Automated Visual Inspection
Illustrative case Fictional example
This fictional business case demonstrates decision analysis and approval-document design. The manufacturer, line, products, defect data, figures, and outcomes are illustrative; they are not evidence of work performed for a real client.
- Domain
- Manufacturing quality automation
- Delivery approach
- Hybrid: predictive equipment, iterative model validation
- Scenario
- Retrofit automated visual inspection on a stable line to reduce escaped defects without disrupting throughput or safe operation.
- Decision sought
- Approve a gated retrofit for one production line, with authority to stop after data feasibility or pilot testing.
- Investment scale
- USD 1.18 million including contingency
- Planning horizon
- Five years from approval
- Primary constraints
- Representative defect data, false rejects, line availability, controls integration, and machine safety.
- Evidence basis
- Illustrative production, defect, inspection, labor, downtime, and model-performance assumptions.
- Demonstrates
- AI investment appraisal under performance uncertainty, gated validation, workforce impact, and measurable acceptance criteria.
Document Control
| Field | Detail |
|---|---|
| Sponsor / benefit owner | VP, Manufacturing |
| Business-case owner | Plant Director |
| Decision body | Capital Investment Committee |
| Version / status | 1.0 / For approval |
| Decision date | 30 September 2026 |
| Appraisal period | Five years from approval |
| Funding requested | USD 1.18 million, including contingency |
| Next review | Completion of data-feasibility gate |
Executive Summary
Approval is requested to retrofit one stable production line with an automated visual-inspection station and controlled manual fallback. The line ships 2.4 million units annually. Manual inspection detects an estimated 90% of visible in-scope defects, allowing about 4,800 defective units (2,000 parts per million, or ppm) to escape and incorrectly rejecting 8,400 conforming units each year.
The recommended retrofit combines proven cameras, lighting, guarding, controls, and reject handling with an iteratively developed vision model. Predictable hardware, procurement, safety, installation, and shutdown work will use stage-based planning; data and model work will use short validation cycles. Expected full-year cash benefit is USD 490,000, offset by USD 145,000 annual operating cost. The five-year base case has a USD 24,000 net present value (NPV) at 8% and a 4.0-year undiscounted payback. A further USD 240,000 annual capacity benefit from redeploying four employees is disclosed but excluded from cash returns.
The economics are marginal and sensitive to performance. Approval should therefore release funding by gates, with authority to stop after feasibility or pilot stages. No autonomous rejection will begin until quality and safety acceptance criteria are met.
Strategic Case and Current State
The investment supports fewer customer escapes, better traceability, and more capacity without added routine inspection labor. The stable line and product family favor retrofit over replacement.
The following twelve-month baseline will be independently validated before the offline pilot:
| Measure | Baseline and assumption |
|---|---|
| Annual shipped production | 2,400,000 units |
| Line takt | 8.0 seconds; inspection is currently off-line and does not constrain takt |
| Visible in-scope defects produced | 48,000 (2.0%): surface 24,000; assembly 14,400; label/marking 9,600 |
| Manual detection | 90% overall; 43,200 defects contained |
| Escaped visible defects | 4,800, or 2,000 ppm |
| False rejects | 8,400 conforming units, or 0.35% of production |
| Inspection staffing | Eight permanent FTE plus USD 120,000 annual agency/overtime spend |
Labels are inconsistent, and rare critical defects have few examples. Automation feasibility remains unproven.
Objectives and Success Measures
Acceptance will require 30 consecutive production days across all approved shifts and product variants, measured against quality-engineer adjudicated samples.
| Outcome | Acceptance target |
|---|---|
| Defect detection | At least 97.0% overall and 99.5% for defined critical classes |
| False reject rate | At most 0.10% of conforming units |
| Cycle time and availability | No increase above 8.0-second takt; inspection availability at least 99.0% |
| Escaped visible defects | At most 750 ppm within six months of acceptance |
| Safety | Zero recordable injuries attributable to the station; validated guarding, interlocks, emergency stops, and safe reject access |
Constraints and Dependencies
- Installation must fit the scheduled 48-hour shutdown; an overrun would affect customer supply.
- Existing programmable controls, quality systems, network segregation, floor space, and utilities must support the station.
- Quality must provide representative, adjudicated images covering shifts, suppliers, normal variation, and every approved variant.
- Procurement lead times for cameras and controls are assumed to be no more than 20 weeks.
- Works council and workforce consultation, role training, cybersecurity approval, and machinery safety assessment are prerequisites.
