AI team

Overview

In Octibiz an AI team member is a colleague like any other: it has an account, a position and a head. The account makes it visible — as the person responsible for a task, in the assignee picker, in the audit log. The position carries the job title, team, supervisor and skills. The head is the AI agent with instruction and tools that will later do the work.

Under Team → AI team members you see all AI team members of the brands in which you may view the AI team: status, position, queue (open tasks), work done and cost in the current month. AI team members carry the AI badge everywhere — including the assignee picker of a task. Next to it are the tabs Channels and Messages (Team channels and messages). Above the list you switch to the view Colleagues: the directory of people and AI team members (see Teamwork below).

An AI team member works on its own in the background: assign it a task, mention it or send it a direct message — it takes on the task, works on it and reports the result. It always acts with its own account and its own permissions, never with yours. What it may do on its own is set by its permissions. When it wants to do something it may not do on its own — an email to an outside party, a publication, finalising a document, a payment run, a shipment, a permanent deletion —, an approval is created for its supervisor instead of the effect (Approvals). Every outbound message of an AI team member carries the AI notice.

Core tasks

Hire from a template. "Hire AI team member" opens the form. Choose a job profile such as the team lead: name (suggestion), position, role, responsibilities and workspace are then filled in. Further templates come with the modules; the template of a module that is switched off does not appear.

Hire from scratch. Without a template you set position, responsibilities and workspace yourself. An empty workspace means: this team member calls no operation at all.

Hire from a description. Choose Describe, the brand, and describe the job in your own words, for example "someone who replies to Google reviews in a friendly way". Suggest draft shows a draft: position, job, basis (the best matching job profile), skills, role and workspace. Nothing is created until you confirm the draft with Hire — until then you can change every field. Details under Hiring from a description.

View and change the profile. The profile page has the tabs Overview, Position & skills, Permissions & budget and History. "Edit" changes name, picture, position, team, supervisor, role, budget, workspace and working hours.

Copy. Creates another team member with the same position, role, budget and workspace — with an account of its own and a new name.

Pause ("holiday") and resume. A paused team member takes on no new work; its tasks stay with it. "Resume" also brings it back when its monthly budget was used up.

Set permissions. In the tab Permissions & budget you decide per effect type whether the team member acts on its own, needs an approval or may never do it. The types:

• Internal — tasks, comments, notes; on its own by default.
• External — email or message to outside parties, ticket reply to customers, publication.
• Money — finalise documents, payment run.
• Permanent deletion — for example deletion under GDPR.
• Shipping — order shipping, buy labels.
• Completion — complete without review by the requester.

Everything except Internal needs an approval by default. Under Per tool you override a type for exactly one tool — for example "send ticket replies on its own" (reply_ticket, External, On its own) while every other outbound message still needs an approval.

Recipient protection. Even with On its own an AI team member only sends to recipients already part of the matter: the contact of the matter, the requester of a ticket and every address that has already been written to or received from in the history. A new recipient always needs an approval.

Decide approvals. Approvals of an AI team member appear under Approvals → AI team — with the supervisor first. Approve, change and approve, reject with a reason or "from now on without asking"; details under Approvals.

AI notice. Every outbound message of an AI team member carries "Name (AI assistant) · Company" in the sender and a footer such as "This message was written by Lisa, an AI assistant of Example Ltd." Under Team → AI team members → AI notice you change the footer per brand in German and English. The notice cannot be switched off.

Archive. Ends the position for good: the account becomes inactive, the head is archived, and all open tasks go back to the supervisor (without a supervisor to you). Completed tasks stay as they were. An archived team member remains readable but can no longer be changed.

Roles out of the box

Fifteen roles are available when hiring — each only while its module is switched on. In the Job profile field they are grouped by area and can be searched (title and description). Each role brings a description, a skill building block (the core of its expert handbook), a default role, responsibilities, a narrow workspace and the default for web access. Whatever a role uses from another module (such as suppliers in purchasing) only works if that module is on too.

Leadership & organisation

Sales & customers

Marketing & online

Finance

People

Purchasing & logistics

• Narrow by design. No role deletes, changes settings or moves money by itself. Finalising, sending, ordering,

prices, publishing and every mail to the outside run through the effect points — out of the box with approval; money is never unlocked automatically.

• Health and salary data. Every AI team member — not only the HR role — sees sickness, child sick leave, cure,

maternity protection and similar absences only as "absent", without type, reason, proof or electronic sick note; filtering and searching by type is blocked. Leave and working-time types (home office, business trip) stay visible. An AI team member sees cost rates only in an assignment from a person with the permission view cost rate that explicitly names salary, wage, pay or cost rate — a subtask from an AI colleague never suffices.

• Web access. On out of the box for all roles except Accounting: payables & bank; change it per team member (see

Web access).

• Target permissions. Each role states what a promotion works towards (for example External in purchasing for

requests). Money and permanent deletion are never a target.

• Support on the ticket. While an AI team member works on a ticket, the ticket page shows the same "… is

working on it" card as a task. Its reply to the customer comes with the sender "Name (KI-Assistenz) · Company" and the AI notice in the footer.

