AI is most useful in operations when it starts from the work. Which tasks are slow, error-prone, or repeated? Which can be standardized and measured? Who will check the result? Those questions come from Lean and quality practice, and they matter more than which tool is chosen.
This guide explains the kinds of AI you will meet (language assistants, classifiers, forecasting, agents, and plain rules-based automation), what each is good at, how to decide whether to automate, augment, or keep a task with a person, how to use the eight wastes and a value stream map to find opportunities, and a five-question readiness check. A worked complaint-handling example shows the method.
It does not cover how to build AI. Product capabilities change fast, so the guide describes types and controls and not named products.
Before You Start
Why Start With Operations, Not Technology
Most organizations that struggle with AI did not choose the wrong tool. They chose the wrong problem, or put AI on a process nobody had standardized, or never decided who would check the output. Operations work, such as process improvement, quality, safety, maintenance, and flow, is a good guide for all three, because it starts from the work and not from the tool.
Start From the Work
Ask which tasks are slow, error-prone, or repeated, then ask whether AI is the best fix. Often it is not.
Standardize First
Automating an unstable process produces faster variation. Lean and Six Sigma discipline comes before AI, not after.
Design the Human Role
Someone must check, correct, and own the output. That role has to be designed, staffed, and measured.
Measure the Result
Baselines, controls, and results are what separate a useful AI use from a demo.
The Kinds of AI You Will Meet in Operations
“AI” covers several different things. They behave differently, fail differently, and need different controls. Naming the type is the first step in choosing the right use.
| Type | What it does | Typical operations uses | Main watch-out |
|---|---|---|---|
| Generative language assistants | Read, write, summarize, translate, and answer questions in natural language | Drafting SOPs, summarizing incident reports, first-draft RCA, emails, meeting notes | Fluent but sometimes wrong, with confident invented details |
| Classifiers and extractors | Put items into categories or pull fields out of documents and images | Routing complaints, classifying near misses, reading invoices, visual inspection | Errors depend on the mix of cases; rare cases are the hardest |
| Forecasting and anomaly detection | Predict values or flag unusual patterns in data | Demand, maintenance, quality drift, energy use | Needs good history; can fail when conditions change |
| Agents | Pursue a goal by planning steps and using tools such as email, tickets, or databases | Triage work orders, chase exceptions, prepare orders | Can take wrong actions at speed; needs permissions and limits |
| Rules-based automation (not AI) | Follows fixed rules exactly | Routing by code, scheduled reports, approvals | Brittle when inputs vary, but predictable and cheap |
If a fixed rule can do the job reliably, use the rule. It is cheaper, easier to test, and gives the same answer every time. Use AI where the input varies and judgment from language, images, or patterns is needed.
What AI Is Good At, and Where It Struggles
| Good at | Often struggles with |
|---|---|
| Reading and summarizing large volumes of text | Precise arithmetic and exact look-ups without a tool to do them |
| Sorting items into categories from examples | Rare events and situations unlike its examples |
| Drafting first versions that a person improves | Knowing when it does not know |
| Finding similar past cases | Facts outside what it was given or trained on |
| Extracting fields from messy documents | Judgments that carry safety, legal, or ethical weight |
| Spotting patterns across many measurements | Explaining reliably why it produced an answer |
Automate, Augment, or Keep With a Person
Decide at the level of the task, not the job. Most jobs combine tasks that suit AI with tasks that do not.
| Choice | When it fits | Example |
|---|---|---|
| Automate with review | High volume, clear rules or examples, cheap to check, mistakes are small and fixable | Labeling and routing incoming complaints |
| Augment a person | Judgment is needed, but gathering and drafting take most of the time | Preparing an investigation summary for an engineer |
| Keep with a person | Serious or irreversible consequences, rare cases, ethical or legal weight, trust that depends on a human | Approving a corrective action that changes a safety control |
| Fix the process instead | The task exists because of waste or rework upstream | Eliminating a report nobody reads |
Human-in-the-Loop Process Design covers how much autonomy each task can carry, and Use-Case Selection and Business Case covers how to rank the candidates.
Lean First: Waste and AI
The eight wastes are a useful checklist for finding AI opportunities, and a reminder that some waste should be removed and not automated.
| Waste | Possible AI use | Caution |
|---|---|---|
| Defects | Flag likely defects or errors for checking; catch mistakes in documents | Measure the AI’s own error rate; see the AI in Quality guide |
| Overproduction | Better demand signals and plan suggestions | Forecasts fail when conditions change; keep a person on the plan |
| Waiting | Faster triage, routing, and first responses | Do not hide a queue problem behind faster routing |
| Non-utilized talent | Take drudge work away so people use their expertise | Involve the people whose work changes |
| Transportation | Route and schedule suggestions | Check the suggestions against constraints |
| Inventory | Spot slow movers and stock-out risk | Poor data gives poor suggestions |
| Motion | Find information faster; reduce searching | Make sure the answer is cited and checkable |
| Extra processing | Draft routine documents and reports | First ask whether the document is needed at all |
See 8 Wastes and Value Stream Mapping. A value stream map shows where time is lost and which steps are candidates, before any tool is chosen.
Finding Opportunities: Break Work Into Tasks
- Pick a value stream or process where there is pain: delay, rework, backlog, or cost.
- List the tasks in the process, in order, with who does each one.
- Record for each task: time per item, items per month, inputs and outputs, how often it goes wrong, and what a mistake costs.
- Ask what the task needs: reading, judging, deciding, or acting? Is there a clear rule or only examples? Is the data available and good enough?
