Discovery field guide / Industry
How to find AI opportunities in Manufacturing
Where should discovery start?
Start at a recurring production disruption and trace its evidence across operators, maintenance, quality, and planning. Look for repeated diagnosis, difficult-to-find instructions, and incomplete handovers before proposing predictive maintenance or autonomous control. A useful candidate must fit the actual asset, revision, shift, and operating conditions. First determine whether the cause is missing data, an unstable process, or interpretive work that assistance could improve.
The shared framework
Fourteen stages, from objective to portfolio.
Every guide uses this sequence to connect a business problem to an accountable investment. Revisit earlier stages when new evidence changes the assessment.
Stage 1: Business Objective
Name the outcome the organization is trying to move, so every later idea has something to be measured against.
Stage 2: Department
Locate the team that owns the objective and the work that feeds it.
Stage 3: Process
Describe how the work actually runs today: steps, handoffs, systems, and exceptions.
Stage 4: People
Capture who performs and depends on the work, and what they know that no document records.
Stage 5: Problem
State the friction in plain terms: delay, rework, error, cost, or risk, with evidence.
Stage 6: Opportunity
Frame a specific improvement that addresses the problem, not yet tied to any technology.
Stage 7: AI/Automation Fit
Compare AI, RPA, IDP, workflow automation, process redesign, and no technology change.
Stage 8: Value
Estimate benefit in the objective's own terms, with assumptions written down.
Stage 9: Feasibility
Check data, systems, skills, and change capacity needed to deliver.
Stage 10: Risk
Surface compliance, quality, security, and adoption risks and the controls they need.
Stage 11: Scoring
Rate each opportunity on the same dimensions so comparisons are consistent and explainable.
Stage 12: Prioritization
Sequence opportunities by score, dependencies, and capacity, with reasons people can challenge.
Stage 13: Investment
Match approved effort to the sequence; people with budget authority decide what gets funded.
Stage 14: Portfolio
Track opportunities over time as a living set: delivered, in flight, waiting, and retired.
Start with one observable journey
Choose one line or asset family and one recurring disruption. Review a normal run, a stoppage, and a quality hold. Walk the physical process with the people responsible and map when the records were created, not just when they were entered later.
Where to look across departments
Maintenance and shift handovers
Inspect fault notes, work orders, manuals, and parts searches. Identify where technicians reconstruct an asset's history or interpret inconsistent failure descriptions. If records lack asset identifiers or timestamps, basic maintenance-data discipline may be the prerequisite rather than an AI diagnostic layer.
Quality and engineering changes
Trace a nonconformance from detection through containment, disposition, and the latest approved instruction. Look for repeated document comparison and evidence assembly. Separate finding an applicable specification from changing the process or deciding whether a part can safely be released.
Who to interview and what each person knows
Operator, technician, and quality reviewer
Ask people on different shifts to show how they resolve the same exception. Interview a maintenance technician and the quality person who signs off disposition. Workarounds may be essential local knowledge, but they may also be unauthorized deviations that must not become automated instructions.
Engineering, safety, and plant systems
Include the process engineer, safety owner, production planner, and operational-technology or manufacturing-system owner. Ask who controls revisions and approves system connections. A corporate IT integration pattern may not be acceptable for an isolated plant network or safety-related equipment.
What to ask: follow the evidence
Which record explained the failure?
Ask what information was available before the stoppage and what was learned only afterward. Which manual revision or setting applied? What did a technician infer from physical inspection? Avoid assuming retrospective maintenance notes constitute an available real-time signal.
What must be verified on the floor?
Ask which recommendations require inspection, lockout procedures, quality approval, or engineering change control. How is an incorrect instruction detected and withdrawn? Capture how production pressure affects escalation and whether staff can safely decline a suggestion during a busy shift.
Signals of opportunity, and reasons to look elsewhere
Repeated interpretation of reliable plant evidence
A bounded search or handover task can be promising when the asset and instruction revision are known and a qualified person can verify the result. Sparse failure histories, changed equipment, and unrecorded conditions weaken claims that a predictive model can generalize.
Preventable investigation delay
Repeated manual reconstruction of a quality or maintenance history may indicate useful assistance. Check whether the delay comes from unavailable spares, staffing, or physical repair instead. A faster diagnostic summary has limited value if it does not change the bottleneck constraining the line.
