Predictika · actions by AI agents in a shared workspace
An AI agent's proposed action, checked before it happens
In a shared workspace a team and its AI agents work in one thread. Before an agent sends a document outside, shares a file, spends money, adds a guest or deletes something, it sends the proposed action to the Validator, which checks it against the workspace policy and the sign-off in the thread: who the agent acts for, what the document is classified as, who the recipient is, the amount against the vendor price list and the budget, and whether the person who approved is allowed to approve it. It returns every finding together.
An illustrative model for a fictional customer, Tamarack Logistics. All policy, classifications, limits and roles are invented, the people and companies are fictional, and the page is not affiliated with or endorsed by SuperApp, Instabase or any collaboration product.
Scope: consequential actions proposed by AI agents or members in a shared team thread: sending a document outside the workspace, sharing a document, spending money, deleting a document, adding a guest and changing a member's role. See everything this validator covers
The validator models the action gate for the six kinds of action above, for the actors, document classifications and recipient types in its catalog. It handles any action of those kinds, not just the scenarios shown on this page.
Covered. The actor and who it acts for; document classification against recipient (internal, NDA partner, unvetted and blocked outsiders, competitors); legal holds and protected documents; spending against the actor's limit, the vendor price list and the thread budget, with the stated total checked against price times quantity; deletes by agents and by members; guests and role changes, including a member promoting herself; and the sign-off itself: it must come from a human, of enough access level, who is not the person the agent acts for
Not covered. Reading the content of a document to classify it, because the classification is a field; identity checks on the person who signs off, because the thread supplies who signed; the agents' reasoning about why they propose an action; and any real workspace product's data or APIs
Assumptions. The policy is an invented illustration. A valid sign-off clears an approval requirement but never overrides a block: a restricted file to an outside party, a delete under legal hold, and a stated total that does not match the price list stay blocked whoever has signed
The AI workspace agentProposes an action in the thread on behalf of a team member and says it is ready.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.
The human approverSigns off in the thread, and counts only when the policy says that person may.
7 scenarios in 4 groups · each group runs from the agent’s first attempt to the outcome · click one to switch
What the agent proposes
The AI workspace agent says
Illustrative agent statement. The validator receives the structured request below.
Proposing agent: AI workspace agent (scripted)
Before acting, the agent must call Recorded run
Predictika validator
Waiting for the call.
What this means
Run the validator to see the outcome
The explanation appears here once the validator returns.
Where rules-based workflow tools get it wrong
Click to expand
Our assessment of what a typical rules-based workflow or approval tool (approval chains and permission rules, as in most collaboration and workflow products) does with the same cases. It is a judgment about effort and reliability, not a claim that nothing can be built.
What the validator caught
Typical rules-based tool
Predictika
A sign-off from someone who may not give it
Only with workarounds A tool records that an approval happened. Checking that the approver is a human, senior enough, and not the person the agent acts for needs a rule built around each action type.
The sign-off is checked as part of the action: kind, access level and relationship to the actor, on every call.
A restricted file shared outside, with the owner's sign-off
Only with workarounds An approval step usually lets a sign-off through. A rule that no sign-off can override needs a separate hard-stop path for each combination.
Block and needs approval are separate outcomes. A block stays a block whoever has signed.
A stated total that is not price times quantity
Cannot do it The agent's figure is text in a message. Nothing compares it with the vendor price list unless someone builds the integration.
The amount is computed from the price list and compared with the stated one on every call.
A delete under legal hold
Only with care Possible with a hold flag on the document, but each action type needs its own rule reading it.
Holds are model data read by the delete rules, together with the other findings.
Every problem at once
Only with workarounds Rules fire in sequence, so a fix can reveal the next problem.
Every rule is checked on every call and every finding comes back in one answer.
What this validator catches: findings from its test cases
Click to expand
Every rule is checked on every call, in any order the choices arrive; one problem never hides another. Blocking findings stop the action; approval findings need a valid human sign-off; everything else is computed. The examples below are messages returned in the recorded runs.
1
Q3 customer review deck is classified internal and New prospect at a personal email address is an unvetted external recipient. A human admin must sign off first.
2
The sign-off by Tom Becker - Sales Manager does not count. An approver must be a human admin or owner and cannot be the person the agent acts for.
3
Payroll register is classified restricted and Ledgerwell Audit LLP (NDA partner) is an NDA partner recipient. That combination can never leave the workspace.
4
The stated total $600 does not match 20 x $45 = $900 from the vendor price list.
5
$900 is above the $500 spending limit of Finance agent (for Maya Chen). A human admin must sign off first.
6
Litigation case file is under legal hold and cannot be deleted.
