base.
3. Read the source and availability shown for each section. Available skills, tools and knowledge describe configuration; the task determines what is selected or used. An unavailable section does not mean that nothing is configured.
4. Press Escape to close the view and return focus to the assistant.

Head uses its local runtime configuration. When no model is pinned, the view says No model pinned; resolved by the local runtime. It does not invent a model name or start the engine to inspect it.

Configured instructions are shown without private conversation, memory or retrieved content. Inspection does not execute tools, test provider connections, change settings or grant access.

Attachments and customer context

MB and GB use binary units here. Exactly 50 MiB keeps existing automatic behavior and reader limits. Large nonvideo originals stay available even when processing is unavailable: PDF readers allow 25 MiB, image readers 8 MiB, Office/XLSX readers 32 MiB, and customer proposal reading 10 MiB. There is no enabled Process action for unsupported large files. Library storage remains limited to 2 GiB per organisation; existing permissions and billing rules apply.

Screen: /assistant

Use case: Give the assistant the material and customer scope it needs.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Use Attach a file beside the composer. Supported original file drops can be up to 1 GB (1 GiB; 1,073,741,824 bytes) per file, inclusive. Existing limits still apply to small inline intake and other attachment surfaces.
2. Select supported text, PDF, image (JPG/PNG/WebP), DOCX, XLS/XLSX or CSV files. Wait for the upload result. Files larger than 50 MB (50 MiB; 52,428,800 bytes) remain in My Docs without automatic processing or review jobs; the assistant cannot read their contents merely because they were uploaded.
3. Mention or pin the intended customer when the conversation offers a customer picker. Check the context chip; remove or change it when moving to another customer.
4. Add a specific question and send. If an intake produces customer drafts, open the proposed draft, inspect extracted fields and sources, then explicitly confirm creation.

Expected result: The message uses the intended attachments and customer context.

Attachment limits are surface-specific. The older inline-intake hint describes its own temporary uploads; completed originals in My Docs retain their normal document lifecycle. Uploading alone is not customer creation.

Resume, revise and delete conversations

Screen: /assistant

Use case: Continue earlier work without losing its context.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Choose an existing conversation from history. Read the recent messages and any pending proposal before continuing.
2. To correct a request, use the edit action on your earlier message and resend the revised prompt. Check the new response instead of assuming it matches the old one.
3. To start independent work choose New chat. Use the conversation actions menu and Delete only for a conversation you intend to remove; read and confirm the dialog.
4. If a draft is pending, resume it through the offered draft link and review saved fields before committing.

Expected result: You continue or remove the intended conversation deliberately.

Unsent text in the composer is preserved across navigation: leave the page mid-draft and return, or open the launcher instead, and the same half-typed message is still there. Editing, deletion and draft commits are source-confirmed; they were not executed against staging conversations in this study.

Review action proposals and contextual work

Screen: /assistant

Use case: Turn an AI suggestion into a deliberate business action.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Open the proposed action and inspect the customer, values, source evidence and expected effect. A generated draft has not automatically been sent.
2. Edit an offered draft or reject it when facts are wrong. Use the confirmation control only after checking the final values; review the recorded result or failure.
3. In customer context, Attach file can include a note and deal. Log email records counterparty, direction, subject, body and optional deal; it is a record of an email.
4. Log meeting records title, type, date/time, duration, attendees and optional deal. Complete the appropriate form and save only the real event you intend to record.

Expected result: You can distinguish a proposal, a saved record and an actual external send.

A consequential change — sending, saving a record or committing an edit — always needs its own separate confirmation step; opening or reading a proposal never performs it by itself. Data-changing and outbound actions were not executed during this study.

Review a proposal and its handoff

1. Read the proposed action, target customer, cited source and exact fields. If the Assistant asks for missing context, supply it or stop; an answer is not an approval.
2. Open the linked draft or task in Workbench when offered. Review recipient, sender, body and cited facts there. An AI draft remains prepared work until a person confirms a send or a record change.
3. After approval, check the visible result or error and the destination record. If the result is uncertain, inspect Workbench or Sent before repeating an outbound action.

The Assistant’s conversation and Workbench task have different roles: chat preserves your working context; the task carries the saved draft, review controls and outcome. Source review of staging ef39dda8b did not execute a tool, approval or send.