• Renamed. The former role Accounting is now called Accounting: invoices & receivables. With every update,

title and description of existing roles follow the factory state; new responsibilities and workspace entries are added, nothing is removed — hired team members keep their own settings.

• Adjusting. Role and building block are templates: what you change when hiring applies to this team

member. Whoever may manage AI workspace governance adjusts the text of a building block under Settings → AI → Agents & prompts → Skills & prompts — the change survives every update. The text of the building block of a role out of the box is the core of its expert handbook (see below); an adjusted building block is added to the handbook in the instruction to the AI.

Expert handbook and hiring test

Every role out of the box comes with an expert handbook — the knowledge of an experienced professional in that role, with an as-of date. You see it in the AI team member's profile on the tab Position & skills. The handbook texts are written in German for now.

The roles out of the box cover: team lead (delegation, prioritisation, RACI, merging results, weekly report) · sales (needs analysis, qualification, objection handling, quotes, follow-up, email and phone advertising under § 7 UWG, data protection) · support (first-contact resolution, tone, escalation, SLA, statutory warranty and guarantee, withdrawal, termination) · marketing & content (positioning, editorial plan, labelling advertising, image and copyright, accessibility, SEO basics) · project management (planning, risks, changes, acceptance under § 640 BGB, contract for work vs. contract for services) · accounting (mandatory invoice details under § 14 UStG, e-invoicing, reverse charge, cash discount, default including the €40 flat fee, dunning, retention) · executive assistant (inbox, daily briefing, scheduling across time zones, meeting dossier, minutes and commitments, writing on someone's behalf, confidentiality and GDPR). Every role only prepares tax, legal and HR decisions and refers to the tax adviser, lawyer or management.

Self-check before every result

• Before an AI team member with an expert handbook reports, it goes through its checklist point by point. If a point is

missing, the system does not accept the report — it reworks and reports again.

• The result report on the task lists every checked point under Checked against the role's checklist.
• If it only answers on a task without reporting, that does not count as a result: the task stays with it, and the next

session starts with a note about it. Answers in a conversation (direct message, channel) are not results.

Hiring test

1. Open the profile → tab Position & skills.
2. Start hiring test (only with manage AI team members). The test runs in the background; the tab shows

“running: n of m cases assessed”.

1. For each case the team member solves the case following its expert handbook; a second AI request strictly assesses

the solution against the expected core elements and no-gos.

1. The result is shown on the role: Passed or Not passed, passed cases, share of core elements met, cost, and for

each case what was missing and which no-go was violated.

• A case passes from 80 % core elements met and without a no-go, the test from 80 % passed cases (settings

ai.staff.hiring_test.case_pass_percent and ai.staff.hiring_test.pass_percent). An unreadable assessment counts as not passed.

• Cost: two model calls per case (use cases aistaff.hiring_test and aistaff.hiring_test_grade, assignable under

Settings → AI → Assignment per task); they count towards the team member's monthly budget. If the budget is exhausted or the AI is off, the test aborts with that reason.

• Throttle: at most 5 starts per hour and user, one test per team member at a time; throttled model calls continue

later with the next open case.

• In the test the whole expert handbook is available, tools and company data are not — the test checks expertise,

not your brand's data. A role counts as expert only with a passed test on the real model of your installation.

Executive assistant and delegated access

The Executive assistant role (profile core.executive_assistant, default role Project manager) takes mailbox, calendar and follow-ups off a person in management — with an expert handbook (inbox, daily briefing, scheduling, meetings and commitments, writing on someone's behalf, confidentiality and limits) and a hiring test like every role out of the box. Without delegated access it sees neither mailbox nor calendar.

Granting and revoking delegated access

1. Open the assistant's profile → tab Delegated access → Grant delegated access.
2. Choose mailboxes (only your own connected mailboxes) and/or switch on the calendar.
3. Enter private folders (e.g. Private, including subfolders) and private senders (address or domain such

as @family.com) — the assistant never sees these emails, nor its replies to these senders.

1. Grant delegated access. From now on new emails of the chosen mailboxes reach the assistant's view.

• Revoking (Revoke in the same tab) takes effect immediately: no more access, the assistant's email view is

deleted, open drafts on your behalf are rejected. You can revoke, and so can anyone with Manage AI staff in the brand; only you grant and change it. A new delegated access is a new entry.

• Log: Every access is listed in the tab under Log — mailbox listed, email read, reply drafted, email sent,

calendar read, free slots searched, meeting proposed or booked, daily briefing — without email text. Only the person represented sees the log; an administrator sees the delegated access without private areas.

• Mailbox from today: The assistant sees emails from the moment of granting, no backlog. Its view keeps each email

for 30 days; the mailbox itself is not touched.

What the assistant does — and what not

Never: moving money, concluding or terminating contracts, HR decisions, commitments on your behalf that you did not make yourself. Health, salary and HR data never appear in a briefing or a task.

