Meet the AI Assistant That Actually Does the Work

Most AI features in work tools are chatbots with a good vocabulary. You ask a question, you get a paragraph, and then you go do the actual work yourself: create the tasks, post the update, send the invites. The AI advised; you clicked.

Openbook's AI assistant is built the other way around. It is a global assistant with around 25 tools that take real actions in your workspace — it creates tasks, posts updates, invites members, and more. You describe the outcome; it does the clicking. And because acting on your workspace is a bigger deal than chatting about it, the design is deliberately conservative where it counts: destructive actions require your confirmation before they happen, and the whole thing runs on a bring-your-own-key model, so you choose the AI provider and hold the account.

This guide covers what the assistant actually does, how to set it up, how the safety model works, the workflows where it earns its keep, and — just as important — its limits. No magic claims. Just what it is good for and how to use it well.

What "takes real actions" means

The difference between a chatbot and an agent is tools. A chatbot can only produce text. An agent has functions it can call — create this, update that, post there — and the judgment to chain them together from a plain-language request.

Openbook's assistant has roughly 25 such tools spanning the things you do in a workspace every day. The brief version: if it is routine workspace work — creating tasks, posting updates, inviting members, and the surrounding read-and-write operations that make those useful — the assistant can probably do it. Because the assistant is global, it works across your whole workspace rather than being trapped in one room. That is what makes multi-step requests possible. Consider a request like:

"Create tasks for the action items from yesterday's retro and post an update to the team feed saying they're on the board."

Fulfilling that means reading in one place, creating cards in another, and posting in a third. A per-feature chatbot cannot cross those lines. A global assistant with tools can — and this is precisely the kind of glue work that eats an hour of a team lead's day: not hard, not interesting, just spread across surfaces.

Three properties are worth internalizing before you start prompting:

• It acts, so treat requests like delegation, not search. The best prompts describe outcomes with enough specifics to act on: what, where, who, when.
• It is conversational, so you can iterate. If the first result is 80% right, say what to change rather than starting over.
• It is bounded by its tools. Around 25 tools cover a lot of workspace ground, but "AI assistant" does not mean "can do anything." The Limits section below is honest about the edges.

Setup: bring your own key

The assistant is BYOK — bring your own key. Instead of paying a markup for a bundled, black-box AI, you connect an API key from one of three supported providers:

Provider	Bring
Anthropic	An Anthropic API key (Claude models)
OpenAI	An OpenAI API key
DeepSeek	A DeepSeek API key

Why this model is worth the five extra minutes of setup:

• You choose the brain. Teams have real preferences about AI providers — capability, cost, data-handling policies, regional considerations. BYOK means that choice stays yours, and you can switch providers without switching workspace platforms.
• You see the real cost. Usage is billed by your provider at your provider's rates, on your own account, with your provider's own dashboards. No bundled "AI credits" abstraction between you and the meter.
• Your existing provider agreements apply. If your company already has terms, data-processing agreements, or usage policies with Anthropic or OpenAI, the assistant runs inside them, because it runs on your account.

Practical setup notes: the AI assistant is a Pro-plan feature (see pricing — Pro is $15 per user per month, $12 annual, with a 14-day free Pro trial, so you can evaluate the assistant on real work before paying anything). Getting a key from any of the three providers takes a few minutes on their platform, most teams set a spending cap on the provider side as basic hygiene, and the key should be handled by an Owner or Admin as workspace configuration — not pasted around in chat.

The safety model: confirmation before consequences

An assistant that can act on your workspace raises the obvious question: what stops it from acting badly? Openbook's answer is a confirmation gate on destructive actions.

Here is the shape of it. Actions divide into two kinds:

• Additive and routine actions — creating a task, drafting a post, that class of operation. These are low-stakes and reversible in the ordinary way (you can edit or remove what got created), so the assistant can carry them out as part of doing its job.
• Destructive actions — the ones you cannot easily take back. These require your explicit confirmation before they execute. The assistant tells you what it intends to do, and it happens only when you approve.

This split matters because it matches how good delegation works between humans. You do not ask a competent assistant to check with you before adding a row to a list — that would defeat the purpose. You absolutely expect them to check before deleting something. The confirmation gate encodes that norm.

Two habits make the safety model work even better in practice:

1. Read confirmations like you mean it. The gate only protects you if the approval step is a real decision, not a reflexive yes. When the assistant asks to confirm something destructive, that is the moment to slow down for five seconds.
2. Review what got created, at least early on. For your first couple of weeks, glance at the tasks and posts the assistant produces. You are calibrating — learning where it is reliably right and where your prompts need more detail. Most teams find the review habit fades naturally as trust builds, the same way it does with a new hire.

