Agentic Workflows vs RPA: Why We Stopped Recommending UiPath

We stopped recommending UiPath to clients in late 2025. Not because it's a bad product — it's excellent at what it was designed for. We stopped because the problems our clients brought us had changed, and RPA was no longer the right tool for most of them.

What Changed Between 2020 and 2026

In 2020, the typical automation request was: 'We have a person who copies data from System A to System B eight hours a day. Automate that.' RPA was the perfect answer. Record the clicks, replay them, scale to 50 bots, save $300K/year in labour costs. UiPath, Automation Anywhere, Blue Prism — all built for this exact problem.

In 2026, the typical automation request is: 'We have a process where someone receives an email, reads the attachment, decides whether it matches one of 15 categories, checks three different systems for context, makes a judgment call about urgency, and then routes it to one of four teams with a summary of what needs to happen.' Try recording that as a click sequence.

The difference isn't complexity for its own sake. It's that the easy automations have already been automated. What's left are the processes that require interpretation, context, and decision-making — exactly the things RPA was never designed to handle.

How RPA Actually Works (And Where It Breaks)

RPA operates on a simple principle: record a sequence of UI interactions and replay them. Click this button. Type in this field. Copy this value. Paste it here. Move to the next record. Repeat.

This works perfectly when:

The UI doesn't change. If the target application moves a button, changes a field name, or updates its layout, the bot breaks. Every RPA team maintains a queue of 'bot fixes' triggered by application updates.

The process is entirely rule-based. If the decision at every step can be expressed as an if/then statement, RPA handles it. 'If the invoice amount is over $10,000, route to senior approval.' Clear rule, clean automation.

The data is structured. RPA reads fields, cells, and form inputs. It doesn't read unstructured text, interpret images, or understand context. A bot can extract the dollar amount from an invoice field. It cannot read a contract paragraph and determine whether the terms are favourable.

The exceptions are rare. RPA handles the happy path. When exceptions occur — and in real business processes, they occur 15–30% of the time — the bot either crashes, produces incorrect output, or routes to a human exception queue that becomes the new bottleneck.

What Agentic Workflows Do Differently

An agentic workflow doesn't replay click sequences. It uses AI models to understand context, make decisions, and take actions across multiple systems — without needing a predefined path for every scenario.

They handle unstructured input. An agent can read an email, understand what's being asked, extract relevant entities (company name, product mentioned, urgency signals), and classify the request — even if the email format varies every time. RPA can't process an email unless it follows a strict template.

They make judgment calls. When a process has 15 categories and the correct one depends on context from three systems, an agent can synthesise that information and choose. Not perfectly — accuracy rates of 85–95% are typical for classification tasks — but fast enough and accurate enough that humans only need to review the uncertain 5–15%.

They recover from exceptions. When something unexpected happens, an agent can assess the situation, try alternative approaches, and escalate with context. An RPA bot that encounters an unexpected dialog box stops. An agent that encounters an unexpected response evaluates it, decides if it can proceed differently, and either handles it or escalates with an explanation of what went wrong.

They improve over time. Every correction a human makes becomes training data. The agent gets more accurate as it processes more cases. RPA bots don't learn — they execute the same recorded sequence forever, or until someone records a new one.

The Exception Problem: Why We Switched

The turning point for us was a client project in late 2025. A manufacturing company needed to automate their purchase order processing. The happy path was straightforward: receive PO, validate against contract terms, check inventory, confirm delivery date, send acknowledgment.

We scoped it as an RPA project. The first version automated the happy path in three weeks. Then reality hit.

12% of POs had line items that didn't match the product catalogue exactly — different naming conventions, abbreviations, or bundled items that needed to be decomposed. The RPA bot flagged these as exceptions.

8% had special pricing that required checking a separate contract database and applying tiered discounts based on volume commitments. The bot couldn't handle conditional pricing logic.

5% came in as PDF attachments with slightly different formats from each customer. The bot could only read the structured EDI format.

Combined exception rate: 25%. One quarter of all POs required human intervention despite the 'automation.' The exception queue became the new bottleneck — worse than before, because now the team had to context-switch between the automated and manual tracks.

We rebuilt it as an agentic workflow. The AI agent read POs regardless of format (PDF, email body, EDI). It fuzzy-matched product names against the catalogue. It checked contract terms and applied conditional pricing. When it genuinely couldn't determine the right action — about 4% of cases after training — it escalated with full context so the human reviewer could resolve it in 30 seconds instead of 5 minutes.

Exception rate went from 25% to 4%. Processing time per PO dropped from 12 minutes (average, including exceptions) to under 2 minutes.

When RPA Is Still the Right Choice

RPA is not dead. It's narrower than vendors sold it as, but the use cases where it fits are real and well-defined.

High-volume, zero-exception structured data transfers. Moving records between two systems with stable APIs or UIs where the data mapping is 1:1 and never changes. Payroll data transfers. Monthly report generation from fixed templates. Inventory count uploads.

Regulated processes where AI judgment is not acceptable. Some compliance workflows require deterministic processing — the same input must always produce the same output, with an auditable trail showing exactly which rules were applied. AI models are probabilistic. When a regulator needs to verify that every step followed a specific rule, RPA's determinism is an advantage.

Legacy system integration where APIs don't exist. Some 20-year-old ERP systems have no API and no database access. The only way to interact is through the UI. RPA was literally built for this — automating UI interactions with systems that offer no other interface. Agentic workflows typically need an API or database connection to function.

If your process fits these criteria, UiPath and its competitors are still good tools. But if your process involves interpretation, judgment, variable formats, or high exception rates — that's agentic territory.

The Cost Reality: RPA vs Agentic

RPA vendors price on a per-bot or per-automation basis. UiPath's enterprise licensing starts at roughly $8,000/year per attended robot and $15,000/year per unattended robot, plus platform fees of $20,000–$40,000/year. A ten-bot deployment costs $170,000–$200,000/year in licensing alone, before development costs.

Add development: each RPA automation typically costs $15,000–$40,000 to build and $5,000–$15,000/year to maintain (handling UI changes, exception tuning, bot fixes). Ten automations: $150,000–$400,000 to build, $50,000–$150,000/year to maintain.

Total first-year cost for a typical enterprise RPA programme: $370,000–$600,000.

An agentic workflow built custom costs more upfront — typically $40,000–$100,000 per workflow depending on complexity — but doesn't carry per-bot licensing. Infrastructure costs run $500–$3,000/month on cloud compute for AI model hosting. Maintenance is lower because the agent handles exceptions that would break an RPA bot.

Total first-year cost for an equivalent agentic deployment: $200,000–$400,000, with lower Year 2+ costs because there's no per-robot licensing and fewer break-fix cycles.

The economics flip in Year 2. RPA licensing is perpetual. Agentic infrastructure costs stay flat or decline as models get cheaper. By Year 3, the agentic approach is typically 40–50% less expensive than the equivalent RPA deployment.

Hybrid Architectures: Using Both

The practical reality in most enterprises is that some processes suit RPA and others suit agentic workflows. A hybrid architecture uses each where it's strongest.

RPA layer: Handles the deterministic, structured, high-volume data movements. No AI needed. No judgment required. Cheap per transaction.

Agentic layer: Handles the processes that involve interpretation, classification, context-dependent decisions, and variable inputs. Uses AI models. Costs more per transaction but handles cases RPA can't.

Orchestration layer: Routes work to the appropriate handler based on characteristics. Simple, structured input → RPA. Complex, variable input → agent. Uncertain → human review queue.

This hybrid model is what we recommend for enterprises that already have RPA investments. You don't throw away working bots. You stop extending RPA into territory where it fails, and use agents for the new automation that RPA can't handle.

How to Decide: A Practical Framework

For any process you're considering automating, answer four questions:

1. What's the exception rate? If less than 5% of cases require human judgment, RPA works. If 15%+ require judgment, you need agentic.

2. Is the input structured or variable? Structured (forms, templates, EDI) → RPA. Variable (emails, PDFs, free text) → agentic.

3. Do decisions depend on context from multiple systems? Single-system rule application → RPA. Cross-system synthesis → agentic.

4. Does the target UI change frequently? Stable UI → RPA works. Frequent updates → agentic (API-based) avoids constant bot maintenance.

If a process scores 'RPA' on all four, use RPA. If it scores 'agentic' on any one, seriously evaluate an agentic approach. If it scores 'agentic' on two or more, RPA will create more problems than it solves.

What We're Building Now

At Madgeek, the shift from RPA to agentic workflows has defined our engineering practice in 2025–2026. The AI call quality monitoring system we built for an operations client — the one that helped scale from 50 to 80+ agents in three months — is fundamentally an agentic system. It listens to calls, understands context, scores quality based on nuanced criteria, and routes coaching recommendations to supervisors. An RPA bot couldn't even begin to process an audio conversation.

The manufacturing cost estimator we built uses ML models to predict material costs based on specifications, historical data, and market conditions. The procurement platform for Tejas Networks digitised an approval workflow with dozens of exception types that would have required a different RPA bot for each variation.

We still use deterministic automation for the parts that don't need AI. Data synchronisation between systems. Scheduled report generation. Structured file processing. But the interesting problems — the ones where automation actually changes how a business operates — are all agentic now.

If you're running an RPA programme that's hitting diminishing returns, or if you're scoping new automation and the exception rate keeps climbing during analysis, that's the signal. The process needs judgment, not playback.