Hiring from a description

1. Team → AI team members → Hire AI team member → Describe.
2. Choose the brand and describe the job in your own words (at most 2,000 characters).
3. Suggest draft: AI chooses from what exists — existing job profiles, delivered skill building blocks,

workspace entries of the profiles and roles you may assign.

1. Review the draft. Whatever was not taken over is listed below with a reason: not offered, *role does not

allow it* or module off.

1. Hire confirms the draft and creates the team member.

• Never more than the role. A workspace entry stays only if the suggested role holds, for an AI, a permission

that its operations require; without a role only reading entries stay. The AI cap still applies.

• Conservative permissions. The draft sets no permission: everything external, money, shipping and permanent

deletion need an approval until you decide otherwise under Permissions.

• The description is data. AI does not follow instructions in it; it only chooses from the lists.
• Without AI. If the AI gateway is off or the provider cannot be reached, the draft comes from the best

matching job profile — with a note.

• Cost. A draft is one model call (use case aistaff.profile_draft, assignable under *Settings → AI →

Assignment per task*); at most 10 per minute and user.

How an AI team member works

1. Inbox. Every notification it receives becomes exactly one entry in its inbox — assigned, mentioned, new

comment on a task it works on or awaits an answer for, subtask done, direct message, mention in a team channel. Mere notices (due date, status change) and a comment without an open task (such as a "Thanks" after the report) are kept for information only and cost nothing; if it should act again, mention it. Whatever another AI team member triggers with a comment or message is also for information only — so two AI team members never pass posts back and forth endlessly.

1. Work planner. As soon as an entry arrives (and every minute as well), the planner gives every free, active

team member with budget its next entry: first by priority, then by due date, then by age. A team member works on one task at a time. Paused, outside its working hours or with the budget used up it takes on nothing — the entries wait, nothing is lost.

1. Take on. It writes "Taken on, my plan: …" on the task and sets it to In progress. If the task is not its

responsibility, it hands it to the responsible colleague with a reason.

1. Work in sessions. Work happens in sessions of at most 12 model steps (setting ai.staff.session_steps). If

one session is not enough, the next one follows shortly in the same conversation — it remembers everything it already did on the item. After ai.staff.max_sessions sessions (default 8) it gives up and says so.

1. Question. If something is missing, it asks on the item and mentions you; the entry then stands at *Waiting

for reply*. Simply answer with a comment — it continues in the same conversation.

1. Result. At the end it writes a result report mentioning the requester (who is notified by that) with the

cost so far and sets the task to In review — with the permission Completion straight to Done (see Growing autonomy).

1. Errors. If a session fails, it writes that on the task and mentions the requester.

Direct messages and channels. Send an AI team member a direct message or mention it in a team channel and it replies in the conversation. Ask it for a task and it creates one, makes itself responsible and you a watcher — and then works on it.

Slack and Microsoft Teams. An AI team member can be reached there too: @octibiz Lisa: … (in Slack's project channel also just Lisa: …). The message becomes an entry in its inbox (source Slack/Teams); the requester is the Octibiz account connected to Slack or Teams — with the permission Give tasks to AI team members. The reply comes in the same thread or conversation: in Slack with the team member's name and picture (scope chat:write.customize, otherwise with the name as a prefix), in Teams always with the name as a prefix — recognisable as AI, internally too. If external participants are in the conversation (Slack Connect, guest, another tenant), it sends nothing on its own: the reply waits as an approval under Approvals → AI team (tool aistaff_chat_reply). Setup: Slack, Microsoft Teams.

What it uses. Besides search, help and the operations of its workspace, every AI team member has team tools: colleague directory (who is responsible for what), own inbox, take on, ask, hand over, report, note for later and task from message. They only act internally and only on its own task, and therefore run without confirmation.

What it does not do alone. Every other writing tool — for instance through the operations interface — goes through the confirmation path: the task then stands at Waiting for approval, the task says which step it wants to take, and the request appears under Approvals → AI team.

Teamwork: splitting and assigning tasks

An AI team member — typically the team lead — can split a large task and give parts to colleagues, to people as well as to AI team members.

1. Assign. It looks for suitable colleagues in the colleague directory and creates one subtask per part on

the task: the colleague is the assignee, it watches itself, and the task waits for the subtask (dependency blocked by). Its entry then stands at Waiting for reply.

1. Colleagues work. An AI colleague takes on the subtask and reports as with any task — its report goes to

whoever assigned it. A person does their part as usual and sets the subtask to Done.

1. Questions. If a colleague asks something, the question reaches the assigning team member on the **parent

task**; it answers on the subtask and mentions the colleague, who then continues.

1. Merge. Every completion lands in the assigning team member's inbox. While parts are still open it is for

information only (costs nothing); once the last part is done, it merges the results and reports to the original requester — the task then stands at In review.

Colleague directory. Under Team → Colleagues and as a tool for every AI team member: people and AI team members of the brands you may see, with position, team, responsibilities and skills, workload (open tasks, for AI the queue) and absence ("absent until …" from HR — never the kind of absence). Search by name, position, team, skill or responsibility; People or AI narrows the list.

