Placed
Every item that reaches that review point on the map.
- Example
- Every vendor bank change, before the bank field is edited.
- Where you set it
- The workflow map, as its own box.
A human review step pauses the agent wherever policy says a person must decide. The agent waits; it never guesses and never times out into proceeding. The reviewer sees the agent's record beside the source document, then approves, rejects with a reason, corrects a value, or supplies a missing fact. The agent resumes on the branch the reviewer chose, and the decision joins the run's record.
The problem
Releasing money. Changing where money goes. Signing for the company. Turning someone down. An agent can be right about all of them and still should not be the one to decide.
Today most of those approvals happen in email. A clerk forwards a PDF, a manager replies "approved," and the evidence lives in three inboxes. The reply cannot show what the manager actually looked at. It can be forwarded, forged, or lost.
LaunchAI treats review as a step type: placed on the map where everyone can see it, given a screen designed for the decision, and recorded down to what the reviewer saw and chose.
A signature is not the same as a review
On August 11, 2020, Citibank meant to send about $7.8 million in interest to Revlon's lenders and instead wired roughly $894 million, the full outstanding principal. Three people reviewed and approved the transaction. The entry lacked settings that would have kept the principal from being sent, and none of the three caught it. Lenders holding about $500 million refused to return their share, and the case ran until the Second Circuit ruled for Citibank in September 2022.
How a review runs
Whether the decision comes in four minutes or the next morning, the run continues as if no time had passed, and the record shows exactly how much did.
It reaches the review step and holds its place in the chain.
It lands in the review queue, and the assigned reviewer is notified.
The agent's record on one side; the source on the other: the PDF, the scan, the portal screenshot.
A corrected value writes through to the shared table, so the agent, the reports, and the next review all see it.
A rejection requires a reason, and the reason becomes part of the record.
On the chosen branch, from exactly where it stopped.
Four doors
A placed review step is the obvious door. There are three more, and a well-built workflow uses all four.
Every item that reaches that review point on the map.
A decision rule sends some items to review and lets others pass.
A model reading falls below the step's threshold.
The agent cannot finish a step safely: an ambiguous target, a missing fact, a write whose result is unknown.
The reviewer workspace
Four decisions
Confirms the prepared action.
Continues on the approved branch.
Stops the action and says why.
Follows the rejection branch to its own endpoint.
Fixes one field, such as a misread quantity.
Continues with the corrected value, written to the shared table.
Adds what the agent could not find, such as a job number left off an invoice.
Continues with the fact in place.
Speed
A review that takes longer than the task is a bottleneck with extra steps. The workspace is built against that.
The agent has already read, matched, and calculated. The reviewer checks a conclusion, not a stack of paper.
Everything needed to decide is on one screen.
Fields are searchable and layouts switchable.
The next review is one key away.
Low-risk items can be approved in a batch, where policy allows.
Reviews can be taken in the Portal or on a phone.
Assignment follows role, region, or amount, and a review that waits too long escalates on its own.
Placement
Placement map
How that framework translates into LaunchAI's four doors. Amounts are illustrations; yours live in a table.
| Action the agent would take | Can it be undone cleanly? | Pattern in LaunchAI |
|---|---|---|
| Convert approved recommendations into purchase orders | Partly; a sent PO reaches the supplier | Placed review before conversion, reviewed as one list |
| Raise a customer's credit limit | Yes, but exposure starts at once | Rule-routed review above the band in the approvals table; below it, no pause |
| Post an inventory adjustment | Only by another adjustment | Rule-routed review outside the count tolerance |
| Submit a pay application to an owner's portal | Placed review, every submission | |
| Send an adverse action notice to an applicant | Placed review, every notice | |
| Update a vendor's remit-to address | Not in practice, once a payment goes out | Placed review, every change |
| Read an invoice and fill a coding worksheet | Confidence gate only |
Routing and escalation
The assignment is a lookup like any other: a table the business owns.
| Condition | Assigned to (illustration) |
|---|---|
| Amount up to $4,999.99 | AP specialist for the entity |
| Amount $5,000.00 and above | AP lead |
| Property in the Southeast region | That region's property accountant |
| Vendor master change of any kind | Controller |
You set how long an item may wait before it moves to the next person on the assignment path, and alerts go out by the channels you choose: the Portal, email, or chat. A typical ladder: the assigned reviewer first; after four business hours, the team lead; by the next morning, the controller. The agent holds its place the whole time.
Batch approval
A speed feature, and also the fastest route to a rubber stamp. Decide per review step whether to permit it.
Batch only items that stopped for the same reason. A batch that mixes reasons hides the one item that stopped for a different one.
Segregation of duties
Accounting has long required that no single person both create and authorize a consequential transaction. NIST's SP 800-53 carries the same idea as control AC-5, separation of duties.
LaunchAI divides the work into separate rights: building workflows, approving them for production, editing rules and tables, reviewing items, running, viewing history, and managing credentials. The backend enforces each one. Three splits matter most for review:
Sensitive actions, such as approving a payment workflow, can demand step-up authentication: a fresh sign-in or a second factor at that moment.
