Accelerated Software Development

min read

RPA in Logistics: Where Bots Hit Their Limit and AI Agents Take Over

Written by

SHIVA SANKAR

Published on

July 4, 2025

Your RPA programme looked like a win for the first eighteen months. Bots handled invoice matching, shipment tracking, and order entry without complaint. Then a carrier portal changed its layout, three bots broke overnight, and your team spent a week on fixes instead of the roadmap.

That's not a one-off. Maintenance now consumes 70% to 75% of total RPA automation budgets industry-wide. The bots that were supposed to free up your ops team are, at scale, quietly employing a growing share of it just to keep the estate running.

This isn't an argument that RPA was a mistake. It solved a real problem, structured, repetitive, rule-based tasks, and did it well. The problem is what happens next: the exceptions, the unstructured supplier emails, the PDFs that don't match last year's template. That's where bots hit their limit, and where AI agents pick up.

RPA Solved One Problem, and Only One Problem

RPA bots mimic human clicks on a screen, following a fixed script: read this field, paste it there, check this condition. That works reliably when the input never changes shape. It breaks the moment it does.

A bot has no way to reason about a "warehouse temporarily closed" message it wasn't scripted for. It can't interpret a supplier email that arrives in a different format than the one it was trained on. It doesn't understand what it's doing; it executes a sequence, and if the sequence doesn't match reality, it fails silently or loudly, depending on how well you've instrumented it.

That's the structural ceiling. It's not a bug to patch. It's what rule-based automation was always going to run into once the easy, stable processes were automated and what remained was the messier 20%.

Where RPA Breaks in a Real Supply Chain

Supply chain data is rarely as clean as the process map suggests. Supplier communications arrive as scanned PDFs, images, emails in different languages, and instant messages with no standard field names. None of it matches the tabular structure RPA and your ERP were built to expect.

Every user interface change is a maintenance event. A carrier portal redesign, an ERP version upgrade, a vendor changing their invoice template — any of these can break a bot's script, and someone has to notice, diagnose, and rebuild it. At scale, across dozens of bots and multiple vendor systems, that becomes a standing maintenance workload, not an occasional fire drill.

Exception handling is the sharpest limit. A bot processing shipment data has no path forward when it hits a scenario nobody scripted for. It stops, or it does the wrong thing confidently. Either way, a person has to step in, which defeats the point of automating the process.

A Bot Estate Where Maintenance Caught Up With the Build Cost

One logistics enterprise we've seen this pattern with built out 40 RPA bots over three years, covering order entry, invoice matching, shipment tracking, and carrier communications across their network. The programme delivered real early savings; the first dozen bots automated genuinely repetitive, stable processes well.

By year three, the maintenance bill told a different story. Applying the industry-typical 70-75% maintenance share of total automation spend to a bot estate that size, the ongoing cost of keeping 40 bots running — fixing broken selectors after every vendor portal update, patching scripts after ERP releases, monitoring for silent failures — was approaching what the original three years of implementation had cost to build.

The decision wasn't to rip out the bot estate. It was to stop trying to make RPA handle what RPA was never built for. The stable, high-volume, rule-based bots — the ones processing consistent invoice formats and structured order data — stayed as RPA. Everything touching unstructured supplier communications, exception routing, and carrier disruption handling moved to an AI agent layer sitting alongside the existing bots, not replacing all of them.

That's the pattern behind AI agent engineering for enterprise supply chain automation: identifying which parts of a bot estate are genuinely stable enough to stay rule-based, and which parts have been quietly consuming your maintenance budget because they were never a good fit for rules in the first place.

Where RPA Still Wins, and Where Agents Take Over

This isn't a case for retiring every bot. It's a case for putting each task in the automation approach that actually fits it.

RPA still wins for high-volume, zero-variation tasks: structured data entry, standardised invoice formats, order records that follow the same shape every time. It's fast, cheap to run, and doesn't need a large model behind it.

AI agents take over where the input varies: unstructured supplier documents, exception handling that requires judgement, and multi-step decisions that span more than one system — rerouting a shipment around a closed warehouse, reconciling a mismatched invoice format, flagging a compliance risk buried in a supplier email.

The strongest deployments we've seen combine both: RPA executing the stable, structured work, with an AI agent layer handling triage, exceptions, and anything that doesn't fit last year's script. Neither technology does the other's job well.

How to Audit Your Own Bot Estate Before Deciding What's Next

1. Track maintenance hours per bot, not just uptime. A bot with 99% uptime that consumes ten engineering hours a month to maintain is a different proposition than one that runs untouched.

2. Map your actual exception rate. If a meaningful share of transactions routes to manual handling because a bot can't process them, that's the segment an AI agent layer is built for, not a scripting problem to solve with more RPA.

3. Separate stable processes from volatile ones. A process built on a vendor portal that changes twice a year needs a different approach than one built on your own stable internal ERP fields.

4. Cost the maintenance burden against the original build, not just against manual labour. The comparison that matters isn't "bot versus person" anymore. It's "cost of maintaining this bot versus the cost of an agent that doesn't break the same way."

Want to see where your bot estate's maintenance cost has quietly overtaken its value? Talk to us about AI agent engineering for enterprise supply chain automation.

FAQs

Does moving to AI agents mean retiring existing RPA bots?

Not necessarily. The processes RPA still handles well — stable, high-volume, structured data — usually stay as-is. AI agents typically get added for the unstructured and exception-heavy work bots were never suited to.

What's the actual cost problem with RPA at scale?

Maintenance, not licensing. Industry data shows maintenance consuming 70% to 75% of total RPA automation budgets, driven by UI changes, vendor format updates, and system upgrades that break bot scripts.

How do you know if a process needs an AI agent instead of another bot?

If the input format varies, if exceptions require judgement rather than a fixed rule, or if the task spans multiple systems with conditional logic, that's outside what a rule-based bot can reliably do.

Is this a full replacement of RPA, or an addition to it?

For most enterprises, an addition. The processes that were always a poor fit for rigid scripts — unstructured supplier communication, exception routing — move to agents. Genuinely stable, structured processes often stay exactly as they are.

How long does it take to add an AI agent layer to an existing RPA estate?

It depends on how cleanly your existing bots are documented and how fragmented your source data is, but it's typically a phased rollout starting with your highest-exception, highest-maintenance processes first.