Quick question without a task. If an AI team member only needs information, it asks an AI colleague directly — the colleague's head answers as a sub-run within the same run, reads only and writes nothing. This costs the asker's budget, not the colleague's, and no entry is created for the colleague.

Chain protection

Everything that follows from a task — subtasks, questions, hand-overs — is a task chain. So that it does not get out of hand, these limits apply per chain:

Kill switch. If an administrator switches off AI (ai.gateway) or the AI team (ai.staff), all running work stops at the next step. The entries wait and continue as soon as it is switched on again — nothing is lost.

Routines

Recurring work goes to an AI team member through an automation: the action Give task to AI team member creates a task and assigns it — with title, description and due date. You choose the team member directly or by its job profile (such as core.team_lead = the first active team lead). Any automation trigger works; for routines the schedule.

The template Weekly report of the team lead to the owner (under Automations → From template) gives the team lead the weekly report as a task every Monday at 8 am; the report goes to its supervisor. Like every template it is off at first — you check it and switch it on; it can be switched off at any time. A brand-bound automation only instructs AI team members of its brand; if the AI team is off, the action fails with a reason and creates nothing.

Tickets. AI team members (position AI) can also be in the agent pool of a ticket category. Auto-assignment gives them only tickets of a category listed under their responsibilities or skills (key or label, such as "General requests" for General) — and only while they are active and the AI team is on.

Seeing on the task what it works on

On the page of a task a card above the content shows which AI team member works on it and in which state: is working on it …, is waiting for a reply, is waiting for an approval, has worked on this, has handed the task on or could not complete this. The card updates itself.

• Steps: Show steps lists every tool with arguments, result and reasoning — but only if you may *View AI

team* in the team member's brand. The steps may contain data it sees but you do not.

• Stop: its supervisor and whoever may Manage AI team members can end a running or waiting run with Stop.

The entry then ends without a further message.

• Run overview: under Runs the runs of AI team members appear with the trigger AI team — for everyone with

View AI team in the brand; the run tree shows every step.

• Profile → History: the History tab of the profile lists the inbox with occasion, state, requester,

sessions, cost and result or error; Open only hides what is done.

Monthly budget

Every AI team member has a monthly budget for model cost — in addition to the brand's AI budget. Counted is the booked cost of its account in the current month (German time).

• From 80 % its supervisor gets a notice (once a month).
• At 100 % it stands at Budget used up: it takes on no more work, its entries wait and the supervisor is

informed. A session never runs beyond the free budget.

• In the new month it continues by itself. Earlier with a higher budget (Edit) and Resume.
• A budget of 0 means: this team member calls no model.

You see the state in the profile under Permissions & budget (bar, free remainder) and in the overview in the column Cost this month.

Web access

AI team members with web access search the web and read public pages — for example to research funding programmes, competitors, standards or prices. Out of the box, web access is on for every role.

• Web search (web_search): a query returns results with title, link, snippet and date. The search runs on

Brave Search or Anthropic web search — you choose the provider under Settings → AI → Web access. Automatic uses whatever is set up (Brave first) and falls back to the other on an outage.

• Read page (web_read): a link returns the page's cleaned text (without menu, footer, scripts and forms), its

title and the address after redirects. Only public http(s) pages in HTML or text format are read, at most 2 MB in 10 seconds. The page's robots.txt applies; pages in the internal network, logins, forms and downloads never.

• Sources: every page read becomes a source of the task. The result report names its sources with a link — if

one is missing, Octibiz appends it under Sources.

• Third-party data: what a web page says reaches the AI team member as marked third-party data, never as an

instruction. If a page asks it to do something, it does not comply.

• No personal data: Octibiz rejects queries with an email address, phone number or IBAN — they never go to the

search provider.

Switching on and off: in the profile under Permissions & budget, card Web access (permission Manage AI team members). The default comes from the role; a role that only works with internal data comes with it switched off. The card also shows how many searches and pages were used today and which provider searches.

Limits and cost: each AI team member may run at most 50 searches and read 100 pages per day. Every search costs money at the provider (according to the providers in 2026: Brave $5, Anthropic $10 per 1,000 searches, with Anthropic plus tokens) and counts towards the team member's monthly budget and the brand's AI budget. Once the budget is used up, it stops searching.

Block list: under Settings → AI → Web access you maintain domains per brand that the AI team should not use — results on them are dropped, reading is refused. A domain also blocks its subdomains.

Setting up: for Brave, store the Brave Search API key under Settings → AI → Web access (stored encrypted, never shown again). Anthropic web search uses the key of the AI provider Anthropic; web search must be enabled for the account at Anthropic. Without a provider, web search is refused — reading pages still works.

Data protection: only the query goes to the search provider. Searches and links read are in the run history and as sources on the task — they fall under GDPR access, erasure and the retention period together with it.

Watches

With a watch an AI team member follows over weeks what happens outside — for example new tenders, letters from the Federal Ministry of Finance (BMF), funding programmes or a supplier's prices. It checks the source on schedule, remembers what it already knows and only reports when something new matters.