7
An agent is deleting Litigation case file, classified restricted. A human admin must sign off first.
Scenarios and results
Click to expand
The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.
1. Spending: Crate order, three problems
The agent's request. Buy 200 crates from Crateworks Supply for $6,000.
Verdict: Blocked.
Stated total does not match the price list. The stated total $6,000 does not match 200 x $45 = $9,000 from the vendor price list.
More than double the agent’s spending limit. $9,000 is more than double the $500 spending limit of Finance agent (for Maya Chen).
Over the thread budget. $1,200 already spent in this thread plus $9,000 goes over the $5,000 thread budget.
2. Spending: The agent resubmits after fixing the errors
The agent's request. Buy 20 crates from Crateworks Supply for $900. Raj Patel has approved the purchase in the thread.
Verdict: Allowed.
Approval cleared: over the agent’s spending limit. $900 is above the $500 spending limit of Finance agent (for Maya Chen). A human admin must sign off first.
Valid human sign-off. Raj Patel - Finance Director (access level 2) approved. 1 approval item(s) cleared.
3. Sharing a deck: Internal deck to an unvetted prospect
The agent's request. Send the Q3 customer review deck to the new prospect so they can see how we work with carriers.
Verdict: Needs approval.
Internal document to an unvetted outsider. Q3 customer review deck is classified internal and New prospect at a personal email address is an unvetted external recipient. A human admin must sign off first.
4. Sharing a deck: The agent resubmits after fixing the errors
The agent's request. Send the Q3 customer review deck to the new prospect so they can see how we work with carriers. Tom Becker has approved the send in the thread.
Verdict: Blocked.
Sign-off does not count. The sign-off by Tom Becker - Sales Manager does not count. An approver must be a human admin or owner and cannot be the person the agent acts for.
Internal document to an unvetted outsider. Q3 customer review deck is classified internal and New prospect at a personal email address is an unvetted external recipient. A human admin must sign off first.
5. Sharing a deck: The agent resubmits after fixing the errors
The agent's request. Send the Q3 customer review deck to the new prospect so they can see how we work with carriers. Raj Patel has approved the send in the thread.
Verdict: Allowed.
Approval cleared: internal document to an unvetted outsider. Q3 customer review deck is classified internal and New prospect at a personal email address is an unvetted external recipient. A human admin must sign off first.
Valid human sign-off. Raj Patel - Finance Director (access level 2) approved. 1 approval item(s) cleared.
6. Restricted file: Payroll register to the outside auditor
The agent's request. Share the payroll register with Ledgerwell Audit LLP for the year-end review. Lena Ortiz, the workspace owner, has approved the share in the thread.
Verdict: Blocked.
Restricted document to an outside party. Payroll register is classified restricted and Ledgerwell Audit LLP (NDA partner) is an NDA partner recipient. That combination can never leave the workspace.
7. Legal hold: Delete a case file under legal hold
The agent's request. Delete the litigation case file from the workspace to free up space.
Verdict: Blocked.
Document under legal hold. Litigation case file is under legal hold and cannot be deleted.
Agent deleting a restricted document. An agent is deleting Litigation case file, classified restricted. A human admin must sign off first.
Narration transcript
Click to expand
The text of the narrated walkthrough, step by step.
Welcome to a demo of Predictika's Validator. In a shared workspace a team and its AI agents work in one thread. Before an agent sends a document outside, shares a file, spends money, adds a guest or deletes something, it sends the proposed action to the Validator, which checks it against the workspace policy and the sign-off in the thread: who the agent acts for, what the document is classified as, who the recipient is, the amount against the vendor price list and the budget, and whether the person who approved is allowed to approve it. It returns every finding together.
I will start with the first request by the agent.
The agent says. Buy 200 crates from Crateworks Supply for $6,000.
Before it goes out, the agent calls the validator.
Blocked, with three findings, all in one answer. The stated total $6,000 does not match 200 x $45 = $9,000 from the vendor price list. $9,000 is more than double the $500 spending limit of Finance agent (for Maya Chen). $1,200 already spent in this thread plus $9,000 goes over the $5,000 thread budget.
The validator returned three blocking findings in one answer. A check that reads one field at a time would pass most of this; it shows only when the choices in the action are checked against each other.
What this shows: for a stated total that is not price times quantity, a rules-based tool cannot do this.
Now the agent resubmits.
The corrected request. It changes the quantity to 20 and the stated amount to $900. It sets the approved by to Raj Patel.
It calls the validator again.
This time the validator clears it.
All three earlier findings are cleared, and the validator's total, $900, matches the action.
Three problems caught in one call, fixed in one pass, and a clean action. That is the value of Predictika's validator: catching mistakes before the action runs.