Muda in AI: 7 Types of Waste When Working with Artificial Intelligence

Some time ago, I wrote about the 7 types of waste in software development. Lean and its concept of Muda originated at Toyota for industrial manufacturing, yet mapping these categories onto software development still works surprisingly well. And now, as companies are learning to work with AI, from ChatGPT in everyone’s pocket to enterprise implementations with custom models — it’s becoming clear that the same seven categories of Muda apply here too. They’re just hiding in new places.

If your company has been asking over the past year, “why isn’t AI delivering the value we expected?”, the answer may well be that you’re producing AI waste faster than you’re producing benefits. Let’s take a look through the lens of Lean.

Waiting

In software, we wait for test environments, releases, and approvals. In the AI era, we mostly wait for someone to tell us we’re allowed to use AI in the first place. Employees wait for official AI guidelines. Teams wait for an approved list of tools. Managers wait for someone else’s pilot project to see whether it works. And while everyone is waiting, the competition has been working with AI for three months.

The paradox is that corporate AI guidelines are supposed to enable AI use and instead, waiting for them becomes the main bottleneck. Meanwhile, employees are using AI anyway, just outside the corporate framework, without audit trails and without sharing best practices.

Examples:

• Waiting for corporate AI guidelines before people are “allowed” to try AI tools
• Waiting for IT to approve a list of tools
• Waiting for usage limits to reset on cheaper licences or free tiers

Inventory

Inventory in software might be analysis sitting in a drawer, or finished but undeployed code. In AI, inventory takes on a new dimension — and it’s expensive in the most literal sense.

Purchased AI tools that nobody uses are today’s quiet epidemic in large companies. A company signs an enterprise contract for an AI tool covering 500 people, and 50 actively use it. Prepaid API access that doesn’t even reach half its capacity. The investment is visible on the balance sheet; the value is nowhere to be found.

A second type of AI inventory is the prompt library without a curator. Someone enthusiastically writes 200 prompts for various use cases at the start. Six months later, half no longer work (the model has changed in the meantime), a third are duplicates, and nobody knows which one is the right one.

Examples:

• Purchased AI tools and licences that nobody uses
• Fine-tuned models that nobody deployed to production
• Prompt libraries that nobody maintains

Defects

In software, defects are production bugs and technical debt. In AI, we get a new type of defect, the hallucination that looks credible.

A classic bug in development usually shows itself: the application crashes, a test fails. An AI defect is insidious because the output looks perfect. Fabricated citations in a legal brief. An incorrect total in a financial report. A made-up link in a marketing email. Nobody checks anything, because “the AI wrote it and it looks right.”

And the deeper AI penetrates into decision-making processes, the more expensive the consequences become. A hallucination in a draft email is an inconvenience. A hallucination in the materials for an investment decision is a problem.

Examples:

• Hallucinations accepted as facts
• AI-generated code or content without review
• Prompt debt, unversioned, untested prompts copied between teams

Overproduction

In Lean philosophy, overproduction is the worst type of waste because it generates all the others. In software, it means building features nobody uses. In AI, this is doubly true and it includes AI features built for their own sake, with no regard for the customer.

“We need to put AI in there” is today’s most expensive sentence in the corporate world. A chatbot on the website nobody uses. An AI summary in an app users immediately skip past. AI in an internal tool that returns worse results than search. An AI-generated personalised onboarding nobody cares about.

The cause is the same as in software, missing customer feedback and the conviction that “we know what the user wants.” Only now it’s amplified by pressure from CEOs and investors who want to see an “AI strategy” in every quarterly report. Features get built because AI is available, not because they solve a real need.

Examples:

• AI features built without regard for the customer — nobody wants them
• Chatbots instead of better design
• Automating processes that should have been eliminated, not automated

Transportation

In software, transportation is about handing information off between people and systems. Every handoff = risk of losing context. In AI, handoffs between models are now added to handoffs between people.

Chaining several models where one would do has been fashionable over the past year. One model generates the text, another translates it, a third fact-checks it, a fourth formats it. Every transition = potential context loss, longer lead time, and higher token cost. Architectures of the “agent calls agent calls agent” variety look impressive on slides, but in practice often introduce errors where there were none before.

The truth is that most of the use cases on which companies deploy multi-agent solutions, some of today’s models can handle in a single pass. All it takes is a well-written prompt and a quality input.

Examples:

• Chaining several models where one would do
• Copy-paste workflows between a tool and an AI chat
• Sensitive data leaking out through the prompt interface

Overprocessing

In software, overprocessing shows up as overly tight processes and workflows that people end up bypassing. In AI, we get new forms of it. A typical example is excessive prompt engineering where simplicity would do.

People spend hours tweaking prompts, adding chain-of-thought, few-shot examples, system messages, role definitions, XML tags, and meta-instructions — for a task that would have needed one clear sentence and a quick check of the output. A cult of complexity emerges: the longer the prompt, the more professional. Yet modern models often respond better to short, precise instructions than to a three-page brief in which even the author gets lost.

Examples:

• Excessive prompt engineering where simplicity would do
• Enterprise AI projects for problems that off-the-shelf tools already solve

Motion

Motion in software = unnecessary steps, approval loops, meetings without an agenda. In AI, switching between tools and repeated re-prompting are added to the mix.

A user writes a prompt. The result still isn’t quite right. They rewrite the prompt. Still no. They switch to a different model. Try again. Go back to the first one. Copy the output into a third tool to clean it up. Instead of focused work, it turns into AI fidgeting, constant switching that looks productive, but where an experienced person would have produced the final output by hand more quickly.

The second type of motion is AI approval loops. A generated draft, then human review, followed by an AI review of the review, then a second human review, then a final AI polish, and of course an approval to top it off. The finished item bounces back and forth several times — only now with even more actors involved.

Examples:

• Switching between multiple AI tools for a single task
• Repeated re-prompting instead of thinking about the brief
• AI reviewing AI output as ritual, not as quality

What to Do About It

As with software — first you have to notice it, then measure it, then remove it. Three practical questions that help in an AI audit:

1. Where are people waiting? (For a tool, for permission, for a model, for a response.)
2. What have we bought or built and aren’t using? (Licences, models, prompts, AI features.)
3. Where is AI adding steps instead of removing them?

AI on its own creates no value. Value comes from removing waste in the process AI serves. If you add AI on top of existing chaos, all you get is faster chaos. And that, to paraphrase a classic, is a more expensive form of waste than anything Toyota could ever have come up with.