Real workflows worth automating

The assistant pays off on work that is frequent, multi-step, and boring. Here are the patterns teams reach for, from simplest to most compound. Each is built entirely on the assistant's documented action set — creating tasks, posting updates, inviting members, and their supporting operations.

1. Capture on the move. You leave a call with four commitments in your head. Instead of opening the Kanban Board, creating four cards, and setting owners one at a time: "Add four tasks to the sprint board: [the four things]. Assign the API ones to Dana, due Friday." Ten seconds of typing, and nothing evaporates between the call and the board.

2. The status update you keep postponing. Narrating work is high-value and chronically skipped. Delegate the posting: "Post an update to the project feed: beta shipped to the first ten customers Tuesday, two bugs found and fixed, on track for the full rollout on the 15th." You supply the facts; the assistant handles the posting. Pairs naturally with a Project Status room, where the update becomes part of the permanent history timeline.

3. Meeting outputs to board, in one pass. The classic failure of retros and planning meetings is that outputs stay in the notes. After a retro in the Retrospective room — which can already produce an AI summary of the meeting — hand the action items to the assistant: "Create a task for each action item from today's retro, assign the owners we agreed, and post the list to the team feed." Actions land on a board where they will be seen again, and the team sees it happened. (Why this loop matters: Action Items That Actually Get Done.)

4. Onboarding motions. New person joining? The assistant can invite members, which turns "I'll add them when I'm at my desk" into a sentence typed from anywhere: "Invite jordan@company.com to the workspace." Combine with a task: "...and add an onboarding task for me to set up their first-week check-in."

5. The Monday sweep. Compound requests are where a global assistant beats per-app AI decisively: "Create this week's tasks from our planning doc list, post the week's priorities to the feed, and add a task for me to review Friday's check-in digest." One instruction, three surfaces, zero tab-hopping.

The five patterns, reduced to a cheat sheet:

Chore	What you say	What the assistant does
Capture	The tasks, the board, owners, dates	Creates the cards
Narrate	The facts of the week, the feed	Posts the update
Meeting outputs	Action items, owners, destination board	Creates tasks, posts the list
Onboarding	The email address, any setup tasks	Invites the member, creates tasks
Weekly sweep	All of the above, in one message	Chains the actions

A prompt pattern that covers all five: outcome, location, specifics. What should exist afterward, in which room, with which owners and dates. The requests that disappoint are the vague ones — "organize my tasks" gives the assistant nothing to act on. The requests that delight are delegations you would happily give a sharp assistant on their third week.

The AI already inside your rooms

The global assistant is not the only AI in Openbook. Three room types have AI built into their own workflows, and they compose well with the assistant:

• Kanban per-board AI chat. Each kanban board has its own AI chat, scoped to that board — useful when the conversation is about this backlog, this sprint, these cards.
• Gantt AI plan generator. The Gantt room can generate a timeline plan with AI, which beats staring at an empty chart when you know the goal but have not structured the phases. Generate, then correct — editing a wrong-ish plan is faster than authoring from nothing.
• Retrospective AI summaries. The Retro room produces an AI summary of the session — the writeup nobody volunteers to do, done by default.

The division of labor: room AI helps within a surface; the global assistant moves work between surfaces. A sprint's rhythm might use all of them — the Gantt generator to rough out the release plan, the board chat while working the sprint, the retro summary at the end, and the global assistant to turn that summary's action items into next sprint's cards.

Search or delegate? The assistant and ⌘K

Openbook gives you two global ways to get things done from anywhere: the AI assistant and the ⌘K command palette. New users sometimes treat them as rivals. They are not — they split the work cleanly, and knowing the split saves time on both sides.

Reach for ⌘K when you know the destination. Jumping to a room, finding a doc by name, pulling up a teammate — the command palette is two keystrokes and instant. Using the assistant to "open the sprint board" is like asking a colleague to press a button for you; the palette is the button.

Reach for the assistant when you know the outcome. Anything that involves creating, posting, inviting, or chaining several steps is delegation, and delegation is the assistant's job. "Find the marketing space" is a ⌘K query. "Post this week's priorities to the marketing feed and add the launch tasks to their board" is an assistant request.

The gray zone is questions. "What's on the board for this sprint?" can be answered by either navigating (⌘K, then look) or asking. Early on, navigate — you will learn your workspace's shape faster. Once the shape is familiar, asking is often quicker, especially for anything you would otherwise have to compile from more than one place.

The rule of thumb that sticks: ⌘K for nouns, the assistant for verbs. Places and things: palette. Actions and outcomes: assistant.