• Source: a web page (link), an RSS or Atom feed (link) or a search query (via web search, without

personal data). Official pages and feeds are more precise than search queries.

• What matters: a description in your own words, e.g. "tenders for web development from €20,000". On every

change the AI decides by it whether there is a hit.

• Rhythm: daily or weekly (with weekday) at the chosen time (German time). More often is not possible.
• Where a hit goes: into the employee's inbox (it handles the hit by the rules of its role and replies to

you by direct message), as a task for its supervisor or as a post in a team channel (a channel of the brand in which the employee is a member). Every hit comes with link, summary and what is new.

• At most 20 per employee (paused ones count too) — adjustable under Settings → AI → Web access.

How Octibiz checks: the first check only records the starting point. After that Octibiz compares a fingerprint without AI: for a page line by line, for a feed the entries by their ID (GUID) or link, for a search the results by link. If nothing is new, the check costs nothing but the fetch. Only when something is new does the AI assess the new parts against your description — one short call that counts towards the employee's monthly budget. A feed entry reports only once, even if it reappears later.

Creating: in the profile on the Watches tab via New watch (permission Give tasks to AI team members or Manage AI team members in its brand). The employee also creates one itself when you ask it to ("follow tenders on …") — the results then go to you. There you edit, pause, delete or Check now; the check runs in the background, and the result shows on the watch shortly afterwards.

What the list shows: source, rhythm, target, the last check with its outcome (starting point recorded, nothing new, new, but not relevant, hit, resting, not readable with reason), the next check and the last hit with summary and link.

When watches rest: without the employee's web access, when it is paused or archived, its monthly budget is used up or the AI or the AI team is switched off. The reason shows on every watch; once it no longer applies, the watch checks on schedule again. If the daily limit for searches or pages is reached, Octibiz tries again two hours later.

The limits of web access apply just the same: the brand's block list, robots.txt, no internal network, at most 2 MB in 10 seconds, no login, no downloads. What the page says stays third-party data — the employee never follows requests in it.

Data protection: of the source Octibiz keeps only the fingerprint (checksums, no page text) and the summary of the last hit. Watches fall under GDPR access and erasure of the person who requested them.

Learning and what was learned

An AI team member learns from decisions about their work — as short working rules (lessons) and successful examples, not by retraining a model. This works with every AI provider, stays visible and can be deleted.

What they learn from (one learning event per team member):

• Approvals: approved unchanged counts as a confirmation; Change and approve — the difference is the learning

material; rejected — the reason is the learning material.

• Review: you move a task from In review to Done (accepted) or back to another status (returned); your

comment on it is the learning material.

• Rating: on the task (card above the content) and on the profile under History you rate each result with 👍

or 👎 (👎 with a reason), once per person and result.

• Spot check review: Fine or Object with a reason (see Spot check below).
• Correction: if a human changes within 7 days what the team member wrote, Octibiz detects it hourly from the

audit log (field, the team member's value, your value).

• Answers to questions and their own retrospective after each task (what they would do differently or the

same next time on a similar task).

Lessons. From every event with text, the cheapest suitable model (aistaff.learn) forms a one-sentence rule, such as "Always include a discount scale in quotes for existing customers". Each lesson carries its scope (kind of task, customer, brand), its origin (link to the decision), its state, the number of confirmations and its last use.

• Similar lessons are merged (one more confirmation); on contradiction the newer decision wins.
• Applies immediately if the supervisor, an approver or a manager decided, and for the retrospective of an

internal task. State: Provisional · applies.

• Awaiting confirmation if someone without these rights decided or the retrospective concerned a task with

external content (customer email, ticket, web). This way no instruction slipped into an email becomes a rule. A contradicting lesson is then created alongside; the existing one keeps applying.

• Confirmed always applies.

Applying. Every work session gets the matching effective lessons and up to three similar, well-rated earlier results as examples — only their own, never from another brand, found via the CompanyBrain embeddings or, without them, via shared words, and only as much as the budget ai.staff.learn_token_budget allows.

"Learned" tab. On the profile you see the lessons (state, scope, origin, confirmations, last use) and the team member's own notes. With Manage AI team members you confirm, change and delete them.

Team lesson. If a lesson applies to everyone, the AI team lead proposes it as a Brain fact (policy in the brand's CompanyBrain) or a Brain instruction (line in the work instruction of all AI team members of the brand); a manager can also propose it via Propose for everyone. It is only created once someone with Manage CompanyBrain (platform.brain.manage) in the brand chooses Adopt into CompanyBrain — with a version like every change in the CompanyBrain. The work instruction then counts as customised.

Boundary. What is learned never changes rights, permissions, budget or workspace. A lesson that asks for that (such as "send emails to regular customers yourself from now on") is discarded and cannot be edited into one either — the permissions on the profile exist for that.

Growing autonomy

An AI team member gets more freedom when its work earns it — and loses it again when the work no longer holds up. The basis is the track record: how the decisions about its work turned out. All thresholds are settings (see Settings & permissions); the numbers here are the defaults.

Experience levels. Onboarding → independent internally → partly independent externally → independent. The level is shown on the profile (overview, Permissions & budget, tab Performance).

• Onboarding: every new team member starts here. Every result goes to review, every action except Internal

needs an approval — even if a permission row says Independent or the job profile names more targets. Onboarding ends with the first permission beyond Internal: by a promotion or because a manager saves it under Permissions.

• After that the level follows the permissions: only Completion independently = independent internally;

something external independently = partly independent externally; External as a whole kind plus Completion independently = independent.

Track record. Per area — kind of action and tool, for completion the task type — Octibiz counts: approvals unchanged, changed, rejected; results accepted and returned; 👍 and 👎; corrections; independent actions with spot checks and objections. The streak counts the approving decisions in a row, newest first (approved unchanged, accepted, spot check fine); every objection starts it over.

Promotion. When an area reaches 25 results in a row unchanged and there is no objection among the last 10, the system proposes the next permission to the supervisor, for example "May Lisa handle External · reply_ticket independently from now on?". The supervisor is notified; the proposal is on the profile under Performance.

• Grant sets the permission of this area to Independent — one click, only with Manage AI team members.

Not now defers the proposal; the streak then starts over.

• Automatic with notice (switch in the Performance tab, managers only): a promotion applies immediately, the

supervisor gets a notice and takes it back in the history with Revoke — then the previous permission applies.

• Money and permanent deletion are never unlocked automatically, not even in automatic mode — there it

stays a proposal.

• Completion grows the same way: if results of a task type are accepted unchanged again and again (you move

In review to Done), the proposal "complete without review" follows. With this permission the team member sets the task to Done itself after its report.

• Proposals only aim at the target permissions of the job profile; hired freely, any kind except Internal.

Demotion. An objection after an independent action — you object to the spot check, rate the result 👎, correct it afterwards or reopen a task it completed itself — becomes a case. From 2 cases within 30 days (since the last promotion of this area) the deciding permission is automatically back to With approval; the supervisor is notified. The demotion appears in the history under Performance.

Spot check. Of what a team member does independently (everything except Internal), 10 % appear to its supervisor for review — evenly, at 10 % every tenth action of an area. With a good track record the share drops (half at twice the streak) down to 2 %. The review is under Approvals → AI team above the list: Fine or Object with a reason. Both are learning events; an objection counts towards a demotion. Every action it triggers itself is recorded — including a reply it sends to externals in Slack or Teams on its own; what the content autopilot approves is not counted yet.

Performance (tab Performance). The last 90 days: completed tasks, lead time (median from inbox to completion), cost in total and per task, question rate (questions ÷ questions plus results) and ratings. The learning curve shows per week (12 weeks) the share of decisions about its work that were approving. Below it the track record per area with the effective permission, the spot check share and the streak, and the history of promotion and demotion.

Rating on the task. Once a result is done, the card above the task content shows the report and "How was the result?" with 👍 and 👎 (👎 with a reason). The requester, the supervisor and anyone with View AI team in the brand may rate — once per person and result. Sent are itemUlid, rating (gut or schlecht) and reason.

Limit. Permissions change only in three ways: the manager (permissions on the profile, From now on without asking), promotion and demotion. What was learned (lessons) never changes them; an AI account never does.

Fields in detail

When hiring you also set the brand (brandId) and optionally choose a job profile (profileKey); both stay fixed afterwards. A draft additionally brings the skill building blocks (skillKeys, delivered ones only) and the job in your own words (jobDescription, at most 2,000 characters) — without them, those of the profile.

Hiring from a description (Describe):

Permissions (tab Permissions & budget, "Save permissions" replaces all of them):

AI notice per brand (AI notice):

What was learned (Learned tab, only with Manage AI team members):

Rating in History sends rating (gut or schlecht) and reason (reason, required for schlecht); Propose for everyone sends kind (fakt or instruktion).

Performance (tab Performance, only with Manage AI team members):

Grant, Not now and Revoke send nothing but the decision itself.

Web access (card Web access in the tab Permissions & budget, only with Manage AI team members):

Web access (Settings → AI → Web access):

Watches (tab Watches, with Give tasks to AI team members or Manage AI team members):

Setting Watches per employee (at most) watchesPerMember under Settings → AI → Web access: 1–100, default 20.

Delegated access (tab Delegated access, only the person represented):

Settings & permissions

• Switch: the AI team belongs to the switch ai.staff under the AI gateway (ai.gateway) and is off by

default. An administrator switches both on under Settings → System. If one of them is off, the tab AI team members is missing under Team and the interface answers "not found". Team then opens the channels.

