Teams use three things to keep humans in the loop with AI agents: mandatory approval gates before the agent takes action, transparent citations that show which documents the agent read, and audit trails that log who approved what and when. The approval gate sits between the agent's suggestion and the action itself—a person sees the proposed change, checks it against context the AI may have missed, and either approves or rejects it. The citation shows the team which source material the agent drew from, so anyone can trace an answer back to the vault document it came from. The audit trail records every decision, making it possible to undo something the assistant filed or restore a prior version of a document. Together, these controls let a team gain speed from AI without surrendering judgment to it.
Key Takeaways
- An approval gate requires a person to review and confirm any action the AI agent proposes before it takes effect, keeping the final decision with the team.
- Transparent citations show which vault documents the assistant read to generate an answer, and any reply that read nothing from the vault is marked that way.
- Audit trails and document history record who changed what and when, with restore, so something the assistant filed can be undone.
- Keeping customer records in their own system—with the AI reading only the shared folders the owner specifies—limits what the agent can see and act on.
- Team chat with @assistant threaded replies brings the assistant into the conversation without moving decisions into a black box.
Why Human Oversight Matters When AI Agents Act on Business Data
An AI agent that can file notes, create tasks, or update records across connected apps becomes useful the moment it saves the team from manual data entry. It becomes a liability the moment it acts on incomplete context, misreads intent, or silently writes over something that mattered. The difference is whether a person checked the work before it became real.
Teams typically expect AI to handle the repetitive parts—summarizing a conversation, drafting a response, pulling facts from past projects—while a human makes the judgment call. That division of labor only holds when the system itself enforces it. An approval gate turns a suggestion into a proposal: the assistant offers to file a decision note or add a task, and nothing happens until a person approves it. The gate is not a speed bump; it is the point where business judgment enters the loop.
Transparent citations show the team what the AI read. Each answer from Mycel's assistant shows which vault documents it consulted, and a reply that read nothing from the vault is marked that way. The citation is not a footnote; it is the evidence a team member uses to decide whether the answer is grounded in the company's own knowledge or needs a second look.
What an Approval Gate Actually Does in Practice
An approval gate is a screen that appears after the AI proposes an action. The proposal includes what the assistant wants to do, which document or task it would create or change, and why it surfaced. A person reads the proposal, decides whether it makes sense, and either approves or dismisses it.
The gate typically appears in three places. In team chat, typing @assistant in a message gets an answer in the thread from the business vault, and the assistant may offer to file a decision note or add a task—nothing happens until a person approves it. In the Approvals section, every pending proposal waits in a single list, with context, so the owner or admin can review a batch at once. In the Today view, proposals that need approval surface alongside other items waiting on a person, with which app it came from and how long it has waited.
Anything Mycel's assistant would change inside another Martello app—Relay, TradesBackbone, ReputeLift, or SynthPrism—waits for an approval first. The assistant reads the folders the owner shares with Relay's and TradesBackbone's AI; ReputeLift's AI reads the business vault; SynthPrism's assistant reads it for brand voice and business facts. The assistant can propose filing what the business learns into the vault, but the proposal appears in Approvals before it takes effect.
The approval gate does not slow down reading. The assistant answers questions from the vault immediately, with citations. The gate only appears when the assistant wants to write, create, or change something.
How Transparent Citations Let a Team Verify What the AI Read
A citation is the list of documents the assistant consulted to generate an answer. Mycel's assistant shows which vault documents it read after every reply, and a reply that read nothing from the vault is marked that way. The citation turns an answer into something a team member can check: open the source documents, see whether the assistant pulled the right detail, and decide whether the answer stands or needs correction.
Transparent citations matter most when the question is specific and the answer needs to be exact. A team member asks the assistant for the warranty terms the company offers, and the assistant answers with a paragraph and a citation showing it read the policies folder. The team member opens that document, confirms the terms, and uses the answer. A week later, someone updates the warranty policy. The next time the assistant answers that question, the citation will show it read the updated document—or, if the update has not yet been filed in the right folder, the citation will reveal the gap.
A reply that read nothing from the vault is marked that way. If the assistant cannot find relevant documents, it says so rather than filling the gap with a guess. That marker tells the team the answer came from the model's general training, not the business vault, and may not reflect the company's actual practice.
Citations also show the team when the assistant is reading too broadly or too narrowly. If the citation list includes folders the person did not expect, it signals that folder permissions or document tags may need adjustment. If the assistant answers a question but the citation shows it missed a key document, the team knows to check the folder structure or the document's metadata.
