Back to all articlesAI action safety

Decide Which AI Tools Can Run Read-Only and Which Need Write Approval

Alactic Editorial9 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

Consider a hypothetical assistant asked to find a customer’s latest conversation. Its tool returns the conversation—but also marks it as read, updates a last-contacted field, or creates a follow-up task. The request sounded informational. Its effects were not.

To classify AI tools as read or write, assess the proposed action’s observable effects: what data it accesses, what state it persists, and what it exposes or triggers beyond the authorised interaction. Allow verified, narrowly scoped reads within existing permissions. Require approval for defined changes or disclosures. Keep actions unavailable when their effects are unknown, their scope cannot be bounded, or they exceed the requester’s authority.

The unit of classification is the action with its arguments and execution context, not the tool name. A tool may support both lookup and consequential updates. If those modes cannot be separated reliably, do not grant the assistant unrestricted access to the whole tool.

1. Inspect effects before assigning a permission label

Start with a description someone else can verify. “Use customer search” is too vague. “Retrieve the latest conversation for this authorised customer record without changing conversation state” establishes a checkable boundary.

For each proposed action, answer four questions:

  • Data access: Which records, fields, attachments, and derived information can it retrieve? Is that access necessary?
  • Persistent effects: What remains changed afterward, including drafts, tasks, ownership, read status, files, or queued work?
  • External effects: Does another person or system receive data, a notification, a message, or an instruction?
  • Authority: May the requester perform this exact action on these exact objects? Is additional approval required?

Separate routine technical logging from business-state changes, but do not ignore its audience. An access log need not make every lookup a business write. Copying a full customer record into a broadly accessible log can nevertheless create an unacceptable disclosure.

Evidence may include implementation documentation, permission definitions, and controlled tests with fictional records. A reassuring tool description is not enough. If the evidence is missing, mark the action unavailable pending verification rather than assuming it is read-only.

2. Use three execution decisions, not a read/write toggle

Read-only within an authorised scope

This means no business-state mutation, no separately created business artifact, and no unapproved disclosure or downstream action. A permitted lookup returned to an authorised assistant interaction may qualify even when that interaction is routinely retained.

Assess conversation retention and audience separately: who can inspect the interaction, where it is stored, and whether its retention is appropriate for the retrieved data. “Only in the chat” is not evidence that information is temporary or private. If returning the result introduces a new disclosure, the lookup does not qualify for this read-only path.

Read-only is still permission-bound. OWASP’s object-level authorisation guidance calls for authorising access to each requested object. Knowing a record identifier does not establish permission.

Approval required for a bounded effect

Use this decision when the intended change or disclosure is understood, permitted, and reviewable. The approver must see the target and proposed effect, not merely a “Continue” button. Saving a draft and sending a customer message require different approval details.

Unavailable to the assistant

Use this decision for unknown side effects, excessive scope, prohibited destinations, or actions beyond the requester’s authority. Approval cannot substitute for underlying permission.

This framework follows the least-privilege, deny-by-default, and per-request checking recommendations in OWASP’s authorisation guidance. These are design recommendations, not evidence of controls in a particular product.

3. Filled decision table: a hypothetical workspace

Fictional worked example: A sales assistant supports staff handling customer records. The table below is a filled, reusable classification artifact for that hypothetical workspace, not a description of existing Alactic controls. Its notification and routing effects are scenario assumptions that another team must verify before reusing the decisions.

Hypothetical action classification and execution requirements
ActionData accessPersistent effectsExternal effectsExecution decision and approval requirement
Record lookupOne permitted record: contact name, owner, latest conversation summary.No business changes; ordinary authorised interaction retention and access logging only.Result stays within the already authorised interaction audience; no notification.Read-only after verification. Check record and field access on every request. No additional write approval.
Saved draft creationSelected customer context and requester instructions.A separate draft is saved against the customer record.Draft becomes visible to authorised colleagues; no notification or customer delivery in this scenario.Approval required for content, linked record, and draft visibility before saving.
Task creationCustomer record, proposed assignee, relevant context.A task and due date enter the work queue.Assignee receives a notification.Approval required for linked record, task text, assignee, due date, and notification.
Owner changeAffected record, current owner, proposed owner.Record ownership changes.Assignment notification and routing change.Authorised approver must confirm old and new values and known consequences. Unavailable if consequences remain unknown.
Scoped exportSpecified records and explicitly selected fields.A separate file is created.Data enters a named destination with its own audience.Approval required for filters, record count, fields, destination, and audience. Unrestricted or unbounded export remains unavailable.
Message sendingRecipient details, relevant context, permission and opt-out information.An outbound message record or send request is created.Customer communication is attempted; delivery remains a separate outcome.Approval required for recipient, channel, exact content, attachments, conditions, and timing. Missing permission or an applicable opt-out blocks sending.