A week with the assistant

To make the workflows concrete, here is what one team lead's week looks like with the assistant handling the glue. Nothing in it is aspirational; every step uses the create-tasks, post-updates, invite-members action set described above.

Monday. Planning happens against the board as usual. Afterward, the compound sweep: sprint tasks created from the agreed list with owners and due dates, the week's priorities posted to the feed, and a personal reminder task for Friday's check-in digest. Ten minutes of post-meeting admin becomes one instruction plus a quick review of what landed.

Tuesday. A customer call surfaces three commitments. From the hallway, before they evaporate: three tasks captured to the board, assigned, dated. The alternative timeline — "I'll add them after lunch" — is where commitments go to die.

Wednesday. A new contractor starts Monday. One sentence invites them to the workspace; a second creates the onboarding checklist task. When they accept, they are already in the people directory and visible to the team.

Thursday. Mid-sprint status ping from a stakeholder. Instead of a meeting, a dictated update posted to the Project Status room: what shipped, what slipped, what is next. Two minutes, written record, history timeline intact.

Friday. Retro. The room produces its AI summary; the assistant turns the action items into cards for next sprint and posts the summary link to the feed. The lead reads the check-in digest (that Monday reminder task fires), spots a yellow mood trend, and books a one-on-one — the thing no assistant can do.

Total assistant time across the week: perhaps five minutes of typing. Total clerical time it displaced: roughly two hours, most of it the kind that otherwise leaks into evenings. That is the realistic scale of the win — not a transformed job, but a reclaimed afternoon every week, compounding.

Limits: what to know before you rely on it

An honest tool guide includes the edges. Here are the assistant's, and how to work with them.

It is bounded by its tool set. Around 25 tools is substantial coverage of routine workspace operations, and it is not infinite. If you ask for something outside what its tools can touch, the right outcome is the assistant telling you rather than improvising. When you hit a boundary, do that step manually — the assistant is an accelerator for the workspace, not a replacement for it.

It is as good as your instructions. Underspecified requests produce plausible-but-wrong results: tasks on the wrong board, an update missing the key number. This is not a flaw to engineer around so much as a skill to build — the same briefing skill delegation to humans requires. Specifics in, quality out.

Model quality is your choice, literally. BYOK means the assistant's intelligence is whichever provider and model your key points at. If results feel weak, the first lever to check is on the provider side, not the workspace side.

Confirmation is friction, on purpose. Occasionally you will wish the destructive-action gate were not there — bulk cleanups feel slower with a checkpoint. That trade is correct. An agent that can delete without asking saves you seconds on the good days and costs you dearly on the one bad day. The gate stays.

It does not replace judgment or presence. The assistant can post your update; it cannot decide what the honest status is. It can create tasks from a retro; it cannot have the retro for you. Teams that thrive with it automate the clerical layer and reinvest the time in the human layer — the writing, deciding, and noticing that no tool does.

Rolling it out to a team

A few norms make the difference between one power user and a team-wide habit:

1. Start with one shared workflow, not a memo. Pick a single recurring chore — say, retro actions to board — and have the team lead visibly run it through the assistant for two weeks. Adoption follows demonstration.
2. Share prompts that worked. When someone finds a phrasing that reliably produces the right result, post it to the feed. A team prompt library beats individual trial and error, and a pinned post is library enough. A Q&A room works well for "how do I get the assistant to..." questions once usage spreads.
3. Agree on review norms early. Decide as a team what gets double-checked (anything client-facing, anything destructive — which the confirmation gate enforces anyway) and what does not (routine task creation). Explicit norms prevent both blind trust and pointless re-checking.
4. Route around it without shame. Some teammates will prefer clicking. Fine. The assistant is a faster path to the same rooms, not a mandate — the board does not care whether a human or a tool call created the card.

Next steps

• Trial it on real work. The 30-day Pro trial plus a provider key gets you a live assistant this afternoon. Give it one week of your team's actual glue work — captures, updates, meeting outputs — and judge it on time saved.
• Set up the basics first. The assistant is most useful when there are boards to create tasks on and feeds to post to. If your space is not composed yet, start with Getting Started with Openbook and Rooms Explained.
• Pick your provider deliberately. Anthropic, OpenAI, or DeepSeek — decide based on your company's existing agreements and preferences, set a spending cap, and hand the key to an Admin.
• Revisit your workflows monthly. The best assistant use cases are discovered, not planned. Ask the team what they delegated last month, promote the winners to shared norms, and keep the boring work moving without you.

The pitch for most AI features is that they will change everything. The pitch here is smaller and more useful: the assistant does the workspace chores you already know how to do but should not be spending your day doing. Delegation, with a confirmation step. That is the whole trick — and it is enough.