- The current process must remain stable; unapproved product, lighting, supplier, or packaging changes invalidate model evidence.
Options
- Enhance manual inspection: improve standards, lighting, training, audits, and sampling. Disruption is low, but subjectivity, fatigue, limited traceability, and recurring labor remain.
- Retrofit an automated station: add controlled imaging, model-assisted classification, reject handling, traceability, and manual fallback. This targets the problem without replacing productive assets.
- Replace the line: procure integrated inspection with a new line. This is disproportionate while the current line is reliable and unconstrained.
Options Appraisal
Scores are 1 (poor) to 5 (strong). Weighted totals are out of 5.
| Criterion | Weight | Manual | Retrofit | Replace line |
|---|---|---|---|---|
| Quality and strategic outcome | 30% | 2 | 5 | 5 |
| Five-year value / affordability | 25% | 2 | 4 | 2 |
| Technical and operational feasibility | 20% | 5 | 4 | 2 |
| Delivery disruption | 15% | 5 | 3 | 1 |
| Flexibility / reversibility | 10% | 3 | 4 | 2 |
| Weighted total | 100% | 3.15 | 4.15 | 2.75 |
Enhanced manual inspection costs an estimated USD 180,000 initially plus continuing staffing and would not reliably meet the escape target. Line replacement is indicatively USD 7–9 million with 9–12 month interruption risk. Retrofit is the strongest proportional response, subject to validation.
Recommended Option and Scope
Approve Option 2 for one line. Scope includes imaging enclosure and lighting, cameras and edge compute, controls and safe reject mechanism, integration to line and quality records, dataset creation, model development, cybersecurity, validation, training, operating procedures, spares, and manual fallback. Product redesign, upstream process correction, other lines, and fully unattended quality release are excluded. Operators remain accountable for safe operation; Quality retains disposition authority.
Financial Case
All figures are rounded USD, exclude tax and financing, and use constant prices.
| Implementation cost | USD thousands |
|---|---|
| Equipment, lighting, compute, and spares | 420 |
| Controls, integration, guarding, and safety validation | 180 |
| Data preparation and model development | 160 |
| Installation, shutdown, and commissioning | 110 |
| Training and workforce transition | 45 |
| Project management and independent assurance | 65 |
| Contingency (20% of USD 980,000 base) | 196 |
| Funding envelope, rounded | 1,180 |
| Full-year recurring cash benefit | Calculation | USD thousands |
|---|---|---|
| Fewer escaped defects | 3,360 additional defects contained × (USD 90 external less USD 12 internal handling) | 262 |
| Fewer false rejects | (0.35% − 0.10%) × 2.4m units × USD 18 | 108 |
| Removed agency/overtime spend | Approved reduction in external and overtime budget | 120 |
| Gross cash benefit | 490 | |
| License, maintenance, calibration, data/model support, and cyber operations | (145) | |
| Net annual cash benefit from year 2 | 345 |
Four permanent inspectors are expected to move into existing production and quality vacancies, creating capacity valued at approximately USD 240,000 annually (4 × USD 60,000 loaded cost). No redundancy is planned, payroll does not fall, and this capacity is not counted as cash benefit. The USD 120,000 benefit is counted only after agency contracts and overtime budgets are actually removed.
| Year from approval | 0 | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|---|
| Implementation | (900) | (280) | — | — | — | — |
| Gross benefit | — | 245 | 490 | 490 | 490 | 490 |
| Operating cost | — | (110) | (145) | (145) | (145) | (145) |
| Net cash flow (USD thousands) | (900) | (145) | 345 | 345 | 345 | 345 |
At an 8% discount rate, five-year NPV is USD 24,000: −900 − 145/1.08 + 345/1.08² + 345/1.08³ + 345/1.08⁴ + 345/1.08⁵. Undiscounted payback is 4.03 years from approval.
| Sensitivity (same cost timing) | Five-year NPV | Payback |
|---|---|---|
| Cash benefits 20% lower | (USD 322,000) | Beyond five years (about 5.4 years if sustained) |
| Base case | USD 24,000 | 4.0 years |
| Cash benefits 20% higher | USD 370,000 | 3.25 years |
A 10% implementation overrun alone reduces NPV to approximately (USD 92,000). Consequently, validated benefit and cost forecasts must preserve non-negative NPV at each investment gate unless the committee records a strategic exception.