The Role of Audit Trails and Document History
An audit trail is a record of who changed what and when. Every document in Mycel's vault keeps a history of edits, with the person who made each change and a timestamp. Something the assistant filed can be undone; a document can be restored to an earlier version. The trail does not require manual logging—the system writes it automatically.
The audit trail serves two purposes. First, it makes it possible to reverse a change that should not have been approved. A team member approves a task the assistant proposed, then realizes it duplicated an existing task or misread the project scope. The team member opens the document history, sees the assistant's edit, and restores the prior version or deletes the new task.
Second, the trail shows the team how the vault is changing over time and who is making decisions. The owner can see which folders are actively maintained, which have gone stale, and whether the assistant is filing notes into the right places. The history does not judge; it records.
The whole vault exports as markdown at any time. Each exported file carries its created and updated dates and metadata in a frontmatter header. The export is the current version of each document, not its version history, so the history remains inside the vault where it can be searched and restored, while the export delivers clean files.
How Folder Permissions Limit What the AI Can See
Folder permissions control which people—and which AI assistants—can read and write to a folder. A folder in Mycel can be kept to everyone in the business, to team members but not viewers, or to owners and admins only. The assistant reads only the folders the owner shares with it, so sensitive material can stay in a restricted folder that the assistant never sees.
Relay's and TradesBackbone's AI read only the folders the owner shares with them. Customer records themselves stay in Relay and TradesBackbone, one click away from the vault but not inside it. ReputeLift's AI reads the business vault; SynthPrism's assistant reads it for brand voice and business facts. The business workspace gives the owner a single place to set which folders each connected app can access.
Limiting what the AI can see prevents two problems. First, it keeps sensitive information—financial records, personnel files, contract negotiations—out of the assistant's training context. Second, it reduces the chance the assistant will answer a question by pulling from the wrong document. A folder that holds old, superseded policies should not be shared with the assistant if the team wants answers grounded in current practice.
A business has four roles: Owner, Admin, Member, and Viewer. Folder permissions and role-based access belong to a business workspace, not to Personal or Personal Plus plans. Team chat also belongs to a business workspace—channels (every business has #general) and direct messages between the people in the business, with @assistant available in threads.
What Today and Approvals Show the Team
Today is a list of what is waiting on a person across the connected apps. It shows which app the item came from, why it surfaced, how long it has waited, and a link straight to it. Proposals from the assistant appear in Today alongside tasks, messages, and other items that need attention. The list updates as items are completed or dismissed.
The Approvals section holds every pending proposal from the assistant in one place. Each proposal includes the action the assistant wants to take, the document or task it would create or change, and the context that led to the proposal. A person can approve or dismiss each one, or approve a batch. The section does not hold approvals from other apps—only the assistant's proposals.
Together, Today and Approvals give the owner or admin a single surface to review what the AI is doing without opening each app separately. The Today view prioritizes by how long something has waited; the Approvals section groups by type of action.
Team Chat with @assistant Keeps the Assistant in the Conversation
Team chat in Mycel includes channels (every business starts with #general) and direct messages between the people in the business. Typing @assistant in a message gets an answer in the thread from the business vault, with citations. The assistant may offer to file a decision note or add a task, which does nothing until a person approves it.
The threaded reply keeps the assistant's answer visible to everyone in the channel, so the team sees not just the answer but also which documents the assistant read. If someone disagrees with the answer or spots a gap, the thread is the place to discuss it. The assistant does not take the conversation private or move the decision into a separate interface.
The assistant in team chat is not a separate bot. It is the same assistant that answers in the Ask section, reading from the same vault and subject to the same folder permissions. A question asked in #general gets the same answer—and the same citations—as the same question asked in Ask. The difference is that the team sees it.
How This Approach Differs from Fully Autonomous AI Agents
A fully autonomous AI agent acts on its own, without waiting for approval. It reads a message, decides what to do, and does it—files a note, updates a record, sends a reply. Autonomy means speed, and it also means the team has no gate between the agent's decision and the action.
The trade-off is control. An autonomous agent is useful when the task is so routine and low-risk that a mistake costs less than the time it would take to review every action. Sorting email, tagging support tickets by category, or logging a timestamp are often good candidates. Updating a customer record, filing a warranty decision, or drafting a policy note are typically not.
