The AI Layer Is Here: Your Entire Agency, Understood by One Model

Every AI product shipped in the last two years has made the same mistake. A project management tool adds AI that knows about tasks. A chat tool adds AI that knows about messages. A CRM adds AI that knows about contacts. Each one is smart about its own slice of your work — and completely blind to everything else.

When you ask Slack's AI why a client went quiet, it cannot tell you — because the tasks, files, and meeting history that would answer the question live somewhere else. When you ask Asana's AI to help you prepare for tomorrow's call, it cannot — because the client conversation and the CRM notes are in different tools.

This is the problem we set out to solve before writing a single line of AI code. The result is the Kobin AI layer — and it is unlike anything else in the agency software category.

The architecture decision that changes everything

We did not add AI features to existing modules. We built a model that has access to the entire workspace data model before it responds to anything.

Every time you interact with Kobin AI — whether through @AI in the inbox or the global command bar — the model assembles a live briefing from your workspace. It pulls your active tasks, your project statuses, your upcoming calendar events, your CRM pipeline, your vault files, your recent team messages, and your client history. All of it lands in context before your question is processed.

Then, instead of guessing, the model uses tools. It calls get_tasks to fetch exactly the tasks it needs. It calls get_team_workload before suggesting who to assign something to. It calls search_contacts to get a full profile before answering a question about a specific person. This is not a chatbot that makes things up — it is a model with a live, structured read on your entire operation.

@AI in the Inbox

The first way to access the AI is the most natural: type @AI in any room. Project channels, group chats, DMs — it works everywhere. The AI responds inline as a new message in the thread, clearly labelled with a purple avatar and an "AI · Command Center" tag.

What makes this different from every other chat AI is what happens before the model responds. When you type @AI where does this project stand? inside a project room, Kobin automatically injects the full project context: every active task, their statuses and priorities, recent messages from the thread, any vault files linked to the project, and the project's timeline. The model does not need to be told any of this. It already knows.

Here is what you can ask, and what the AI does with it:

Responses stream in real time. There is a blinking cursor while the model is thinking. And unlike other inline AI tools, the model's responses are saved as messages in the thread — so your whole team can see what was asked and what was answered.

The Global AI Command Bar (⌘K)

The second entry point is the command bar: press ⌘K (or Ctrl+K) from anywhere in Kobin to open a floating AI panel. Unlike @AI in rooms, which is scoped to a single project or conversation, the command bar has full workspace access — every project, every client, every task across all time horizons, the complete CRM, the entire calendar, all vault files.

The command bar also keeps a persistent chat history. Every conversation is saved, titled, and listed in the sidebar. You can pick up where you left off, review what the AI told you last week, or delete sessions you no longer need. It is a proper interface, not a modal that vanishes when you close it.

Here is what you can ask the command bar that you cannot reasonably ask a per-module AI:

The AI can act, not just answer

Most AI tools respond. Kobin AI responds and executes.

When you ask it to create a task, it creates the task. When you ask it to assign something to the least busy team member, it checks workload, resolves the name, and assigns it. When you ask it to attach vault files to a task, it fuzzy-matches the file titles, resolves which vault folder they live in, and attaches them — all in a single action, confirmed with a green success card below the AI's response.

The action layer supports five operations today, with more on the way:

• Create task — Full task creation with title, notes, priority, status, due date, assignee (by name), project (by name), vault file attachments, external links, and deliverable requirements. Everything in one call, nothing left blank.
• Update task — Change any field on an existing task, found by fuzzy title match. Add vault files, add links, reassign, reschedule — partial updates work fine.
• Delete task — Finds the task by title, shows a confirmation card in the UI, deletes only after you confirm. No accidental deletions.
• Create project — New project with name, description, priority, status, start date, and end date.
• Update project — Update any field on an existing project found by fuzzy name match.

Every action is idempotent — the AI will never create the same task twice even if you ask the same question twice. And every action is transparent: the response always tells you exactly what was created or changed.

Eight read tools. Every layer of your workspace.

The AI does not dump your entire workspace into a prompt and hope for the best. It uses structured read tools — functions it can call to fetch exactly the data it needs, when it needs it. This keeps responses precise and prevents hallucination from stale or irrelevant context.

The eight read tools available today cover every major module:

• get_workspace_overview — High-level stats across all modules: task counts, overdue figures, pipeline value, team count, upcoming events this week.
• get_tasks — Fetch tasks with filter presets: all active, overdue, blocked, due today, due this week, recently completed. Can filter by project name or assignee name.
• get_projects — All projects with task completion stats, priorities, and deadlines. Filterable by status or name.
• get_team_workload — Every team member with their active task count and a workload label (FREE / LIGHT / MODERATE / HEAVY). Used automatically before assigning tasks.
• get_crm_pipeline — Full pipeline grouped by stage, with deal values, weighted values, days in stage, and stale flags. Can include clients.
• get_calendar — Upcoming or recent events with flexible range presets: today, this week, next 7 days, past 30 days.
• get_vault_files — Files and documents from the project vault, filterable by project or searchable by title. Used to resolve file names before attaching them to tasks.
• search_contacts — Full contact profile by name: pipeline stage, deal value, upcoming meetings, recent past meetings with outcomes, recent email threads.

What makes this fundamentally different from every other AI tool

The question we get most often is: "How is this different from ChatGPT or Claude?"

The answer is context. Generic AI models know nothing about your agency. You would have to paste in your tasks, your client list, your pipeline, your calendar, and your messages every single time — and even then, the model only knows what you copy-pasted, which is always a fraction of what actually matters.

Kobin AI is pre-loaded with your live workspace. Every response is grounded in real data from your Supabase instance, not from training data or hallucinated guesses. When it tells you a project is at risk, it is because it pulled the actual task completion rate and compared it to the actual end date. When it suggests assigning to a specific team member, it is because it fetched their real current workload.

The second difference is multi-step reasoning. A simple AI assistant can answer "what tasks are overdue?" in one step. But answering "assign the most urgent overdue task to whoever has the lightest workload and set the due date to end of week" requires the model to first fetch overdue tasks, then fetch team workload, then resolve names, then create the update — all in sequence, with the output of each step informing the next. Kobin AI does this natively, using a structured tool loop that runs up to four reasoning steps before generating a response.

What we are building next

The AI layer ships today with task and project actions. The modules coming next are already in development:

• @AI draft a follow-up to [contact] — Pull their full CRM profile, last email thread, and meeting outcomes, then draft the message in context.
• @AI schedule a meeting with [client] — Find a free slot on your calendar, create the event, send the invite card in chat.
• Daily morning brief — A push notification every morning at 8am: overdue tasks, today's meetings, clients awaiting response, projects needing attention, and quick wins.
• Pre-meeting brief — 10 minutes before every calendar event: client profile, last conversation summary, open tasks, suggested talking points.
• Client silence detection — Background scan every 6 hours. If a client with open tasks hasn't replied in 4+ days, the AI drafts a follow-up in your voice for one-tap review.
• Weekly client report draft — Every Friday at 5pm, a draft report per active client: work completed, project status, pending client actions, next week's focus.

Frequently asked questions