Use the AI Workspace

A question about the business often needs both an answer and a way to inspect its evidence. Open AI Workspace to ask, review record links and see the tools and results in run history. Use Ask AI when you want to stay on your current screen.

Ask a question the system can verify

Name the record or time period and say what you need. For example:

Find open opportunities with no next action. Show the company, owner and last recorded contact. Do not change any records.

Read the returned record links and dates. If the answer lacks a needed source, ask for that source or inspect the record yourself before acting. An incomplete answer should stay incomplete until the missing information is available.

Request a change

Ask for one bounded action, such as preparing a reply to a named conversation. Choose Review beneath the answer to inspect the exact recipient, content and changes in chat. Approve or reject there, or use Approvals. After approval, inspect the execution result; approval alone does not establish that the operation succeeded.

Set AI costs and recurring work

Open AI Workspace → Cost controls as a workspace administrator. Model requests and recurring coworker work start paused. Set a daily budget, monthly budget, maximum cost per request and maximum requests per day, then turn on the work you want and choose Save AI controls. A zero dollar budget permits only a model with verified zero pricing; requests still count toward the daily request limit.

For example, a fictional service team might allow 20 requests a day, reserve up to $0.10 per request and set a $1 daily and $10 monthly budget. Keep recurring work off if the team wants AI only when someone asks for it. Individual coworkers lists each registered coworker and its kinds of work. Pause an individual coworker to prevent its queued execution, or choose Pause all AI now to stop new model requests throughout the workspace. A request already running may finish.

Each provider attempt reserves its bounded token cost before it runs. A confirmed final charge releases the unused part of that reservation. An interrupted request without a final charge keeps its reservation until the period resets, preventing an uncertain outcome from spending the same budget again. Daily periods reset at midnight UTC; monthly periods reset on the first day of the month.

Current usage separates budget used or held from provider charges reported for this workspace. Recent requests identify their job, model, reservation and reported charge. These records begin when cost controls are installed. They do not include another application's usage, credit purchase fees or hosting. Provider credits and automatic top-up remain in the provider's billing settings; saving these controls never purchases credits. Health monitoring continues without model calls when recurring coworker work is paused. Paused work is not marked overdue; previously failed or stalled runs remain visible. If recurring work is on while model requests are paused, supported coworkers use their existing checks on saved records.

If another administrator saves first, refresh the controls and review their current values before saving again. If controls or current prices cannot be read, the app refuses the model request. The fictional demo saves these choices only in its browser session and cannot start real scheduled work or purchases.

Inspect a run

Use run history to find which tools were called, their outcomes and reported usage. This helps distinguish a missing connection, a denied operation and a provider failure.

If a run stops after an external request, inspect its action result before retrying. An uncertain result needs reconciliation to avoid a second send.

When a tool is unavailable

Check its connection requirement in the AI tool reference, then inspect Setup; module settings and your own permissions can restrict access too, independently of whether the connection itself is working, and enabling a connection never grants you additional authority on its own.

Model jobs and evaluation

AI requests name a registered job with capability and context requirements. An operator-selected model must satisfy those requirements. Cost controls pin each request to that model and its verified price ceiling. Free and low-cost models need a recorded passing evaluation before running consequential jobs such as the command agent. A fresh installation does not assume that evaluation has happened.

The model-call activity records the requested model, reported serving model and execution state. A started call without a terminal receipt needs investigation; it is not proof of a completed answer. Model availability and evaluation are separate from permission to change business records or send messages, which still follows the approval workflow.

Example: turn a review into an approved next action

Maya at the fictional Northline business wants to identify an opportunity that needs attention.

1. Ask: “Show open opportunities with no next action. Include record links and the date of the last recorded contact. Do not change anything.”
2. Open a returned opportunity and compare the answer with its saved history. If two records could match a name, resolve the identity before requesting a change.
3. Ask for a specific supported change to the chosen record. For example: “Prepare a next-action update for this opportunity: Maya will call on Tuesday, 15 September 2026. Show me the proposal before applying it.”
4. Open Review in the conversation. Confirm the opportunity and date, then approve only if the change is correct.
5. Inspect the execution result and reopen the opportunity. A proposed change or an approval is not yet proof that the update saved.

The example's names and dates are fictional; use your actual record and intended date. Available tools depend on the enabled modules, your access and the configured model. If the operation is unavailable, use the supported admin control or ask your administrator about the missing capability.

If the run stops or access changes

A denied operation must stay denied; retrying does not grant access. Ask the administrator to check your current membership and record permissions. If the model or provider times out, inspect the run and any staged approval before asking it to repeat the work. A stopped run may already have prepared a proposal. An uncertain external action needs receipt review before another attempt, especially for email or invoicing.

Tool discovery can find operations outside the screen you started from. Loading a tool group does not grant permission to execute it. Groups activated during one command run are not automatically remembered in a later run; ask for the task again when needed and review the newly prepared change.

Architect and Blueprints

Choose Architect to describe how your business works and supply the sources needed for one workflow. This opens the Architect conversation in the AI Workspace. Use Help beside the heading for the next steps and this guide. On a phone, scroll inside longer guidance; press Escape or select outside the panel to close it.

Choose Blueprints to review saved proposals, their versions and validation results. Compare the proposed changes with your actual business before approving an exact version. Approval and application are separate steps; a saved proposal does not mean the workspace has changed.

Review a Blueprint in Architect

In Architect, expand Blueprint review and choose the draft you want to change. Enter a proposal such as “Set the business summary to Require a written scope before scheduling work,” then give the change a short summary. Select Preview patch to compare the current and proposed values. Review draft save opens the exact approval. Approving saves a new draft version; it does not apply the workspace.

Select Simulate to inspect workflow steps and capability checks. Leave the event blank to inspect all workflows, or enter a trigger to narrow the result. Simulation never sends messages, enables modules or applies the Blueprint.

If another administrator saves first, reload the current draft and prepare a new preview. After an uncertain save, inspect the original approval and receipt before creating another request. Each saved version keeps its history and requires a new approval before the workspace can apply it.

Save the operating setup

A new workflow often needs the same decisions repeated across several screens. The Blueprint review collects those decisions in one proposal, so you can keep available board columns and the remaining implementation work together.

1. Review the current version

   Open Blueprints, choose the saved proposal and read Operating setup. Each board identifies its existing destination, while workflows and Coworkers name the capabilities they require. Resolve business assumptions with your team.

2. Approve the proposal

   Choose Approve for the version you reviewed. If you edit the business summary, save it as a new version and review that version before approving again.

3. Save operating setup

   Choose Save operating setup to add available columns to the existing Pipeline or Content board. The saved result keeps navigation, views, workflows and Coworkers as recommendations for review, alongside any custom development briefs.

4. Check the saved result

   Read the number of columns added, then reload the review and check Setup saved with an audit record. Repeating the save returns the same version's result, including when its columns already existed.

For example, the fictional manufacturing Blueprint reuses Pipeline for sales, while its production-order lifecycle becomes a Custom App Brief, a description of the extension still needed. Give that brief and the extension guide to Claude Code, Codex or your team's coding agent, so it can build on the existing records and permissions through the repository's development and review workflow.

What to check

• If Operating setup could not be confirmed appears, reload the review to check its saved result or retry the same save. A dropped connection can hide a completed save; retrying returns that version's original result. A rejected database transaction rolls its columns, saved result and audit back together.
• If an earlier save needs maintainer reconciliation, ask the installation maintainer to inspect its saved result and audit. Keep the existing columns while the maintainer establishes what happened; that version stays unavailable for another setup save.
• Read navigation, view, workflow and Coworker entries as recommendations. Review and implement each supported change through its existing configuration or extension path. An email workflow still needs a configured provider and its required action approval.
• If setup is unavailable after an upgrade, ask the maintainer to apply the release's migrations and run schema verification before retrying.