The ten AI step types
generateText, classifyText, extractInformation, summarizeText, aiAgent, generateImage, documentOcr, invoiceParser, expenseParser and extractDocumentData.
AI & human decisions
AI handles the routine. Your people decide the exceptions.
Ten AI step types read invoices, classify requests, summarize, draft and call your other systems inside the workflow. A rule you set decides which cases need a person. That case waits for the assigned person or role, with the context in front of them, and the same run continues from their decision.
30-day trial with Pro capabilities · no card required
- 10
- AI step types, from document extraction to agents with tools
- 4
- human-in-the-loop step types: two pause the run for a decision, two create and update tasks
- Per field
- confidence scores on invoice, expense and structured document extraction
- 0
- database tools in an agent's toolset. Records change in the workflow's data steps, with business rules enforced in the write.
Anatomy of a decision
What the AI did, why the run stopped, and what happened next.
A supplier invoice is recorded with its PDF. The invoice parser reads it field by field, and the total comes back below the confidence the process requires. The same run waits for Accounts payable, then posts the invoice with their correction.
accounts_payable / invoice-intake · run 4821 Diagram
-
Data change trigger
An invoice is recorded
A new row in invoices, with the supplier and the PDF, starts the run.
-
Invoice parser
AI reads the invoice
Each extracted field comes back with its own confidence score.
- Supplier
- Northwind Metals
- 0.98
- Invoice no.
- INV-20931
- 0.97
- Due date
- Oct 24, 2026
- 0.95
- Total
- €18,420.00
- 0.61
- Below threshold: needs a person
-
If
A rule checks the threshold
Any field below 0.90 confidence goes to a person. Here, only the total does.
- All fields ≥ 0.90 → post automatically
-
Interactive form (a person decides)
The run pauses for Accounts payable
Assigned to a role, with the context the reviewer needs.
Assigned to Accounts payable lead (role)
Decision required
Why the run paused: The invoice total was read with 0.61 confidence, below the 0.90 this process requires.
Northwind Metals · INV-20931 · due Oct 24, 2026 · the extracted total as an editable field
Permitted decisions
- Approve the total as read
- Approve with a corrected total
- Reject the invoice
A Switch on the submitted decision picks the next branch of this same run.
No answer in 48 hours: escalate to the finance manager.
-
Transaction
The same run posts the invoice
The corrected total, the approved status and the line items are written in one transaction, with each table's business rules checked inside that write.
-
Record history
The change is on the record
Written in the same transaction as the change, with the actor and the before-and-after values. The reviewer's decision stays on the completed task.
- actor
- Automation: Invoice intake
- table
- invoices
- total
- ∅ → €18,470.00
- status
- received → approved
Run 4821, start to finishRecorded → read → flagged → decided → posted → audited
Diagram built from step types in the product: a data change trigger, the invoice parser with per-field confidence, an If rule, a role-assigned interactive form with an expiry branch and a Switch on the decision, a transaction, and the row audit that write produces. Names and values are sample data.
Want to see this on your own process? Our team will walk through it with you.
Walk through your process with usWhat it does
What the AI does, and where your people decide.
-
Ten AI steps, each with a defined job
Generate and summarize text, classify it, extract information and read documents, each as its own workflow step. Each step works on the inputs you map to it and returns output the next step checks or routes on.
- Text generation and summaries
- Classification and information extraction
- OCR and structured document extraction
- Invoice and expense parsing
-
Confidence you can route on, field by field
Invoice, expense and structured document extraction return an overall score and a score for each extracted field. An If step compares those scores with the threshold you set. Invoices that clear it continue on their own, and an invoice with an uncertain field goes to the assigned reviewer, instead of someone checking every invoice or none.
-
Agents act only through their configured tools
An AI agent step works with the tools configured on it. By default that is HTTP requests, sandboxed JavaScript and the output of earlier steps, and you can remove any of them from the step's tool list. It has no database tool: records are read and written by the data steps you place around it, with business rules enforced in the write. Outgoing email goes through the workflow's Email step, placed where you want it.