Find your AI team and sales assistants

Screen: /superpower/ai-teams

Use case: Find an agent and understand who it works for.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Open Superpower → AI Teams (badged Preview / 内测). Old Assistant → AI team links open the same page. Search by agent or customer, and filter All, Draft, Active, Paused or Retired; Group by customer changes the listing layout.
2. Read the roster's Performance and Run health tabs. Run health states, for a trailing 7-day window, how many runs finished, how many produced no written review, and how many failed — a run still queued or running is excluded, and a trial run is never counted.
3. Read the Sales Assistant and Business Analyst cards at the foot of the roster. Sales Assistant reports how many rep assistants are already set up for the organisation; Business Analyst is a single organisation-wide hire that only an owner or administrator can make.
4. Open an available assistant card to inspect its supervisor, customers owned, schedule and last-run outcome. It can read only its supervisor's accessible customer scope.
5. For a customer-specific account manager, choose Hire an agent and follow the setup section below.

Expected result: You understand each assistant’s owner, scope, schedule and current state.

A degraded run still completes and keeps the records it gathered, but produces no written review and no suggested actions — open the hire's work history to read it. AI team stays labelled Preview / 内测 in this build; visibility and account-manager setup still depend on release controls and permissions.

Hire and configure an account manager

Screen: /superpower/ai-teams/new

Use case: Set a bounded assignment before activating an agent.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Choose Hire an agent → Role → Continue. In Assignment select the customer and supervisor, and enter an agent name (up to 80 characters).
2. In Brief enter goals and instructions, choose the agent's skills and knowledge, then select English or Chinese. Configure time, IANA timezone and weekdays; explicitly confirm the timezone.
3. Set monthly credits and per-run credits. Both are positive whole numbers; per-run cannot exceed monthly, and monthly cannot exceed 1,000,000.
4. Run the required Preview, inspect the proposed work and evidence, and correct the brief if necessary. Activate only after the preview is acceptable.
5. When editing an existing hire, retain its customer assignment; use Configure for the editable fields. Changing any configuration on a portfolio Sales Assistant returns it to Draft status until a new preview succeeds.

Expected result: The agent is configured for one customer with explicit schedule, budget and reviewed preview.

Hiring an agent, running its required preview and every scheduled or on-demand run afterwards consume credits from the organisation's monthly allowance — the same allowance set here as monthly credits and per-run credits. Role and Assignment forms were inspected; later setup and activation steps were checked in deployed source without creating an agent. Goals and instructions follow the limits shown by their fields.

Agent lifecycle, work history and skills

Screen: /superpower/ai-teams

Use case: Supervise an agent after setup.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Open the agent. Use Overview to check configuration and current status; Configure, Preview and Activate appear only when allowed.
2. Use Pause to stop scheduled work, Resume to restart it, or Retire to end the assignment when appropriate. Run now starts real work immediately and consumes credits from the organisation's allowance, the same as a scheduled run.
3. Open Work history, choose a run, and inspect results, evidence and failure reasons. Review prepared work before accepting any business action. Check the roster's Run health tab first for the finished, no-review and failed counts over the last 7 days before treating a queue as clean.
4. Open Skills to inspect the agent’s skills and versions. Where teaching is offered, review evaluation evidence before publishing or rejecting a learned version.
5. Scheduled optimization drafts, if shown, are suggestions drafted from this account's run statistics only (counts and error classes — never customer content). Adopting one opens a teaching proposal prefilled from the draft; nothing changes for runs until that proposal passes evaluation and publication.
6. Teaching proposals can name where a lesson comes from: teaching.create accepts an optional feedback anchor ({"source":"knowledge_sandbox"|"agent_run","id"}) that Atlas verifies against the Sandbox rating or agent-run rating it names — a missing, withdrawn or cross-tenant anchor refuses and saves nothing, and a rating alone never creates a proposal.
7. To stop using a published lesson, open Skills and choose Withdraw this version on the active version (administrators only). The next run, draft and chat use the base skills only; the version stays in the history marked Withdrawn and cannot be restored. To change the guidance again, publish a replacement through teaching, which is evaluated before it goes live. If someone else changed the version first, Atlas refuses and nothing changes; refresh and try again.
8. To see which lessons shaped a result, open the run on Settings → Traces (administrators and managers) or choose Review on an AI-team item in Today. Guidance used lists the learned guidance the run followed and the reviewed lessons it recalled, each with its version, scope and why it applied; guidance withdrawn later is marked Withdrawn. A lesson’s title appears only if you can already read that lesson — otherwise it says Not available to you — and recalled lessons are advisory: customer facts, permissions and company policy still win.

Expected result: Agent status changes and work decisions remain inspectable.

Source-confirmed detail flow; the inspected production team was empty, so no lifecycle or teaching action was executed.

Personal and team retrospectives

Screen: /assistant/retrospectives

Use case: Review a completed week or month and decide what to improve.

Before you begin: Use an account with access to this feature in the correct organisation.

