A process-led field guide for internal forward deployed engineers. Map the operation, investigate the causes of lost value, size the opportunity and inspect how the work could change.
By the Forward Deployed Team · Updated 24 October 2026 · Illustrative scenarios; not customer results.
Start with the operating outcome. Follow the end-to-end process, identify a specific cause of lost value, then test the intervention. Eight process families below contain 24 working hypotheses; the visual investigations that follow illustrate selected mechanisms.
Demand, specification, purchasing, process effort and loss. Distinguish spend removed from capacity released.
03 / Capital resilience
Improve capital decisions
Asset use, investment discipline, liquidity and resilience. Preserve service and essential controls.
Evaluate every opportunity through: people and responsibilities · process and standard work · data and information · controls and decision rights · technology and adoption. AI is one possible enabler.
These are starting hypotheses, not diagnoses. Establish the baseline and the burden of proof before selecting a tool. All process families stay open as you scroll.
Manufacturing / 01
Plan and schedule production
01Demand
02finite capacity
03sequence
04dispatch
05reschedule
01.01
Schedule instability may create unnecessary setups
01.02
Material readiness gaps may starve scheduled lines
01.03
Product mix may overload the actual bottleneck
Proof or rejection: what evidence is needed?
Replay schedules against actual constraints, material availability and good output; reject plans that assume unlimited labor or machine capacity.
Where AI could contribute
Constraint simulation and schedule explanation; planners and operators validate feasible sequences.
Size the opportunity / Bottleneck capacity contribution
Recovered bottleneck hours × good units per hour × contribution per unit
Low / base / high illustration: 100 / 200 / 300 hours × 50 good units × $8 contribution = $40k / $80k / $120k potential annual contribution.
Value requires saleable demand and no downstream constraint. Do not count the same hours again as labor savings or downtime savings.
Unstable setup methods may require repeated first-off adjustments
Proof or rejection: what evidence is needed?
Observe complete changeovers across crews and product families; separate safe external preparation from work requiring isolation. Reject unsafe parallel work.
Where AI could contribute
Analyze task timing and retrieve approved setup methods; camera observations can establish sequence with consent and validation.
Size the opportunity / Bottleneck capacity contribution
Recovered bottleneck hours × good units per hour × contribution per unit
Low / base / high illustration: 100 / 200 / 300 hours × 50 good units × $8 contribution = $40k / $80k / $120k potential annual contribution.
Value requires saleable demand and no downstream constraint. Do not count the same hours again as labor savings or downtime savings.
Overfill and trim loss may exceed specification needs
04.03
Late defect detection may magnify downstream rework
Proof or rejection: what evidence is needed?
Reconcile material mass balance, defect labels, process settings and customer outcomes. Reject visual checks that miss hidden defects or create excessive false rejects.
Where AI could contribute
Vision supports visible-defect review; process models investigate drift; validated controls govern containment.
Size the opportunity / Annual material-cost opportunity
Annual input units × verified reject-rate reduction × avoidable material cost
Low / base / high illustration: 1m units × 0.5 / 1.0 / 1.5 percentage points × $4 = $20k / $40k / $60k gross annual material opportunity.
Deduct false rejects, extra inspection, rework and operating costs. Do not count contribution and material saving twice.
Analyze failure history, work orders, condition and load; compare avoided work with inspection and false-alarm cost. Reject diagnosis unsupported by technician evidence.
Where AI could contribute
Anomaly detection and source-linked repair guidance; technicians approve interventions.
Size the opportunity / Bottleneck capacity contribution
Recovered bottleneck hours × good units per hour × contribution per unit
Low / base / high illustration: 100 / 200 / 300 hours × 50 good units × $8 contribution = $40k / $80k / $120k potential annual contribution.
Value requires saleable demand and no downstream constraint. Do not count the same hours again as labor savings or downtime savings.
Manual video review may delay usable event evidence
08.02
Unclear policy thresholds may cause inconsistent escalation
08.03
Disconnected observations may require repeated investigation
Proof or rejection: what evidence is needed?
Use independently reviewed recordings and explicit policy rules; measure missed events, false alerts and time to usable evidence. Reject unsupported event interpretations.
Where AI could contribute
Visual perception provides observations; explicit policies evaluate them; supervisors review evidence and decide the response.
Estimate released processing capacity from your own workload. This is an effort model, not a prediction of AI savings. Use the separate formulas above for cash, capital, margin, materials and energy.
Formula: tasks × minutes ÷ 60 × realization × hourly cost − running cost; subtract one-time cost in year one. Financial savings require an actual budget change. No inputs leave your browser.
Method and sizing notes
The opportunity structure combines three value lenses—cash release, cost optimization and capital resilience—with process, contract, working-capital, customer-management and investment analysis. Historical cost-transformation and finance operating-model references inform the process coverage. Every hypothesis requires evidence, a test that could reject it and accountable validation.
Historical transformation benefit ranges are not current benchmarks or AI effect sizes. This guide uses transparent low/base/high sensitivity examples instead. The eligible baseline, proposed change and realization route must be established for the organization. The examples are not additive, forecasts or customer results.
Prioritize by net value, strength of evidence, feasibility, time to value and service or control consequences. Deduct implementation, transition and recurring costs; assign one owner and one benefit record to prevent double counting.
01
Plant operations
Direct labor & changeovers
The working hypothesis
Overtime, uneven staffing, long changeovers and micro-stoppages may constrain output more than installed capacity.
Follow this example
A changeover consumes the production window. Use the control to inspect what changes, then decide what evidence would be needed to test the hypothesis.
What would establish the case?
Observe shifts and changeovers; identify causes of lost minutes and test standard work, staffing or task redesign.
Enablers to investigate
Operator interviews · work observation · time study · vision-assisted event tagging · simulation
WORKED EXAMPLE 01Synthetic scenario
A changeover consumes the production window
BASELINE / 60 MINIsolateFind toolsMaterialsSetup
REVISED / 30 MINIsolateSetup30 min released
Baseline60
→
Revised / examined30
changeover minutes
Illustrative process animation
Task
Baseline minutes
Revised minutes
Stop + isolate
10
10
Fetch tools
15
0
Locate materials
15
0
Change + verify setup
20
20
Task
Baseline minutes
Revised minutes
Before stop: stage tools
15
External preparation
Before stop: stage material
15
External preparation
Stop + isolate
10
10
Change + verify setup
20
20
The scenario removes 30 minutes of searching from machine downtime. Validate the task sequence and isolation procedure before changing work.
Explore the applications
AI & physical AI use cases
4 visual examples · inputs, analysis and operational action
+
These are candidate applications. The useful intervention depends on the task, data quality, variation and operating constraints. All miniatures below are illustrative mockups.
AI · records, interpretation & decisions
Shift allocation scenario
INPUT
Demand + skills + availability
ILLUSTRATIVE RESULT
Compare staffing and overtime constraints
AI · records, interpretation & decisions
Changeover instruction assistant
SOURCE EVIDENCE
→
REVIEW NOTE
Retrieve correct setup checklist
CHECK SOURCE ↗
INPUT
Approved SOP + equipment / product
ILLUSTRATIVE RESULT
Retrieve correct setup checklist
Physical AI · sensing, perception & simulation
Micro-stoppage classification
INPUT
Line video + machine-state timestamps
ILLUSTRATIVE RESULT
Tag candidate stop causes for validation
Physical AI · sensing, perception & simulation
Cobot task feasibility
INPUT
Task path + payload + tolerances
ILLUSTRATIVE RESULT
Simulate reach / cycle time before pilot
MEASURE IN THE REAL OPERATION
Good units per labor hour · overtime · changeover duration
02
Plant operations
Manufacturing overhead
The working hypothesis
Idle equipment, peak loads, fragmented indirect work and consumable variability may create avoidable overhead.
Follow this example
The line is idle; the compressor is not. Use the control to inspect what changes, then decide what evidence would be needed to test the hypothesis.
What would establish the case?
Build a driver-level baseline by line, product and operating state; distinguish production needs from idle consumption.
Enablers to investigate
Submetering · machine-state signals · operating envelopes · energy analysis
WORKED EXAMPLE 02Synthetic scenario
The line is idle; the compressor is not
Baseline90
→
Revised / examined60
kWh across the sample hour
Illustrative process animation
Operating state
Minutes
Power
Production
30
120 kW
Idle
30
60 kW
Operating state
Minutes
Power
Production
30
120 kW
Idle / approved standby
30
0 kW in this simplified case
Baseline: 120 × 0.5 + 60 × 0.5 = 90 kWh. Revised: 60 kWh. A zero-power standby assumption needs equipment and process validation.
Explore the applications
AI & physical AI use cases
4 visual examples · inputs, analysis and operational action
+
These are candidate applications. The useful intervention depends on the task, data quality, variation and operating constraints. All miniatures below are illustrative mockups.
AI · records, interpretation & decisions
Energy scheduling scenarios
Expected
Observed
Explain gap
Illustrative comparison → inspect drivers
INPUT
Production plan + tariff periods
ILLUSTRATIVE RESULT
Move eligible load outside peak window
AI · records, interpretation & decisions
Overhead driver explanation
RECORDRULE / EXPECTEDREVIEW
Sample 01Consistent✓
Sample 02Difference found⚑
Sample 03Consistent✓
INPUT
Spend + line / product activity
ILLUSTRATIVE RESULT
Separate volume effects from rate changes
Physical AI · sensing, perception & simulation
Compressed-air loss detection
INPUT
Acoustic / pressure measurements
ILLUSTRATIVE RESULT
Leak candidate → maintenance survey
Physical AI · sensing, perception & simulation
Idle-state supervision
INPUT
Machine state + power draw
ILLUSTRATIVE RESULT
Idle with high power → approved review
MEASURE IN THE REAL OPERATION
Energy per good unit · peak demand · overhead per unit
03
Plant operations
Value leakage & quality
The working hypothesis
Material loss, process variation, overfill and rework may erode value between specification and actual execution.
Follow this example
A seal defect passes a busy inspection point. Use the control to inspect what changes, then decide what evidence would be needed to test the hypothesis.
What would establish the case?
Trace material mass balance, quality defects and process deviations; verify the highest-value loss points on the floor.
Enablers to investigate
Line camera · lighting · labeled defects · vision AI · operator review / validated reject logic
WORKED EXAMPLE 03Synthetic scenario
A seal defect passes a busy inspection point
P-101PASS
P-102PASS
P-103REVIEW
P-104PASS
P-105REVIEW
P-106PASS
P-107REVIEW
P-108PASS
Sample packs8
→
Review candidates3
suspect packs in sample
Illustrative process animation
Pack
Visible feature
Disposition
P-101
Seal aligned
Pending
P-102
Seal aligned
Pending
P-103
Open edge
Pending
P-104
Seal aligned
Pending
P-105
Wrinkled edge
Pending
P-106
Seal aligned
Pending
P-107
Missing seal
Pending
P-108
Seal aligned
Pending
Pack
Visible feature
Disposition
P-101
Seal aligned
Pass
P-102
Seal aligned
Pass
P-103
Open edge
Review
P-104
Seal aligned
Pass
P-105
Wrinkled edge
Review
P-106
Seal aligned
Pass
P-107
Missing seal
Review
P-108
Seal aligned
Pass
The highlighted packs are predefined example labels, not computer-vision inference. A real model needs labeled samples, false-reject measurement and validation at line speed.
Explore the applications
AI & physical AI use cases
4 visual examples · inputs, analysis and operational action
+
These are candidate applications. The useful intervention depends on the task, data quality, variation and operating constraints. All miniatures below are illustrative mockups.
AI · records, interpretation & decisions
Scrap root-cause exploration
Expected
Observed
Explain gap
Illustrative comparison → inspect drivers
INPUT
Scrap by batch / shift / recipe
ILLUSTRATIVE RESULT
Association to investigate, not causation
AI · records, interpretation & decisions
Supplier quality evidence review
SOURCE EVIDENCE
→
REVIEW NOTE
Missing certificate / recurring defect
CHECK SOURCE ↗
INPUT
Inspection notes + certificates
ILLUSTRATIVE RESULT
Missing certificate / recurring defect
Physical AI · sensing, perception & simulation
Seal and surface inspection
INPUT
Line images + labeled defect library
ILLUSTRATIVE RESULT
Suspect pack → reviewed disposition
Physical AI · sensing, perception & simulation
Overfill and yield control
INPUT
Checkweigher + target fill + variation
ILLUSTRATIVE RESULT
Persistent giveaway → process review
MEASURE IN THE REAL OPERATION
First-pass yield · scrap · giveaway · rework
04
Plant operations
Maintenance & repair
The working hypothesis
Reactive repairs, inconsistent procedures and poorly matched spares may increase downtime and maintenance cost.
Follow this example
A pump trend changes before a work order is raised. Use the control to inspect what changes, then decide what evidence would be needed to test the hypothesis.
What would establish the case?
Review failure history, critical assets and repair-versus-replace decisions; pilot risk-based maintenance on selected assets.
A pump trend changes before a work order is raised
08:00 reading4.1
→
14:00 reading6.3
mm/s vibration in sample
Illustrative process animation
Sample
Vibration
Context
08:00
4.1 mm/s
Normal load
10:00
4.5 mm/s
Normal load
12:00
5.8 mm/s
Normal load
14:00
6.3 mm/s
Normal load
Sample
Vibration
Context
Review flag
Trend change
Not a diagnosis
Evidence to collect
Bearing / alignment / load
Technician inspection
Next action
Planned assessment
No automatic shutdown
The example flags a change for inspection. It does not diagnose a bearing failure or define a safe operating threshold.
