Lean AI automation: fix the process, then add the agent.

Most writing on AI automation starts with tools: which model, which platform, which integration. I start with the process. Over twenty years in production, quality and operations taught me one thing: automate a bad process and you just produce waste faster. For an overview, see the AI automation guide.

What is lean AI automation?

It is an automation approach that combines lean thinking's hunt for waste with the speed of AI agents. The goal is not to "let AI do everything" but to remove repetitive work that adds no value and hand the remaining repetitive work to an agent.

The difference: the classic approach automates the existing workflow as it is. The lean approach first asks whether that workflow is needed at all.

What are the steps of the method?

I work in five steps:

• Go and see: see how the work is really done and who spends how much time on it, on site or over screen sharing.
• Map the flow: show the steps from request to delivery, waiting times and repetitive work on one page.
• Remove the waste: remove unnecessary approvals, copy-paste work, double entries and work that sits waiting.
• Set up the agent: set up an AI agent for the remaining repetitive work; the agent drafts, the team decides.
• Measure and standardize: measure time, errors and cost before and after; turn the working method into standard work.

How much authority does an agent get?

Authority grows step by step, and every step is based on measurement. I use four levels:

• 0. Shadow mode: the agent runs, but its output goes nowhere. It is only compared with the work people do.
• 1. Suggest: the agent prepares a draft; a person edits and approves it. Most setups start here, and many stay here.
• 2. Limited authority: the agent handles reversible, low-risk internal tasks (tagging, opening records, reminders) itself. Even at this level, nothing goes outside without approval.
• 3. Supervised authority: the agent runs broader tasks within written rules, monitored through regular spot checks and log reviews.
• At every level: limits are written into the system itself, not just the agent's instructions; every step goes into a log that cannot be edited; and there is a defined way to undo. You decide on raising the level, based on the measurement records.

Why process first, then AI?

Because an AI agent speeds up the flow you give it. If the flow contains an unnecessary approval, a double entry or work that sits waiting, the agent speeds that up too, or traps it in an endless loop.

What I often see on site: half of a task is preparing a report nobody reads or entering the same data into two systems. Removing that is cheaper and lasts longer than automating it.

Why does multidisciplinary engineering make a difference here?

AI automation looks like a software job, but it is really a process job. And a process rarely stays in one department: a customer complaint starts in quality and touches production, suppliers, purchasing and the ERP.

I have held direct responsibility across disciplines:

• production and operations management (lean manufacturing, TPS, OEE, line balancing)
• quality (IATF 16949, 8D, FMEA, SPC, APQP and PPAP, supplier audits)
• R&D to series production (RFQ, SOP timelines, setting up a plant from scratch)
• enterprise systems (SAP and Oracle ERP, machine monitoring, ERP and MES integration, Power BI)
• regulation and market readiness (CE and FDA roadmaps, GDPR, entering EU and US markets)
• company management (building ways of working with AI in finance, HR and business development)

Which industries do you have experience in?

I have worked on the ground in Türkiye, China and Bangladesh. My experience covers:

• automotive OEMs and suppliers (Hyundai and Ford vehicle programs, interior trim, seat covers)
• textiles and apparel (setting up a plant abroad, quality assurance)
• advanced technology and wearables, industrial IoT
• robotics and heavy industry
• medical IoT and academic R&D projects

Which work fits lean AI automation?

Work that repeats every week, cannot be fully written as rules, and has results that are easy to check. Examples:

• quality: drafting the problem description and timeline of an 8D report, FMEA and control plan documentation
• supply chain: supplier follow-up emails, lists of late orders
• operations: shift summaries, OEE summaries from downtime records
• sales and quoting: quote drafts from RFQ files, summaries from call notes
• management: weekly KPI reports, meeting summaries
• experts and small firms: inquiry follow-up, first replies and content drafts

How do you measure the result?

Before setup we measure the current state (the baseline); after setup we repeat the same measurement. We talk about results based on before-and-after measurement, not estimates. A few rules:

• One outcome metric, at most two supporting metrics: for example, proposal preparation time as the main metric, with rework rate and weekly usage as support. The more metrics, the less focus.
• Outcomes, not activity: "the agent wrote 300 drafts" is activity. "Proposals reach the customer faster" is an outcome.
• Lean measures: cycle time, waiting time, first-time-right rate, how many times a task changes hands, and the share of work completed without human intervention.
• Acceptance threshold: AI output is not identical every time. So instead of "it works or it doesn't", we accept against a threshold written up front, for example the share of drafts approved without edits. If the model or tool changes, we measure again.
• Gains reported separately: time saved, capacity gained and risk reduced are reported separately, not added together to inflate the result.

Who is this method for?

Production and technology companies, automotive suppliers, manufacturers exporting to the EU, and boutique firms and experts whose repetitive work keeps growing. I work across Türkiye and the European Union, remotely or on site; meetings in Turkish or English.

I have applied the method in my own work.

In automotive, working with lean manufacturing and TPS, I focused on raising line efficiency and cutting lead times. Today, at INVAMAR, I have set up a way of working with AI agents in finance, HR, operations and business development. Details and evidence levels are on the About and Case studies pages.