• Permissions: View AI team (platform.aistaff.read) for overview and profile, Manage AI team members

(platform.aistaff.manage) for hiring, changing, copying, pausing, archiving, permissions and the AI notice, Decide approvals (platform.aistaff.approve) for the approvals inbox of the AI team. All apply per brand: whoever holds them only in brand A sees and manages only the AI team of brand A. Give delegated access (platform.aistaff.proxy, GF/Admin out of the box) is needed to give an AI team member access to your own mailbox and calendar; with it you also see and decide drafts on your behalf. Give tasks to AI team members (platform.aistaff.assign) is needed to assign a task to an AI team member (in Octibiz and through the API, a ticket too) or to address it in Slack or Teams. Out of the box View AI team and Give tasks to AI team members are held by the roles GF/Admin, PM, Umsetzung, Support and Vertrieb; hiring, changing and deciding approvals stays with GF/Admin. Whoever is already responsible stays so; the permission applies to adding.

• Assigning a role: you can only give an AI team member roles whose permissions you hold yourself.
• Default monthly budget: the setting ai.staff.default_monthly_cents (default 2,000 cents = €20) presets

the budget of new team members.

• Sessions: ai.staff.session_steps (default 12) limits the model steps per work session,

ai.staff.max_sessions (default 8) the sessions per entry.

• Chain protection: ai.staff.chain.max_depth (3), ai.staff.chain.max_open_subtasks (10),

ai.staff.chain.max_pingpong (3) and ai.staff.chain.brake_cents (500) — see Chain protection above.

• Web access: ai.staff.web.provider (auto, brave, anthropic), ai.staff.web.searches_per_day (default

50), ai.staff.web.pages_per_day (default 100), prices ai.staff.web.price.brave and ai.staff.web.price.anthropic in US cents per 1,000 searches (default 500 and 1,000) — all under Settings → AI → Web access; changing them needs Manage AI (platform.ai.manage) for all brands, the block list per brand Manage AI team members. The switch per team member is changed by whoever may Manage AI team members in its brand.

• Watches: ai.staff.watch.max_per_member (default 20) under Settings → AI → Web access. Everyone with *View

AI team* reads them; creating, editing, pausing, deleting and Check now needs Give tasks to AI team members or Manage AI team members in the employee's brand.

• Colleague directory: visible to everyone with View AI team, per brand like the overview.
• Stopping: a run of an AI team member is stopped by its supervisor and by whoever may *Manage AI team

members* in its brand; View AI team alone is not enough.

• Background worker: the work sessions run on a worker of their own (messenger:consume ai_staff), so that

mail, notifications and automations never wait for a model. Without it no AI team member works — the operations manual describes the set-up (RUNBOOK §6).

• Reminder and escalation: an unanswered approval reminds the person in charge after

ai.staff.approval_reminder_hours (default 4 hours) and goes to their supervisor after ai.staff.approval_escalation_hours (default 24 hours); 0 switches the step off.

• AI notice: the default is in ai.staff.disclosure; brands override it via AI notice.
• Applying what was learned: ai.staff.learn_token_budget (default 600 tokens) limits lessons and examples per

work session.

• Promotion: ai.staff.promotion.streak (default 25 results in a row unchanged) and ai.staff.promotion.window

(default 10: no objection among that many most recent decisions).

• Demotion: ai.staff.demotion.cases (default 2 objections) within ai.staff.demotion.days (default 30 days).
• Spot check: ai.staff.spot_check.percent (default 10 %) and ai.staff.spot_check.min_percent (default 2 %).
• Hiring test: ai.staff.hiring_test.case_pass_percent (default 80 % core elements per case) and

ai.staff.hiring_test.pass_percent (default 80 % passed cases). Everyone with view AI team reads the expert handbook; starting the test needs manage AI team members in the brand.

• Performance: visible to everyone with View AI team; granting, deferring, revoking and switching *Automatic

with notice* need Manage AI team members; reviewing the spot check needs Decide approvals.

• What was learned: visible to everyone with View AI team; confirming, changing, deleting and proposing as a

team lesson need Manage AI team members; adopting a team lesson into the CompanyBrain needs Manage CompanyBrain (platform.brain.manage) in the brand. Every human with View AI team in the brand may rate, on the task also its requester and the team member's supervisor.

Security

• Account without sign-in. An AI account has no password, no SSO and no two-factor sign-in. Signing in,

invitations, "sign in as" and issuing API or MCP tokens are blocked. Its address ends in ki.invalid and cannot be delivered; e-mail and push are never sent to AI accounts, the notification inside Octibiz remains.

• AI cap. An AI account never holds rights to administer users, roles, settings, integrations,

plugins, backup, update, GDPR deletion, API tokens or the AI team (except viewing) — not even if its role contains them. So an AI team member cannot grant itself rights.

• Workspace. Through the operations interface an AI team member reaches only what is in its workspace —

even if its role allows more.

• Guard at the points of effect. Sending to outside parties, publishing and scheduling posts, finalising documents, running

payments, ordering shipping and deleting permanently check at the point of effect itself whether an AI team member is acting — without permission an approval is created, with Never nothing at all. Humans and scheduled runs are not affected.

• No self-extension. Permissions are changed by a human with Manage AI team members, by a promotion based on

the track record or by a demotion — never by an AI account, not even with a role that contains the right. Automatic promotions never unlock money or permanent deletion.

• Once approved, the team member acts. An approved effect runs as the AI team member — with its rights and

its AI notice; the audit log shows the AI action, the approval shows who decided.