Commercial and Hybrid Delivery Approach
Procurement will seek three integrator bids. Fixed-price milestones cover hardware, guarding, controls, installation, documentation, and shutdown commitments; capped time-and-materials covers labeling and iterative model development. Payments retain 15% until acceptance.
Contracts must grant the buyer ownership or perpetual use of labeled data, model artifacts, interfaces, and configuration; require export in documented formats; prohibit vendor use of production images without consent; define security, patching, spare-part, response-time, and exit obligations; and price five-year support. Open interfaces, source escrow for custom control code, and a tested manual mode reduce vendor lock-in.
Phases and Gates
| Phase | Evidence and gate decision |
|---|---|
| Data feasibility | Collect representative images; proceed only with agreed labels, at least 90% reviewer agreement, coverage of all variants, and sufficient critical-defect evidence or an approved test-generation plan. |
| Offline pilot | Blind holdout test; proceed only if detection and false-reject targets are statistically supported and inference fits the takt budget. |
| Line pilot | Run in shadow mode, then supervised reject mode; proceed only after no throughput loss above 1%, stable shift/variant performance, and approved safety design. |
| Installation | Execute rehearsed 48-hour plan with rollback, spares, and manual inspection ready; energize only after safety and cyber authorization. |
| Acceptance | Complete the 30-day test against all success measures, operator competency, maintainability, traceability, and disaster-recovery checks. |
Stop the project if representative data cannot support critical-class validation, a safety requirement cannot be engineered, forecast NPV becomes negative without approved exception, or the shutdown plan lacks a tested rollback. During operation, immediately disable automated rejection and revert to manual inspection after an interlock failure, a confirmed critical escape, false rejects above 0.5% for two consecutive hours, or throughput degradation above 5% for 30 minutes.
Risks, Disbenefits, and Responses
| Risk or expected disbenefit | Response |
|---|---|
| Rare or changing defects reduce detection | Golden samples, adjudication, conservative manual review, and targeted data collection |
| Model drift from suppliers, lighting, wear, or variants | Drift dashboard; weekly sampling; monthly quality review; revalidation before changes |
| False rejects increase scrap or stoppages | Soft alerts first, reject-rate alarms, quarantine rather than scrap, manual override |
| Guarding/reject mechanism creates safety hazards | Formal risk assessment, validated safety functions, lockout/tagout, ergonomic review |
| Vendor lock-in and support failure | Data portability, open interfaces, escrow, priced exit assistance, critical spares |
| Shutdown overrun or integration failure | Factory acceptance test, simulation, rehearsed rollback, manual contingency, schedule reserve |
| Surveillance concerns, anxiety, and deskilling | No people analytics; consult workers; publish data purpose; retrain and redeploy; retain diagnostic skills |
| Added maintenance and review workload | Fund named controls, quality, IT, and model owners within recurring operating cost |
Benefits Realization
Quality owns detection, false rejects, and escapes; Production owns takt, availability, and capacity; Finance verifies avoided costs and removed agency/overtime budgets; EHS owns safety. Baselines freeze before pilot. Reporting is weekly for 90 days, monthly through year one, then quarterly. Finance recognizes only ledger-evidenced savings; redeployed labor remains a separate capacity measure.
Governance and Review Triggers
The sponsor chairs a monthly steering group with Quality, Production, EHS, Engineering, IT/cybersecurity, Finance, Procurement, and workforce representation. The committee approves each gate, contingency draw above USD 50,000, and acceptance. Reopen this case as an early warning if forecast cost rises by more than 5%, schedule slips beyond the shutdown window, expected benefit falls by 15%, NPV turns negative, product mix changes materially, a critical safety or quality event occurs, or the vendor/support model changes. Any forecast above the not-to-exceed USD 1.18 million approval envelope requires explicit committee reapproval before further commitment or expenditure; gate approval, contingency authority, or a strategic NPV exception does not authorize overspend.
Decision Required and Approval Conditions
Approve a not-to-exceed USD 1.18 million gated envelope for Option 2 and authorize procurement and data feasibility only. Release later phases when the committee confirms: validated baseline and representative data; acceptable workforce consultation; non-negative updated NPV or documented strategic exception; forecast cost at or below the approved envelope, or explicit reapproval of any higher forecast; fixed commercial protections; approved cyber and machinery-safety designs; a 48-hour installation and rollback plan; named operational owners; and funded manual fallback. Acceptance transfers the station to operations only after all 30-day criteria pass; otherwise the committee must remediate, narrow scope, revert, or stop.