How to do it

1. Choose More tools → Retrospectives, then Week or Month. Open a dated personal period; confirm the period before comparing results.
2. Read recorded touches, stage changes, decided suggestions, completed/overdue follow-ups and recommendations. Each period shows when it was prepared and whether you have acknowledged it yet, for example Not acknowledged. Open evidence where available.
3. Use Accept for a recommendation only after reviewing its effect. Acknowledge, add a comment or explain why it does not apply where those controls are offered.
4. Managers with Team access can inspect representative performance, ageing or at-risk deals and conversions. Review proposed standard changes before Publish or Reject; select the sharing scope for manager notes.

Expected result: You review a dated evidence set rather than treating missing activity as proof of no work.

Lists and periods were inspected. Data-changing acknowledgement, recommendations and standard publication were not executed.

Ask about current policies and procedures

Screen: /assistant

Use case: Answer a guidance question from current published originals and inspect conflicting statements with their sources.

Before you begin: Sign in as an ordinary member with knowledge read access. Relevant policies or procedures must be published and eligible for answers. Use a text-only question.

These are the same published originals as Knowledge → Policies (/knowledge/policies, /knowledge/workflows). In Assistant, open Sources beside the composer and select Policies, Procedures or both. Both start unchecked. Enter one text question and send; the selected sources apply only to that question and reset after dispatch or changing conversation. File attachments and starter tasks use ordinary chat; finish or remove them before selecting operational sources.

1. Choose the sources explicitly for the question. Ordinary chat stays the default.
2. Wait for the final check. The answer shows complete original passages, policy headings or whole workflow conditions, document identity, release/version and occurrence, checked time and bounded coverage. Original languages are preserved.
3. Read both linked originals when conflicting statements are listed. Atlas does not invent precedence or silently remove a counterpart.
4. Select Inspect sources. Atlas reruns the same question and source scope and opens details only when all returned evidence and conflicts still match. Changes, lost access or unavailable verification clear the old evidence and require Retry.

Coverage describes the 25-release scan cap per operational kind, oversized snapshots/passages excluded, and returned limits (8 operational passages, 5 registered document matches, 5 Library matches). Reaching a limit can omit relevant material. An Unknown result is not a compliance verification. Oversized or unresolved evidence is not shown as a partial answer.

Retry and Regenerate run fresh checks. Editing a question retains its chosen source kinds. Branch carries question/scope/time markers only. Reopened conversations require revalidation: original guidance, citations and model-only text are neither saved in browser archives nor included in later chat history. Copy originals copies only a currently validated in-memory answer when clicked. Operational replies have no rating controls and never initiate actions.

Sources, inspection and recovery controls support keyboard access in English and Chinese. Escape closes the source details and restores focus. Source text is displayed as plain escaped text; Library metadata does not assert an immutable document version.

Expected result: originals, conflicts and a checked time appear only after the final authority and freshness checks pass. Failed checks and reopened history show a recovery state without old evidence.

Ask the Assistant how to use Atlas

Screen: /assistant

Use case: Find out how to do something in Atlas, or where a feature is, and open the exact User Manual section behind the answer.

Before you begin: Sign in with any account. Ask in the language you want the answer in; English questions link to this manual and Chinese questions link to the Chinese manual.

1. Open the Assistant, in the chat window or on the Assistant page, and type a how-to question such as "How do I create a customer?" or "Where do I upload a document?". Send it.
2. Read the answer. Each instruction ends with a small raised number such as ¹. The list at the end of the answer names the manual section each number points to.
3. Select a number to open that manual section in a new browser tab. Your conversation stays open in the original tab.
4. If the manual does not cover your question, the Assistant says it is not sure and does not guess steps. It offers to prepare a report; see Report an issue through the Assistant.

The Assistant answers how-to questions only from this manual and links only to sections that exist in it. It does not change any record when you ask how to do something.

Expected result: a short step-by-step answer with numbered references, and each reference opens the matching manual section. When the manual has no answer, you see a plain "not sure" reply and an offer to report it.

Report an issue through the Assistant

Screen: /assistant and /feedback

Use case: Tell the Atlas team about a problem, or ask for a feature, by letting the Assistant fill in the report form for you to review.

Before you begin: Sign in with any account. Know roughly what went wrong or what you want. You do not need to describe it perfectly: you can edit everything before sending.

1. In the Assistant, say what you want to report, for example "Report an issue: the export button does nothing", or answer yes when the Assistant offers to report something it could not answer.
2. The Send feedback form opens with the type, severity and message already filled in, and the cursor in the message box.
3. Check each field. Change the type or severity, edit the message, and attach screenshots or files if they help.
4. Press Send when you are happy with the report. Nothing is sent until you press it, and the Assistant never sends a report for you.
5. To follow the report afterwards, open your account menu, then Support, then My issues. It appears there like any report you wrote by hand, with its status and replies.

Expected result: the prefilled form waits for you; after you press Send, the confirmation appears and the report is listed under My issues. If you close the form without sending, no report is created.