I’ve spent a good chunk of time in the pure RPA world before joining Nintex and I’ve had to build a ton of these solutions. Bank Reconcilliations are one of the go-to use cases in the RPA world to show off value. Everyone’s done it. Everyone’s overcomplicated it.
With the new AI actions launched earlier this year in the Nintex CE Platform, comparing toolsets and Enterprise ProductsI’ve leveraged in the past this is the easiest, fastest and most importantly most accurate approach to Intelligent Document Automation and does not require any developer to get involved in the process.
The build Process
- Create a new workflow
- Add the Action
- Curate and Provide your AI prompt based on your requirement
- Add the additional logic in order to complete the workflow (convert the file from JSON to a document / save the file to SharePoint / manipulate the file format / make calculations / etc.)
Bank Reconcilliation
My task was to produce a custom POC build showcasing how Nintex can automate the Bank Recon process, and the basic ask was this
- Use bank statements as an input
- Extract the transactions
- Convert the extracted content to Excel
- Do some excel formatting and calculations
- Save the output to a document library in SharePoint and email a link to the finance team
I did have to leverage a basic API to convert the JSON-formatted extractions into a workable Excel file and create a custom Nintex Xtension as this action isn’t available out of the box, but this was pain free (and again I leverages the Nintex AI Xtension builder to help with that and was a matter of minutes – search for a public API, generate the Xtension and it was a 10 minute task).
The most complex part of the workflow build was the prompt engineering the Nintex AI Action uses to extract the content to ensure consistency in the formatting of the JSON response. What worked best for me was this:
Extract ONLY the FIRST 25 transaction rows from the bank statement and output them in the following exact JSON structure.
Do not include any additional wrapper objects, labels, summaries, or metadata.
{
"transactions": [
{
"description": "",
"currency": "",
"transaction_date": "",
"payment_date": "",
"debits": 0,
"credits": 0
}
]
}
Do not nest the transactions under ‘value’, ‘bank_statement’, ‘account’, ‘data’, or any other wrapper.
This being a process that used to be a nightmare of layout handling and days of OCR and classic OCR Document Layout training, testing and tweaking, and now it’s basically done in less than a day in Nintex CE.
The Gotchas
- Invisible Characters: Watch out for Unicode gremlins in your prompt. Strip your prompt down to plain text first.
- Repetitive Prompts: If you hammer the same prompt over and over in testing, you might get a “too repetitive” warning. Just vary it a bit or slow down your tests.
1. AI extraction doesn’t care about document layouts and one action can cater for many
This kills 90% of traditional recon complexity.
2. Your AI prompt is crucial
NAC AI follows your schema exactly, but this means prompts must be:
- Strict
- Descriptive
- Clean
- Free of invisible characters
3. Repetition throttling
This is not a massive issue, but worth planning for in testing.
4. Post-processing is still required
The AI document piece gets you 80–90% there.
You still need:
- Date normalization
- Personal Preference? YYYY/MM/DD (Besides for the fact that it’s my local Standard regional standard, it’s best for sort as well.
- Amount parsing
- Edge-case handling
- Duplicate suppression
5. Error-handling is crucial
In a production environment your workflow should include;
- Soft-fail paths
- Partial extractions
- Retry logic
- AI confidence validation
Input/Output Check Flags to consider
Input check flags:
- Invisible/control Unicode characters.
- Override/ignore system-instruction phrases.
- Forced-output commands.
- “System breached/override/developer mode” wording (multilingual).
- Template/markup/code injection markers.
- Encoding/exfiltration hints.
- Social-engineering bypass wording.
- Context/role confusion cues.
- Overly long inputs (>100,000 chars).
- Repeated chunks (≤1,000 chars repeated >1,000 times).
- Base64-like payloads without spaces.
- Response flags: breach wording, unexpected code/HTML, suspicious single-token outputs.
Output checks flag:
- Breach or override wording (e.g., “system breached”, “admin access granted”).
- Multilingual breach phrases (French/Spanish variants).
- Unexpected markup or code blocks (<script>, HTML comments, fenced code).
- Suspicious single-token strings (numeric-only responses).
In closing
In the automation world (and even more specifically RPA), bank recon is one of the go-to use cases because it’s a pain point every company feels and can relate to. I’ve done dozens of these POCs in the past, and this is by far the easiest extraction I’ve ever run. It’s not just a little better — it changes the whole game. The AI does the heavy lifting so you can focus on the business logic instead of wrestling with document layouts. If you’ve been through the old-school nightmare of building recon workflows, you’ll appreciate just how much smoother it is now.