- Mark each task as automate with review, augment, keep, or fix the process.
- List the candidates and score them with the AI Use-Case Prioritization Matrix.
Readiness: Five Questions Before You Start
| Question | Why it matters |
|---|---|
| Is the process standardized and documented? | AI learns and follows the pattern you give it. A process that differs by person gives inconsistent results and no way to judge them |
| Do we have the data, and can we use it? | Examples, history, and documents must be accessible, accurate enough, and allowed for this use |
| Is there a baseline measure? | Without today’s time, error rate, and cost you cannot show improvement |
| Who owns the result, and who checks it? | Every output needs an owner and a reviewer, named in advance |
| What happens when it is wrong? | If you cannot tolerate or catch the mistake, the use is not ready |
AI Across the Rest of the Site
| Area | Where AI helps | Where to go |
|---|---|---|
| Lean and Six Sigma | Collecting and summarizing data, drafting charters and A3s, finding patterns in the Measure and Analyze phases | AI Applications in Six Sigma; DMAIC Toolbox |
| Quality and inspection | Visual inspection, anomaly detection, document checks | AI in Manufacturing Quality |
| Root cause analysis | Drafting causal hypotheses, finding similar cases, summarizing evidence | AI for Root Cause Analysis |
| Safety | Classifying near misses, reading reports, spotting patterns | Safety Hub |
| Maintenance and reliability | Triage of work orders, prediction of failures | Reliability-Centered Maintenance |
| Leadership and change | Adoption, skills, trust, and the changing shape of jobs | Change Management for Improvement |
Worked Example: Complaint Handling at Riverside Plant
Riverside Plant, an illustrative 140-employee metal fabrication and assembly plant, handles about 600 customer complaints a month. A value stream map of the process shows 55 minutes of handling time per complaint, of which 30 minutes is investigation. The team marked each step. All figures are illustrative.
| Step | Minutes | Choice | Why |
|---|---|---|---|
| Receive and log | 2 | AI handles | Pulling fields from messy emails is repetitive; errors are easy to see |
| Classify and route | 4 | AI handles, person reviews | Clear categories and past examples; a wrong route is fixable, but safety complaints need a rule and review |
| Investigate | 30 | AI assists | Judgment is needed; AI can gather similar cases and history |
| Draft reply | 10 | AI assists | A first draft saves time; a person owns the wording |
| Approve | 6 | Keep with a person | Cost, legal, and safety consequences |
| Send and log | 3 | AI handles | Rule-like and checkable |
What the team decided. Before any tool was chosen they fixed two things the map exposed: three complaint forms that collected different information, and a rework loop where replies were sent back for missing facts. Then they ranked the candidates, and the first pilot was classification and routing, because it was high volume, checkable, and low in risk if safety complaints get their own rule. The use-case ranking and the business case are in the next guide.
Common Mistakes
Tool First
Buying or building a tool, then looking for something to do with it.
Automating a Mess
Putting AI on a process that varies by person, with no baseline.
Replacing Whole Jobs
Thinking in jobs and not tasks, and missing the tasks that should stay with people.
No Owner
Nobody checks the output or owns the result.
Counting Only Savings
Ignoring review time, error cost, and running cost.
Skipping the People
Changing someone’s work without involving them.
Self-Assessment Questions
- Can we name the tasks, not just the jobs, where AI might help, and what each costs today?
- Have we removed waste and standardized the process before adding AI?
- For each candidate, do we know whether to automate, augment, keep, or fix?
- Do we know who will check the output and what happens when it is wrong?
- Do we have a baseline to compare against?
Where AI Fits in Operations: Frequently Asked Questions
Where should an operations team start with AI?
With a process that has a clear pain, such as delay, rework, or backlog, and a baseline measure. Map the work, remove obvious waste, break it into tasks, and pick a task that is high volume, easy to check, and low in risk if it goes wrong. Pilot it with a named owner and reviewer.
What is the difference between automation and AI?
Rules-based automation follows fixed instructions exactly and is predictable. AI handles variation, such as reading free text or images, by working from examples, and it can be wrong in ways a fixed rule would not be. Use a rule when a rule works, and AI where the input varies and judgment from language, images, or patterns is needed.
Should we fix the process before using AI?
Usually yes. AI on an unstable or undocumented process gives inconsistent results that you cannot judge. Standardizing the process and removing waste first often delivers most of the benefit and makes the remaining AI task simpler.
Which tasks should stay with people?
Those with serious or irreversible consequences, rare or unprecedented cases, legal or ethical weight, or where trust depends on a human. AI can still help such tasks by gathering information and drafting, while a person decides.
Do we need data scientists to use AI in operations?
Not for every use. Many operations uses rely on ready-made assistants and tools configured by process owners. You do still need people who can define the task, measure results, manage risk, and review outputs, which are operations skills.
Sources and Further Reading
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (January 2023).
- ISO/IEC 22989:2022, Artificial intelligence: concepts and terminology (check current edition).
- Womack, J. P. and Jones, D. T., Lean Thinking (Free Press, 2003); Ohno, T., Toyota Production System (Productivity Press, 1988), on the seven wastes and standardization.
- Bainbridge, L., “Ironies of automation,” Automatica 19(6), 1983.
- Parasuraman, R., Sheridan, T. B. and Wickens, C. D., “A model for types and levels of human interaction with automation,” IEEE Transactions on Systems, Man, and Cybernetics, Part A 30(3), 2000.
- OECD, Recommendation of the Council on Artificial Intelligence (2019, updated 2024), for the definition of an AI system.