Start — business fit, ROI, and pilot approval

Last updated: 2026-09 · owner: Youngjin · volatility: medium

L0 TL;DR: Define the business task and current performance first. Compare existing automation, commercial solutions, and a system integrator (SI)¹; choose data, simulation, or VLA only when learning is needed. Customers and AWS staff complete this worksheet together.

Paths by role

Role	Read first	Meeting output
Customer decision maker	This page → Executive Brief	One task, spending cap, proceed/hold criteria
Customer engineering team	Execution paths → relevant pillar → Operations and recovery	Prerequisites, experiment plan, owners
AWS staff	Conversation guide → Decisions	Discovery questions, alternatives, support/partner roles

Staff guidance is also public. Do not record internal contacts or confidential customer material on this site.

Business fit — eight discovery questions

Question	Complete with the customer
What task needs improvement?	Process, objects, environment, and start/completion conditions in one sentence
How well does it work today?	Completions/hour, cycle time, failure/rework rate, human interventions
Why is the current approach insufficient?	Compare rule-based automation, controller improvements, products, and SI quotes
Which variation needs learning?	Object, lighting, layout, or instruction changes and their actual frequency
What is already available?	Robot, sensors, logs, demonstrations, CAD, usage rights, staff
What constrains the site?	Human access, outages, latency, power, countries processing data
Who operates and recovers it?	Named site, robotics/SI, ML, and cloud owners
What would stop the pilot?	Budget/time caps, minimum improvement over baseline, safety/recovery conditions

Proceed when the task and baseline are measurable and there is an improvement hypothesis, experimental capacity, and an operating owner. Hold when rights, site responsibility, or spending limits are unresolved. Choose an alternative when a product or existing automation meets the need at lower total cost. Record the choice in Build vs Buy.

Total cost and ROI — beyond the GPU quote

Model low/base/high scenarios. Mark unknown costs as ‘quote pending’, not zero. Copy the budget CSV and fill quantities, rates, and periods (USD worksheet, not a price list).

Cost group	Include
Initial fixed	Robot, gripper, sensors, installation/SI, data collection/cleaning, environment creation, safety review/validation
Experiment variable	GPU instance hours, storage/transfer/logs, models/APIs, failed runs/retries, engineering time
Recurring operations	Maintenance, parts, supervision, interventions, inference/connectivity, retraining/site revalidation
Expected benefit	Actual reductions in rework/downtime, realizable throughput value, staff time that can be reassigned

Period net benefit = period benefit − period operating cost

Payback period = initial investment / monthly net benefit (only when monthly net benefit is positive)

Period ROI = (period benefit − initial investment − period operating cost) / (initial investment + period operating cost)

Simulation can reduce experimentation costs, but does not eliminate environment creation, calibration, or physical validation. Do not count productivity and labor savings twice for the same benefit. Record sources and check dates for estimates and benefit assumptions.

Pilot approval card

Copy this block into the project document. Duration, demonstration counts, and success rates are customer-specific agreements, not universal promises.

Task / exclusions:
Current approach / alternatives / baseline measurement period:
Target metric / minimum improvement / measurement conditions:
Evaluation sample size / repeated runs / unseen environments:
Data, code, and weight usage rights:
Owners: site / robotics-SI / ML / cloud / safety:
Cost: initial / experiment / monthly operations / low-base-high:
Limits: maximum spend / end date / extension approver:
Stop: target missed / budget exceeded / safety-recovery test failed:
Evidence and approver required for the next stage:

Stage gates: establish a baseline → isolated experiment → limited site validation → expand operations. Review metrics, cost, interventions, and recovery at every gate. A technology's GA status or a generic ‘12–24 months’ forecast does not approve the next stage.

➡️ Next action: complete one card in the next customer meeting and decide proceed, hold, or an alternative. If proceeding, select just one of the three execution paths for the current bottleneck.

owner: Youngjin · updated: 2026-09 · volatility: medium

---

1. SI (System Integrator) — A provider integrating robots, equipment, controls, and software at the customer's site. Cloud service support and responsibility for cell integration/safety must be assigned separately. ↩