Where RPA Breaks and AI Takes Over in Operations

An operator logs in on a Monday morning to find 400 exceptions in the processing queue. The culprit? A major freight forwarder changed their commercial invoice template by adding a new column and shifting the totals box down by two centimetres.

If you run Robotic Process Automation (RPA) in your operations, you know this pain. RPA mimics human clicks and keystrokes. It is brilliant for moving data between a spreadsheet and a legacy system that lacks an API, but it demands absolute, pixel-perfect predictability. The moment reality deviates from the rule, the bot throws an error or, worse, keys the wrong data into your ERP.

Operations is rarely predictable. Exhibitors reply to onboarding emails with badly scanned PDFs. Suppliers send packing lists in the body of an email instead of an attachment. Customs brokers use different naming conventions for the same commodities.

Trying to force RPA to handle this variance requires writing endless, brittle rules. Trying to let AI execute data entry directly into your database risks hallucinations and corrupted records. The practical answer for UK and Australian operations teams is a strict division of labour: AI handles the reading, a human handles the decision, and RPA handles the heavy lifting.

The Trap of Rule-Based Automation

Many businesses buy an RPA tool to fix a data entry bottleneck. You map out a process: open the email, download the attachment, read cell B4, copy to field 1, read cell C7, copy to field 2, click submit.

It works perfectly in the pilot. Then it meets the real world.

A trade show organiser moving exhibitor health and safety documents from an inbox into a portal quickly finds that no two contractors submit the same layout. An importer trying to automate customs declarations finds that their Chinese suppliers and Australian freight forwarders use entirely different formats for commercial invoices.

To fix this, IT teams try to write more rules. They build a template for Supplier A, a template for Supplier B, and a template for Supplier C. When Supplier A updates their software, the template breaks. The ops team ends up spending more time managing the bot exceptions than they did doing the manual data entry. This is when Excel automation stops scaling for ops teams.

RPA has no ability to infer meaning. It does not know that "Total Wgt" and "Gross Weight" mean the same thing. It only knows what you explicitly told it to look for.

AI as the Translator

Instead of treating variance as an error, generative AI expects it. Large Language Models (LLMs) are exceptionally good at taking unstructured, messy inputs and structuring them into a predictable format.

In a modern operations workflow, AI sits at the very front of the process. Its only job is to act as a translator between the chaotic outside world and your internal systems.

Take an incoming email from an exhibitor requesting a change to their shell scheme booking. The email thread is seven replies deep, includes two different company names, and has a hand-annotated PDF attached.

An RPA bot cannot read this. An AI extraction workflow can. It reads the thread, ignores the signature blocks, identifies the core request (add two spotlights, change the fascia name to 'Tech Solutions Ltd'), and structures that request into a standard JSON file.

It does not act on the request. It just organises it.

The Handoff: Where the Human Steps In

This is the most critical part of the workflow. AI must never have direct write access to your core operational systems. If an AI misreads a currency symbol and logs a £5,000 order as a £50,000 order, and a bot immediately executes that in your ERP, you have a massive commercial problem.

The system must force a human review. The handoff problem is where most theoretical AI projects fail, because they assume the AI is replacing the operator entirely. It is not. It is replacing the operator's typing time, not their judgement.

The practical setup looks like this:

1. The Input: The messy email and PDF arrive.
2. The AI Draft: The AI reads the documents and extracts the relevant fields (supplier name, invoice number, line items, totals).
3. The Interface: An operator opens a dashboard. On the left side of the screen is the original PDF. On the right side are the data fields the AI extracted. If the AI is unsure about a specific field (low confidence score), it highlights it in red.
4. The Decision: The operator quickly scans the document, corrects the red field, and clicks 'Approve'.

The operator's job shifts from data entry to data verification. They can clear 100 invoices in the time it used to take to manually type 15.

RPA as the Courier

Once the human clicks 'Approve', the data is now perfectly structured, verified, and guaranteed to match your system's required format.

Now, the RPA bot wakes up. Because the data is uniform, the bot will not break. It takes the approved JSON file, opens your legacy warehousing software, clicks the exact buttons, and types the exact fields at superhuman speed.

You get the best of both tools. AI handles the variance of the real world. RPA handles the rigid demands of your legacy systems.

Costing the Dual Approach

Building this architecture sounds expensive, but it is entirely achievable for mid-sized operations. You do not need to buy a monolithic enterprise platform.

If you already pay for an RPA licence (like UiPath or Power Automate) to move data, you keep it. You are simply adding a lightweight AI triage layer in front of it. The costs involve paying for the API calls to an LLM (fractions of a penny per document) and the initial build of the human-in-the-loop review interface.

Operators often ask if they should just hire another administrator instead. You can run the numbers through an Automation vs Hire Comparator to check your specific volume, but generally, if a team of three is spending more than 40% of their week rekeying data from emails, the AI triage layer pays for itself within the first year.

For a detailed breakdown of what that build looks like, look at the real cost of an AI project for B2B ops software.

How AI-Assisted Product Delivery Speeds Up the Build

Historically, standing up a new interface to sit between your email server and your RPA bots took months of scoping, wireframing, and custom coding. Today, AI-assisted product delivery dramatically shortens that cycle.

Development teams use AI coding tools to rapidly generate the boilerplate code for API integrations, the data extraction logic, and the React frontend for the human review dashboard. Instead of spending six weeks writing the routing logic to handle different types of attachments, a development partner can stand up a working prototype in days.

This means you get a functional pilot in the hands of your operations team faster. You can run 50 real commercial invoices through the extraction layer, see exactly where the AI struggles, adjust the prompts, and measure the actual time saved by the human reviewer before committing to the final build.

It shifts the project from a massive, risky IT overhaul into a tight, measurable workflow improvement. You define the exact documents you want to process, you build the extraction layer, you test the human handoff, and you pipe the approved data into your existing RPA setup.

Start small. Pick one supplier template that breaks your bots every month, route it through an AI extraction layer, and let your operators review the output.

Useful tool

Try Samvara's Document Readiness Checklist — Export/import docs by mode.