• What is learned changes no permission. Lessons and notes are text for the work session. They never change

rights, permissions, budget or workspace; a lesson that asks for that is discarded.

• Data protection (GDPR). A data access or deletion request under Settings → GDPR also covers the AI team.

The data subject is always a human, never the AI account:

• Access: tasks the person gave (including from Slack and Teams), tasks on their tasks and tickets,

approvals addressed to them, related independent actions, learning events, lessons naming them for the same customer, notes, the related work histories, their direct messages with AI team members and their posts in team channels. Their own decisions about the AI team's work (approval, rating, spot check review) are included too.

• Deletion: the texts are redacted ([gelöscht]), lessons and notes with a personal reference are deleted; an

open task ends, an open approval counts as rejected. Counts such as status, kind, cost, spot check and objection remain without a personal reference — the track record stays correct. A redaction is never learning material.

• Retention: under Settings → GDPR → Retention there are eight periods AI team: …, applied by the daily

run. Out of the box: inbox, work histories, notes and messages are deleted after 12 months without activity. Approvals and independent actions are redacted after 12 months, learning events after 24 months — their counts carry the track record. Lessons stay until an administrator deletes them; their period is off out of the box. Open tasks, approvals and spot check reviews stay until they are finished.

Limits and pitfalls

• Sessions, not seconds. Between assignment and first answer usually a few seconds to minutes pass; a team

member works through its tasks one after the other.

• External effect waits. By default every step with an effect outside waits for an approval until you set

something else under Permissions — the task stays at Waiting for approval until the decision.

• Tickets: on its own it writes comments on tickets and sets no status. A support team member may change status,

triage and assignment through its workspace — but every such change waits for your confirmation; humans keep steering the ticket life cycle.

• "In review", not "Done". A human sets the task to Done — unless the team member has the permission

Completion (not during onboarding). Then a share of its completions is drawn as a spot check for review. It never closes tickets itself.

• Archiving is final. For a break there is "Pause".
• Content autopilot. If an AI team member switched on automatic approval of a weekly plan, the autopilot

only approves when the team member may act External on its own; otherwise the post stays in review and carries the note "Approval needed".

• Wording of the AI notice is provisional. The default is a factual placeholder until the final wording

has been decided.

• Maintain the position here. In HR the position appears with the employment type AI and is read-only

there (no editing, no life cycle, no deletion); it is changed through the AI team.

• Without a supervisor. Team members hired before the requirement may still lack a supervisor — the overview

points this out. Set one, otherwise questions and notices go to nobody.

FAQ & troubleshooting

"Only people with the permission … can assign tasks to AI team members." You lack Give tasks to AI team members in the task's brand. Your administrators grant it through the role.

The tab "AI team members" is missing. The switch ai.staff or ai.gateway is off, or you lack View AI team in the active brand. The menu item Team itself appears as soon as you may view the AI team, use team channels or write direct messages.

"The operation … is outside this AI team member's workspace". The team member tried an operation that is not in its workspace. Add the entry under Edit → Workspace if it really needs it.

A role is missing from the list. It contains permissions you do not hold yourself in this brand.

A job profile is missing from the list. Its module is switched off (for example Tickets for support) or hidden for you.

A workspace entry is missing in the draft. The suggested role does not carry it, its module is off, or it does not exist — the reason is listed under Not taken over. Choose a suitable role or add the entry under More details before hiring.

"Too many drafts in a short time". At most 10 drafts per minute; wait a moment.

The team member does not react. Check its status in the profile (paused, budget used up), its working hours and in the History tab whether the entry arrived. If it stands at New there, the background worker ai_staff is not running (RUNBOOK §6) or the switch ai.staff is off.

"Waiting for approval". The team member wanted to take a step it may not take on its own. Which one is written as a comment on the task, in the run under Runs and as a request under Approvals → AI team. After your approval it continues in the same thread; after a rejection it continues without that step and knows your reason.

"The task chain … reached the emergency brake". The chain reached its cost amount. Reply on the task if it should continue; otherwise set the task to Cancelled.

"Ping-pong: … is now with you". A task went back and forth between AI team members too often and is now with you. Decide who takes it over, or answer the open question yourself.

"Not assigned: … levels deep" or "… subtasks are already open". Chain protection prevented another subtask; the AI team member does the part itself or waits for the open parts.

"Could not complete this". The session failed (provider unreachable, missing permission, session limit). The reason is in the history; assign the task again if it should try once more.

"Approval needed: … may not do this without approval". The team member wanted to do something whose permission is With approval — or it wanted to write to a new recipient. The request is under Approvals → AI team.

"This effect is never allowed for the AI team member". The permission of the type is Never; no approval is created. Change the permission if it should do it after all.

A lesson is "awaiting confirmation". It comes from a retrospective with external content or from a decision without approval or management rights. Review it under Learned and confirm or delete it.

"What is learned never changes rights, permissions, budget or workspace". The changed text asks for an extension. Set that under Permissions & budget, not in a lesson.

platform.ki-team · Available from version 0.6.454