-
Ambiguous matches go to a person
When an incoming name, email or reference matches more than one record, the resolve-record step assigns the choice to the person or role you set, with a preview of each candidate. A single match on a strong signal, such as an exact reference, resolves on its own. The run continues on its resolved, unresolved or expired branch.
-
Follow-up tasks that don't hold up the run
A create-task step puts follow-up work in My Work for one user or role and can link it to the record it concerns. The run keeps going. A later update-task step reassigns, completes or cancels that task.
The decision stays inside the run
An interactive form assigns the decision to a user or role and shows the reason, context and choices you configure, filled from the run's data. It can show AI output as editable fields, so the reviewer corrects one value instead of rejecting the work. A Switch on the submitted decision picks the next branch of the same run, and with a deadline set, an unanswered task takes its expired branch.
client_delivery / workflows / prepare-a-client-status-update Product · captured Sep 2026
How AI runs in Bondi
AI works inside the rules.
Every AI step runs inside a workflow your team designed, with the inputs you map and the tools configured on each step. Records change in the data steps around it, under business rules enforced in the write.
- Agents act only through the tools enabled on their step
- 11 rule types enforced in the write transaction
- A rule you set pauses the run for the assigned person or role
Under the hood
The mechanisms, in detail.
The mechanisms behind the page, for the architects and security reviewers evaluating it.
Read the technical documentationWhere confidence comes from
Invoice, expense and structured document extraction return confidence.overall and confidence.perField, with a score for each extracted field. Classification, summaries, extractInformation and extractDocumentData in flexible mode return no score: route those on rules and on the output itself, such as an empty required field.
Agent tools
http_request, run_javascript in an isolated QuickJS sandbox with a one-second limit, and read_workflow_data, chosen per agent step. send_email is off by default and cannot be turned on from a workflow; send email with the workflow's email step. There is no database tool.
Human-task lifecycle
interactiveForm suspends the run until the assigned task is submitted or expires, with separate submitted and expired outputs. resolveRecord ends as resolved, unresolved or expired. createTask assigns work without suspending the run.
Designed boundaries
Where it starts and where it stops.
-
The agent's HTTP tool is on by default and is not limited to a fixed list of endpoints. Remove it from agent steps that don't need outside calls.
-
Workflow data steps run under the workflow's own automation identity, which role, row and field permissions do not limit. Business rules still apply in the write. A workflow launched by hand can run as the person who launched it, under their permissions.
How access control works -
You decide where a person is required. Place consequential writes after the decision that authorizes them.
How workflows are built
Availability
- Standard
- Every AI and human-in-the-loop step, with monthly AI credit
- Pro
- More AI credit and Automation Runs, plus a Development environment to build and review workflow changes before they reach Production
- Enterprise
- Custom capacity and commercial terms
Questions buyers ask
Straight answers.
Can the AI act on its own?
Only through the tools configured on the agent step and the steps you place around it. By default the agent can make HTTP requests, run sandboxed JavaScript and read earlier steps' output, and you can remove any of these from the step; it has no database tool. Records change in the workflow's data steps, where business rules are enforced in the write transaction, and you decide where the run stops for a person.
Which AI steps return confidence scores?
Invoice, expense and structured document extraction return an overall score and a score for each extracted field. Classification, summaries, information extraction and flexible document extraction return none, so route those on rules and on the output itself, such as an empty required field.
What happens if nobody responds to a decision?
With a deadline, from one hour to 365 days, the task takes its expired branch, such as escalating to another role or ending the run. Without a deadline, it waits until someone completes or cancels it.
Can a reviewer correct the AI's output?
Yes. An interactive form can present extracted values as editable fields. The reviewer corrects them, and the run continues with the corrected values.
How is AI usage billed?
AI steps draw on the monthly AI credit included in every plan. Text, agent and flexible document steps are charged by usage at the AI level you choose for the step. OCR, invoice, expense and structured extraction are charged per document or page, and image generation per image. Usage beyond the credit comes from your Wallet.
Next step
Show us the work your team checks by hand.
We'll map it with you: the AI steps, the rule that stops the run, the role that decides, and what the run writes. Your process, not a demo script.