Explore the applications
AI & physical AI use cases
4 visual examples · inputs, analysis and operational action
+
These are candidate applications. The useful intervention depends on the task, data quality, variation and operating constraints. All miniatures below are illustrative mockups.
The route lengths are illustrative layout assumptions. Actual benefits need order baskets, travel paths, replenishment demand and congestion.
Explore the applications
AI & physical AI use cases
4 visual examples · inputs, analysis and operational action
+
These are candidate applications. The useful intervention depends on the task, data quality, variation and operating constraints. All miniatures below are illustrative mockups.
AI · records, interpretation & decisions
Slotting recommendation
INPUT
SKU velocity + basket combinations
ILLUSTRATIVE RESULT
Fast movers closer to dispatch
AI · records, interpretation & decisions
Replenishment exception planning
RECORDRULE / EXPECTEDREVIEW
Sample 01Consistent✓
Sample 02Difference found⚑
Sample 03Consistent✓
INPUT
Pick demand + current slot stock
ILLUSTRATIVE RESULT
Prioritize likely stockout locations
Physical AI · sensing, perception & simulation
AMR material movement
INPUT
Layout + payload + crossing traffic
ILLUSTRATIVE RESULT
Simulate replenishment route and queue
Physical AI · sensing, perception & simulation
Pick and pallet verification
INPUT
Barcode + package / pallet image
ILLUSTRATIVE RESULT
Count / identity mismatch → review
MEASURE IN THE REAL OPERATION
Lines per hour · travel · accuracy · cost per order
Visual observations linked to explicit operating policies may help supervisors identify events that need review and assemble their evidence faster.
A camera feed is an input. The work is to interpret what is visible, evaluate an applicable policy, preserve the evidence and route the result for human review. The robot recordings below illustrate this sequence.
01
Observe
Video frames and their timestamps
02
Interpret
Visible people, equipment and interactions
03
Check policy
Evaluate the attached rules against the observation
04
Review evidence
Inspect the result, receipt status and alert
What would establish the case?
Use a representative, independently reviewed event set. Measure missed events, false alerts, time to a usable review and whether the evidence supports the policy decision. Compare the full workflow with the current review process.
Enablers
Task and policy interviews · representative video · a validated perception model · explicit operating rules · timestamped evidence · accountable reviewers
Interactive visual-perception example
Video → perception → policy → evidence
Choose the industrial work cell or robot-interaction recording, start analysis, then inspect the scene interpretation, rule results and evidence queue.
The video inputs are recordings; perception and policy responses are requested from the connected service. If the service is unavailable, the workspace shows that state. Results support investigation and human review; this example does not command a robot or certify its safety.
Understand the intervention
Perception and policy have different jobs
Two operating examples · observation, decision and evidence
+
Visual perception / interpret an observation
Work-cell activity review
Interpret the visible relationship between personnel and equipment using a validated model and the current video window. Keep uncertainty and the source timestamp with the observation.
VIDEO WINDOW → SCENE INTERPRETATION
The connected model supplies the observation. No person, object box or interaction is pre-filled on this page.
Policy evaluation / apply an explicit rule
Operating-policy review
Evaluate the observation against the applicable rule pack. Keep the matched rule, evidence status and human review step distinct from the perception result.
OBSERVATION + RULE → REVIEW QUEUE
A policy result appears only when the service returns one. A service error is not evidence of a policy breach.
SIZE AND VALIDATE THE OPPORTUNITY
Review hours × verified effort reduction × loaded hourly cost, less operating cost. Track time to usable evidence, missed events and false alerts separately. Any downtime or risk benefit needs its own validated business case.
Quick answers
From an example to an investigation.
What is forward deployment in manufacturing?
It is the work of taking a specific operating problem from hypothesis through evidence, process design, pilot and adoption with the people who own the operation. It includes responsibilities, controls and measured outcomes.
How do AI and physical AI contribute to manufacturing improvement?
AI can help interpret records, compare scenarios and prepare decisions. Physical AI combines sensing, perception and potentially physical action. Both need validation against the actual task, environment and operating constraints.
How should a manufacturing AI opportunity be validated?
Define the task and baseline, gather transaction data and work observations, test variation and exceptions, run a bounded pilot and measure outcomes after implementation and operating costs.
Are the numbers on this page customer results?
No. The interactive operating scenarios use synthetic data and simplified assumptions. The videos are example inputs. The integrated workspace requests scene interpretation and policy evaluation from the connected service and reports service failures explicitly.
Does every opportunity need a robot or AI model?
No. Interviews, clearer responsibilities, standard work or better data may be the most useful intervention. Technology selection follows the validated operating need.