Teams that want to keep humans in the loop choose systems where the assistant proposes and the person decides. The assistant reads the vault, drafts the note, and shows the draft with context. The person checks it, approves it, or edits it before it becomes part of the business record. The system does not ask the person to build the proposal from scratch, and it does not let the proposal become real without confirmation.
Mycel is built around that model: the assistant answers from the vault with citations, and anything the assistant would file or change inside a connected Martello app waits for approval first. The assistant can work in team chat, responding to @assistant in a thread and offering to file a note, but the offer does nothing until a person approves it. The business workspace holds the approvals queue, the folder permissions, and the audit trail, so the owner controls what the assistant reads and what it writes.
A Comparison of Common Human-in-the-Loop Controls
| Control Mechanism | What It Does | When to Use It | What It Does Not Do | |-------------------|--------------|----------------|---------------------| | Approval gate | Requires a person to review and confirm any action the AI agent proposes before it takes effect | When the AI would create, update, or delete a record, task, or document | Does not apply to read-only queries; the AI answers immediately from the vault with citations | | Transparent citations | Shows which documents the assistant read to generate an answer, and marks replies that read nothing from the vault | On every answer, so the team can verify the source material and trace the reasoning | Does not prove the answer is correct—only that it is grounded in the cited documents | | Folder permissions | Limits which folders the AI assistant can read, keeping sensitive material out of the assistant's context | When the vault holds confidential files, superseded policies, or material that should not inform general answers | Does not restrict which people can share or forward an answer once the assistant has given it | | Audit trails and document history | Records who changed what and when, with restore, so any edit can be undone | After an action is approved, to track decisions over time and reverse mistakes | Does not prevent the original mistake—only makes it possible to fix | | Role-based access | Assigns Owner, Admin, Member, or Viewer roles, controlling who can approve proposals and access private folders | In a business workspace with multiple people | Does not replace folder permissions; roles control who can manage settings, not which documents the AI reads |
Frequently Asked Questions
What happens if the AI assistant proposes an action that is wrong?
The person reviewing the proposal dismisses it, and nothing happens. The assistant's proposal does not take effect until someone approves it, so a wrong suggestion stays a suggestion. The document history also records every approved change, so if a mistake is approved by accident, the document can be restored to the prior version or the change can be undone.
How does the team know which documents the assistant read to generate an answer?
Each answer from the assistant shows which vault documents it consulted, listed as citations after the reply. A reply that read nothing from the vault is marked that way. The team can open the cited documents to verify the assistant pulled the right detail and decide whether the answer stands or needs correction.
Can folder permissions prevent the AI from reading sensitive files?
Yes. A folder can be kept to everyone, to team members but not viewers, or to owners and admins only. The assistant reads only the folders the owner shares with it. Relay's and TradesBackbone's AI read only the folders the owner shares with them; customer records themselves stay in Relay and TradesBackbone, not in the vault.
Does every change the assistant makes wait for approval?
Any action the assistant would take inside a connected Martello app—filing a note, adding a task, updating a record—waits for approval first. The assistant answers questions from the vault immediately, with citations, because reading does not change anything. The approval gate appears only when the assistant wants to write, create, or change something.
How do team members see what the assistant is doing in team chat?
Typing @assistant in a team chat message gets an answer in the thread from the business vault, with citations visible to everyone in the channel. If the assistant offers to file a decision note or add a task, the offer appears in the thread, and nothing happens until a person approves it. The assistant does not take the conversation private or move the decision outside the channel.
What does the audit trail record and who can see it?
Every document keeps a history of who changed what and when. Something the assistant filed can be undone, and a document can be restored to an earlier version. The history is visible to anyone with access to the document, so the team can see how the vault is changing over time and which proposals were approved.
Keeping Humans in Control Without Slowing the Team Down
The goal of keeping humans in the loop is not to add friction—it is to keep judgment where it belongs. An approval gate, transparent citations, and an audit trail together let a team gain speed from AI without surrendering the final decision. The assistant drafts, proposes, and retrieves; the person checks, approves, and corrects. The division of labor holds only when the system enforces it, and the system enforces it by making every action visible and every proposal subject to confirmation before it takes effect.
For more on how Mycel works and the difference between Personal and Business workspaces, see pricing. The Business plan at $119 per seat per month for up to 20 seats ($1,190 per seat per year) includes everything in Personal per seat, plus connecting TradesBackbone, Relay, ReputeLift and SynthPrism, Today, team roles, private folders, document history with restore, and automations.