Industry-specific risks and constraints
Physical safety and release authority
Wrong instructions can damage equipment or injure people. Keep established safety systems and professional authorization intact. A trial that helps locate a manual is not evidence for autonomous machine control, parameter changes, or quality release; those need separately defined engineering and safety assessment.
Plant variation and sensitive know-how
Evaluate network isolation, supplier permissions, intellectual property, and local data availability. Processes differ by site, machine revision, and product mix. Confirm that a proposed provider may receive the relevant drawings or process notes and that the workflow has an offline fallback where required.
How to evaluate what you find
Test across the actual operating envelope
Replay cases from different shifts and relevant asset revisions, including missing records and ambiguous symptoms. Check source accuracy, unsafe suggestions, abstention, and technician verification effort. Avoid evaluating with post-failure information that would not exist when the recommendation was needed.
Measure line-level consequences
Compare rework, diagnostic effort, quality disposition time, and the constraint affecting output. Include integration, sensor or record upkeep, and training costs. A local task improvement is not a throughput benefit unless the team can show how it changes the limiting step.
How to compare, prioritize, and invest
Start with advisory, inspectable work
Prefer one asset family's evidence retrieval or handover preparation where supervisors can review every result. Sequence records and revision-control repairs first. Treat safety approval and plant change windows as dependencies rather than assuming a high score makes the pilot ready.
Require a site-owned investment case
Compare candidates on reliable flow, reduced investigation and rework, maintenance capacity, and quality. Define stop conditions for incorrect asset matching or unsafe guidance. A plant owner should approve expansion only after local evidence supports it, not because another site's pilot looked promising.
Turn observations into a comparable opportunity record
Describe the current process before naming a solution. Record the trigger, owner, input, steps, handoffs, systems, output, and exception path. Separate what you observed from what someone recalled and what you still need to verify. A vivid interview story is a useful lead, not proof that every case behaves the same way. Validate it against a sample of ordinary work and difficult exceptions, with the limits of that sample made explicit.
Measure active work separately from waiting. Note the observation period, case mix, frequency, and source of each estimate. Include the receiving team's effort, checking, correction, and escalation so a local improvement does not merely transfer work. Describe the outcome in operational terms before attaching a monetary value. Capacity released is not automatically a budget saving; explain what can actually change and which manager has authority to make that change.
Compare interventions before selecting a tool
If a rule is explicit and the input is structured, test ordinary workflow automation first. If the task repeats a stable interface action, examine RPA and an integration alternative. If the difficulty is reading documents, consider IDP with source checks. AI may help where language or variable evidence requires interpretation, but its output still needs a task-specific acceptance standard. Process redesign or no technology change may be the right answer when policy, ownership, or an unnecessary step causes the problem.
Write down why each alternative would or would not address the observed cause. Do not compare a carefully controlled AI proposal with an artificially weak version of the current process. Include data repair, access approval, integration, training, review, and upkeep in each feasible option. An unavailable prerequisite is a reason to sequence work or pause, not a detail to hide inside a feasibility score.
Agree on a decision, not just a demonstration
Before a trial, name the business owner, evaluator, permitted data, review boundary, and person who can stop it. Define acceptable outputs and failures that must trigger escalation. Use examples held out from development, including missing information and uncommon cases. Compare the complete task with the baseline, including human verification and downstream consequences. Preserve the evidence behind the result so another owner can challenge the conclusion.
Keep value, feasibility, risk, and evidence confidence visible separately. A candidate with an attractive benefit estimate is not ready if its owner cannot obtain the data or operate the controls. Record the decision to investigate, pilot, wait, reject, or retire, with a reason and a condition for revisiting it. After delivery, compare observed results with assumptions rather than leaving the original estimate in the portfolio as fact.
Use the evaluation guide for the evidence standard and the prioritization guide to compare qualified candidates. Your leaders make funding decisions; a score or a discovery tool supplies structured evidence, not investment authority.
Continue through the discovery framework
Browse all departments and industriesRelated practice notes
For focused interview techniques, examples, and scoring detail alongside this field guide:
Map a plant exception in IntakeOS
Bring an approved process description, a sanitized incident, and the asset, safety, and sign-off boundaries. Explore how IntakeOS qualifies the improvement before deciding whether it needs AI, automation, or a foundational plant-process fix.
IntakeOS supports structured interviews, qualification, and comparison; it does not make funding decisions or replace professional review. Learn about VARA, review the trust page, or explore the platform before sharing sensitive information.
Explore discovery in an IntakeOS demo