Handoff
Handoff runs both ways. The agent hands a person a decision with the context attached, then takes the work back the moment the decision is made.
A person can also step in mid-run, do part of the job by hand, and hand back: a portal asks for a one-time code only a person can supply, or a record needs a judgment the workflow never anticipated. The agent resumes from the screen as the person left it. No values are copied between windows, and nothing has to be explained twice.
Worked example
Fraudsters target this workflow because one unverified change redirects the next payment run.
The FBI's IC3 counted $55.5 billion in exposed losses across 305,033 business email compromise incidents between October 2013 and December 2023, and advises verifying account-change requests through a second channel. The pressure has not eased.
Since June 19, 2026, Nacha's fraud monitoring rules require every non-consumer ACH originator to keep risk-based processes reasonably intended to identify entries initiated due to fraud, reviewed at least annually. A forged bank-change request fits Nacha's definition of false pretenses. A documented, recorded call-back is one part of such a process; it does not by itself make a company compliant.
Manufacturers verify supplier banking changes and property managers verify vendor banking changes, often many times a day. Here is the chain:
Whether an email, a PDF form, or a portal notice: vendor, new bank, routing and account numbers.
Capture the current bank details and the phone number on file. That number comes from the vendor master or the original contract, never from the request, because whoever sent the request controls every phone number and email address in it.
No exception for amount, vendor tenure, or urgency. Because it is a rule, nobody can argue it away on a busy Friday.
With the number on file, the request, and the prior account side by side.
The reviewer calls the number on file, logs who answered, when, and what they confirmed, then approves or rejects. The review sits after the call-back task, so the change reaches a reviewer only once the call is logged.
On approval, the agent edits the bank fields, captures the screen before and after, and notifies accounts payable. On rejection, it records the reason and routes the request to whoever handles suspected fraud.
A Action D Decision H Human review
Step 5
Variations
Insert a review step after step 5, assigned to the controller, so the person who made the call is not the only one who approved the change.
A rule flags the first payment to any account changed in the last 30 days, and a placed review confirms it before release. The window lives in a table.
The reviewer places the call; an inbound call “from the vendor” never counts. A cloned voice can answer a phone, but it cannot choose which number your reviewer dials.
The review record
Who looked, what they saw, what they changed, what they decided, when, and how long the item waited. The record is filed beside the agent's own screenshots in the audit trail, so an auditor sees the machine's work and the person's judgment in one place.
That record answers the three questions a control test asks. Each answer is a field, not a recollection.
What to measure
Track them per review step, because each step guards something different. Override rate and queue budget are defined in our human-in-the-loop guide; these are the operating measures around them.
| Measure | How to compute it | What it tells you |
|---|---|---|
| Items per step per week | Count of items reaching each review step | Whether a step's volume fits the people assigned to it |
| Decision mix | Share approved, rejected, corrected, and supplied | Which review steps are doing real work |
| Wait by reviewer and hour | Median and slowest decile of wait time, grouped | Where escalation or a second reviewer is needed |
| Escalations fired | Items that moved up the ladder ÷ items | Whether the first assignee is overloaded or absent |
| Corrections by field | Corrected values, grouped by field name | Which reading or rule to fix next |
| Batch share | Items approved in batches ÷ items approved | Whether batching stays within policy |
| Takeovers | Mid-run takeovers per 100 runs | Where the workflow meets situations it was not built for |
Connections
Governance and security
Limits
Review is not free, and we don't pretend it is.
An item that pauses for a person finishes later than one that does not. The remedy is placement, assignment, and escalation, not removing the review from steps that move money.
It can make the evidence hard to skip and the decision impossible to deny afterward. Attention is still the reviewer's job.
For a simple yes or no on a flow that already runs through APIs. LaunchAI can run that same approval, plus the screen work around it, with the evidence and the record in one place.
The goal is fewer, better reviews: rules for the common cases, people for the ones that need judgment.
No. Reviews are handled in the Portal, the shared web surface, from a browser or a phone. Only the machines that run agents need the desktop application.
No. They are two different approvals. A review step is a decision inside a run: this change, this record, today. Approving a workflow version is a decision about the procedure itself, made once before it goes to production. An approved version is read-only, so the chain a reviewer works inside cannot shift under them.
Reject it with a reason, and let the rejection branch carry it to whoever handles suspected fraud. Do not correct values taken from a suspect document; a corrected field on a forged request is still a forged request. The record keeps the document as it arrived, which is what an investigator will ask for.
No. A correction changes the value for that record in the shared table. Rules change only when someone with the right to edit rules and tables changes them. A correction that repeats every week is the evidence that a rule should.
The permissions allow it, and your policy decides. For workflows that move money or change master data, keep them apart: the builder shapes what the agent does, so a second person should judge what it produced.
No prompts. No babysitting. No dumb questions.