What is AI RPA?
AI RPA (also called agentic RPA or agentic automation) is software that completes a business task by understanding it, rather than by replaying fixed steps. Classic RPA repeats what someone recorded; an AI agent reads the input and the screen, acts, and checks the result.
The term is used loosely. Much of what is sold as intelligent automation is classic RPA with one AI step added: a model reads the document and hands fields to the same script as before. The part that breaks most often on legacy systems, the screen logic, is still a recorded sequence someone maintains.
Classic RPA
A bot replays recorded steps, finding fields by selectors or coordinates. Validation and exceptions are whatever the developer built.
Intelligent automation: RPA plus an AI step
An AI model reads the document and returns data, often sold as RPA AI. The screen work is still a script, and checking the values is still your job.
Agentic RPA
One AI agent does the whole job: reads and checks the document, fills the screens by meaning, reads the host's reply and confirms the record.
Classic RPA vs AI RPA vs legacy bridge
General-purpose RPA platforms are capable toolkits; with them you build and maintain the screen logic, validation and exception handling yourself.
| Criterion | legacy bridge | RPA + an AI step | Classic RPA |
|---|---|---|---|
| Finding fields on screen | Reads the screen by meaning: labels, messages, field roles | Recorded steps, selectors or coordinates | Recorded steps, selectors or coordinates |
| Reading documents | AI vision built in; no template per layout; multi-record files split | AI extraction component returns data | Separate OCR tool and templates |
| When a screen changes | Finds the field by what it means, not where it was | Script breaks; a developer updates it | Script breaks; a developer updates it |
| Checking the data | Against the document, your plain-English rules and what the host holds | Validation you design and build | Validation you design and build |
| Unclear values and host errors | Flagged, never guessed; the person sees the field, the image and the host's message | Exception paths you script in advance | Exception paths you script in advance |
| Proof the record landed | Reads the record back from the host and logs the result | Whatever check the script includes | Whatever check the script includes |
| Audit trail | Every keystroke and screen, SHA-256 hash chain, session replay | Run logs, plus what you add | Run logs, plus what you add |
| What you get | A finished product for document entry, run as a cloud service | A toolkit with AI parts; you build the workflow | A toolkit; you build the workflow |
RPA vs AI: why recorded scripts struggle on legacy screens
Recorded scripts struggle because a green screen gives a bot little to hold on to: no element IDs, just rows of text where a field is a position. More in RPA vs AI for IBM i.
Positions move
A new field or a longer message line shifts what sits where. A script typing at row 12, column 30 now hits the wrong field.
The host talks back
Legacy applications answer with messages, warnings and lookups. A script handles the ones someone predicted; an agent reads what the host said.
Documents vary
Every supplier and customer sends a different layout. Template OCR needs one per layout; AI extraction still leaves the checking to you.
How the AI agent automation works, step by step
legacy bridge runs the document-to-system job in six steps; a person stays in control of submission.
Documents arrive
By email, an SFTP/FTP folder or an API upload: PDF, scans, photos, faxes, Word and Excel.
AI vision reads them
Fields are extracted without OCR templates. A file holding several records becomes one record each, and a document your team already keyed is caught as a duplicate. More on this in document AI.
Values are checked
Against the document itself, your process rules written in plain English, and what the host already holds.
The agent works the screens
It signs on, navigates the menus, fills the right fields, handles lookups and function keys, and reads the host's own messages — see TN5250 automation.
A person approves
By default every transaction waits for an employee's approval in the web console or from an email link. Once trusted, a workflow can run on its own.
The record is confirmed
The agent reads the record back from the host and logs the result. Billing counts only records filed and confirmed.
Agentic automation needs guardrails
An agent that types into your system of record must be easy to stop, easy to check and limited to what you agreed. These controls are built in.
An approval gate
Every transaction can wait for a person's approval. You decide when a workflow may run without it.
Flags, not guesses
Unclear, missing or hand-corrected values hold the record with the exact field and the document image. If the host returns an error, the person sees the error and the screen; nothing is submitted.
A full audit trail
Every keystroke and screen is recorded in a tamper-evident SHA-256 hash chain; sessions can be replayed.
Scoped permissions
Environment, access, data handling and approval are agreed up front; the agent performs only those actions.
AI agents for legacy systems: what it works with
legacy bridge connects the way a terminal emulator does, so nothing is installed on the host and no program changes. For the wider picture, see legacy system automation.
In production
IBM i (AS/400, iSeries, System i), including custom RPG and COBOL applications and packaged ERPs on IBM i
Also supported
TN3270 for IBM Z (z/OS, CICS, IMS, TSO screens) — see mainframe automation — VT for OpenVMS, Unix and Pick, and HP 6530 for HP NonStop; scoped per workflow with the same controlled proof
Beyond terminals
Legacy web portals in a browser, and desktop applications reached over Citrix or RDP, driven by vision
Connection
TN5250 from the cloud; hosts not exposed to the internet through a secure SSH tunnel. No new APIs, no database access
Documents
Invoices, purchase and sales orders, claims, work orders, bills of lading, remittances
An RPA alternative that can also sit beside your RPA
For document entry into legacy screens, legacy bridge is an RPA alternative: the workflow is built and maintained as a product, not as scripts in your team's backlog. Our IBM i RPA alternative page compares the two in detail.
It does not ask you to retire an existing RPA estate. Bots that work keep running; legacy bridge takes the document-heavy workflows where scripts break most. Start with one document type and move work across only where the evidence supports it.
UiPath, Blue Prism or Power Automate for green screens
They can drive terminal sessions; you build and maintain the screen logic and exceptions. That suits teams with RPA developers and many varied processes.
Operations and finance teams
Teams keying invoices, orders or claims by hand who want the work done, not another platform to staff. See data entry automation.
IT and RPA owners
Teams tired of fixing bots after screen changes, who need an audit trail and approval step to show compliance.