Boundary conditions: A lookup that marks a conversation as read is not the lookup classified here. Text generated in the current interaction is non-committing preparation only if it creates no separate saved draft, triggers no notification, and introduces no new disclosure. Its conversation retention and audience must still be acceptable. Suggesting a follow-up does not authorise task creation, and selecting a template does not authorise sending.

Acceptance checklist for each row:

  • Evidence supports the stated access, persistent effects, and external effects.
  • The record scope and interaction audience are authorised.
  • Every committing effect has a concrete approval preview and an authorised approver.
  • Unknown effects or unenforceable boundaries keep the action unavailable.

To reuse this artifact, replace each scenario assumption with verified behaviour. Preserve distinctions between actions rather than assigning one permission label to the whole assistant.

4. Bind approval to the exact action

An approval request should describe a transaction, not grant a vague mandate. Bind it to the action type, target record or record set, proposed values or content, destination, affected audience, requester, authorised approver, and execution conditions, including expiry.

Continue the hypothetical example: a salesperson asks the assistant to find a customer’s open issue, draft a reply, and create a follow-up. The verified lookup may run within scope. The assistant can then present proposed reply text within the authorised interaction, without saving it. Task creation remains blocked until approval.

A filled approval preview might read: “Create a task linked to fictional customer Cedar Workshop: ‘Review the unresolved delivery question.’ Assign to Maya Chen, due 14 May 2027 at 10:00 Europe/London. This creates a work-queue entry and notifies Maya. It does not save or send the proposed customer reply. Approval expires before the stated due time or if these details change.” All names and dates in this example are hypothetical.

The observable outcome of that approval is one specified task and its disclosed notification—not permission to save the reply, change ownership, or send a message. If the assignee or task text changes, seek approval again.

Check permission again at execution. An earlier approval must not override revoked access or changed circumstances. Where a change depends on a displayed current value, verify that value has not changed before committing. Otherwise, an approved reassignment could overwrite a legitimate later decision.

5. Separate preparation from commitment

Break compound requests into distinct stages: inspect permitted context, prepare a proposal, obtain approval for concrete effects, then execute only the approved action and verify its result. Permission does not automatically carry from one stage to the next.

If a tool inseparably combines searching, saving, and notifying, apply the boundary required by its most consequential effect. Do not call a bundled “find and follow up” operation read-only because searching happens first. Either narrow the available action or require the appropriate approval for the whole verified bundle.

For WhatsApp, keep campaign preparation, template review, and actual delivery as separate states. Recipient permission matters, and opt-outs must be honoured. Approved templates are used to initiate Platform conversations and outside the 24-hour customer-service window, which opens or resets with a user message. Template approval does not guarantee delivery. Check the current official policy for precise requirements, and provide a clear human escalation path for automation.

Apply these distinctions to the product actually in use. Alactic offers CRM and WhatsApp business communications products; ACID supports business-assistant workflows. Do not assume that a CRM action exists in Communications or that a web capability exists in the Android app. Check the workspace and current plan for availability. The classification method defines operating boundaries, not exact interface steps or proof that a proposed control exists.

6. Test the failure cases that names conceal

Use fictional records in a controlled environment to test each proposed classification. Compare relevant state before and after execution, including notifications, files, queues, and downstream changes—not just the assistant’s visible response.

  • A lookup changes read status: Isolate a verified non-mutating retrieval mode or reclassify the action as a write.
  • A draft triggers a notification: Include that effect in the approval preview. Keep unsaved preparation separate if notification is unnecessary.
  • A task routes work elsewhere: Assess the resulting assignment and notification, not merely the new record.
  • An export includes extra fields: Stop until field selection is enforceable and reviewable.
  • A request accepts another record identifier: Verify object-level permission rather than trusting possession of the identifier.
  • An approved request times out: Establish whether it completed before retrying.

These tests can expose a mismatch between the intended boundary and actual behaviour. Record the mismatch as a reason to narrow, reclassify, or disable the action—not merely to revise its description. Reduce approval friction through narrower actions and clearer previews, not by hiding effects.

7. Recover without compounding the original effect

When an unexpected change occurs, stop further execution of the affected action. Record the requested operation, actual target, approval evidence, and observed result. Establish what changed before proposing repair.

Do not blindly repeat an operation after an ambiguous response. A task may already exist; a message request may already have been accepted. Check resulting state or operation status where available. If completion cannot be established, escalate rather than treating uncertainty as failure.

Reversal also requires authority. Restoring an owner must not overwrite a legitimate later change. Deleting an exported file does not establish that no copy remains. A correction message is a new external communication, not an automatic undo.

Update the affected table row with the discovered effect. Keep the action unavailable until its boundary can be verified and any necessary approval requirement can be enforced.

AI assistantsAuthorisationWorkflow designData access

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications