Process Automation: AI - to use, or not to use

If you're an operations leader thinking about how — or indeed whether — to use AI to help optimize your team’s workflows, hopefully this article will be helpful. We'll walk through practical considerations for where AI is a good fit, where it might not be, and how to think about different ‘automation’ options.

It starts with a way of thinking about processes that I picked up twenty-five years ago as a graduate working in IT for a global car manufacturer. We were obsessed with one thing: understanding production line processes and using IT to remove friction. We implemented lean manufacturing principles using systems to automate Just-In-Time parts supply, and constantly asked: where's the waste in the process, and how do we eliminate it?

Later, as I was leading a transformational change for a large bank it was clear that whether you're optimizing a car production line or automating lending processes, the thinking is the same.

To automate lending, we used factory floor thinking applied to knowledge work. We optimised by replacing the wasteful routine, repeatable parts of the lending workflows with automation and preserved human judgment for the most nuanced decisions. We leveraged IT systems to collect data consistently, triage applications reliably, and let the team focus on the edge cases.

Three Ways to Automate

Today there’s essentially three fundamentally different approaches to using technology to automate workflows. Each solves a different problem, at a different cost.

Code-Based Automation: The Precision Manufacturing System Hard coded systems are ‘deterministic’. They make sense under specific conditions: high-volume, fully repeatable tasks with minimal variability. Once you've built it, it runs the same way reliably. But the upfront cost is high - you need developers, infrastructure, and time. If business logic changes, you need to rework it. ROI only works if the volume justifies the investment and the process stays stable.

RPA: The Mechanical Robot RPA - Robotic Process Automation - in simple terms you train a computer to reproduce the steps a human would. It's still deterministic so relatively expensive to implement and maintain. It made sense as a last resort for organizations stuck with legacy systems that couldn't support API connectivity. As AI agents evolve, they'll do what RPA does - but better, faster, and with far less upfront configuration. I'd be cautious about long-term RPA investments.

AI-Powered Automation: The Intelligent Worker AI is fundamentally different. It’s capable of handling ambiguity, nuances, and variability at scale. It can extract unstructured data from documents, handle nuanced edge cases, and it can adapt.

It's not without downsides though. There's some latency, token costs might add up and there's the hallucination problem - AI can produce confident-sounding mistakes. So tool choice matters (not every AI tool is built equally - some tools will give more consistent results thanks to better memory management capabilities or support for multi-agent workflows etc).

The real strength for AI solutions is speed to implementation and ability to deal with ambiguity. Using the right tool, you won’t need a developer and you can get a solution live in hours or days, then test it, learn from it, and then decide what comes next.

Variability Is the Killer for Determinism Few real workflows are simple linear processes like a car production line. Data comes in different shapes. Customers have edge cases. System changes upstream break assumptions downstream. You might be able to handle 80 percent of a workflow easily, but that remaining 20 percent, where the exceptions live, that’s where complexity explodes.

Deterministic systems might be reliable but are expensive to build, difficult to extend and optimized for certainty. Most of your workflows aren't certain.

Why AI Handles Uncertainty Better

An AI native system doesn’t box you in to rigid ‘if this, then that’ sort of logic. It’s probabilistic - so when you've got edge cases, different rules and different handling required an AI system is better equipped for dealing with the messiness without requiring you to have coded for every scenario.

It feels like magic, but it's not. An LLM has been trained on enormous amounts of data. When you give it something to do, it infers the most probable outcome. Immense computational power is why results are often really good. But it's also why AI is relatively expensive and slower compared to simple if/then/else logic.

And so it's not perfect. AI makes mistakes. Sometimes the variance in outputs is fine. Sometimes it's not. But here's the thing - humans aren't perfect either. Give the same loan application to ten credit officers and you'll get variances in their scores. So the question isn't whether AI is flawless. It's whether the results you achieve are acceptable for what you're doing, and whether it's cheaper and faster than building a rigid deterministic system that tries to handle every variation.

The Smart Approach

Here's the smart strategy: trial with AI in minutes, have visibility into costs and accuracy, learn fast. Maybe you never need to optimize. Or maybe you find certain tasks that are repetitive and low-variance - perfect for deterministic code - and you migrate those to a hybrid approach. But you'll only know if you've actually run it.

The simple decision: if you wanted to build this with code, could you do it in under an hour and be confident it'll handle all scenarios? If the answer is no, start with AI.

Case Study: How It Works

An operations manager in financial services had a workflow eating time. Emails came in with attachments containing customer data. The format varied wildly with spreadsheets, PDFs and text. The downstream CRM system had a rigid schema. Data had to be validated, standardized. First, check: does the customer exist? Update or create new? Then send a Slack notification confirming what was processed and flagging exceptions.

A human was spending 30-45 minutes per batch. The variability in input data meant hard coding rules wouldn't work. Building a deterministic solution would have taken weeks - integration work, data mapping, business logic configuration.

Instead, they signed up for a QuivaWorks account and built an AI Assistant in under 20 minutes.

Here's the key: the user didn't code business logic or map data schemas. They described the business process in natural language. The QuivaWorks assistant already knew what the CRM required - it had that knowledge built in through its smart integrations tooling. So when the user said "check if customer exists, update if they do, create if they don't," the assistant automatically figured out the right API calls without the user needing to do any configuration or data mapping.

The assistant handled input ambiguity, parsed accurately, made decisions, pushed data to the CRM and sent Slack notifications. All in seconds. Yes, token costs were higher than a deterministic solution would be to run per batch. But, there was zero capital cost to build and zero maintenance overhead.

This is why we built QuivaWorks - to remove friction between a business problem and a solution.

Closing

Automation isn't one-size-fits-all. The smart operations leader doesn't start with the tool. They start with the problem. How much variability does it have? What improvement are you looking for? When do you need it? Then they fit the solution to the problem.

Next time you're automating an operational workflow, invest an hour trying out an AI Assistant as your first step. You might solve it faster and cheaper than expected. Or you may quickly learn enough to make a smarter decision about what comes next.

Either way, you'll be thinking about it the right way.