The Supply Chain AI Starter Guide
Start with one operational decision. Prove what works. Know what to do next.
Use this guide to select one use case, understand the work and evidence, test a bounded change, decide whether the operation may rely on it, and choose the next investment. It works with an internal team, an incumbent provider, or SiteTrax.io.
“Real-world activity must become timely, usable, unit-level data before visibility, analytics, automation, and AI can reliably improve physical operations.”
A decision method, not a sales pitch
Built for the person responsible for the next decision — a disciplined way to investigate, test, authorize, and revisit one AI use case.
The Supply Chain AI Starter Guide
Start with one operational decision. Prove what works. Know what to do next.
What you will learn inside:
Research that frames the problem
McKinsey, IBM, and Project NANDA figures — each dated, each its own population. AI activity is not operational value.
Four stages, three permission boundaries
Find one use case → understand reality → test & operationalize → revisit. A test result does not itself authorize a live change.
A reusable Use-Case Record
A two-page decision record for each gate, plus a 15-question customer & vendor checklist for operating fit.
Ready to make the next operational decision better?
Download the guide (PDF)AI activity is not operational value
Individual productivity, financial impact, and scale are different outcomes. Use research to frame the problem, then measure your own operation.
The productivity-to-earnings gap
In McKinsey’s August 2026 global survey, 80% of respondents reported improved individual productivity from AI, while 37% attributed at least some enterprise EBIT impact to AI. These are separate self-reported measures, not a conversion rate or audited supply-chain ROI.
Expected return and scale
In IBM’s May 2025 survey of 2,000 CEOs, respondents reported that only 25% of AI initiatives delivered expected ROI over the preceding few years and only 16% had scaled enterprise-wide. This is a different population and measure from McKinsey; do not combine the percentages into a funnel.
A widely quoted number, read carefully
Project NANDA’s preliminary July 2025 GenAI report said 95% of organizations in its research were getting “zero return.” Its mixed-method findings are not a universal failure rate for all AI projects, and do not establish a physical-AI success rate.
A pilot is not value until it changes a measured decision or workflow under ordinary conditions.
Four stages. Three permission boundaries.
A useful test result does not itself authorize a change to a live operation.
One evolving Use-Case Record
At each gate, record the decision, date, owner, evidence, and exclusions. The two-page template near the end is a decision record, not a substitute for operating procedures or technical logs.
Do not let a pilot drift
Do not let a technically successful pilot drift into routine use without an explicit operating decision.
Three frameworks, three jobs
The SiteTrax.io capability framework, IANA journey, and this guide’s method reinforce one another without sharing authorship.
SiteTrax.io Four Pillars of AI
Generative AI / LLMs (Interpreter)
Interprets unstructured documents, manifests, and complex inputs into structured, usable data.
Agentic AI (Conductor)
Orchestrates multi-step workflows and dispatches tasks with bounded human oversight.
Applied / Physical AI (Optimizer)
Turns physical activity into timely, unit-level evidence through capture methods that adapt to existing workflows.
Analytical AI (Sentinel)
Monitors turn times, dwell, and patterns on trusted unit-level data.
They describe possible capabilities, not a required deployment sequence. [4]
IANA Intelligent Container Journey
[5]IANA developed its intermodal journey in partnership with SiteTrax.io and industry participants. It references the Four Pillars across origin, gate and terminal, rail linehaul, end ramp, final mile, and empty return.
This Starter Guide
A decision method for investigating, testing, authorizing, and revisiting a use case. One pillar, several pillars, conventional automation, a process change, or no AI may be the right answer.
PILLARS = WHAT CAPABILITY | JOURNEY = WHERE IN AN INTERMODAL MOVE | GUIDE = HOW TO DECIDE
From one decision to routine reliance
Four stages, three gates. A useful test result does not itself authorize a change to a live operation.
Find a decision, not an AI project
Start with a recurring uncertainty, exception, or failure that affects an operational commitment.
- Listen for the workarounds — where do people call, walk, search, re-enter, reconcile, photograph, or check another system before acting? Ask operators what happens when the normal process fails.
- Write one bounded question: “When [trigger] occurs at [site or handoff], [role] must decide [action], but lacks [evidence]. This creates [consequence]. We want to test whether [bounded change] improves [outcome].”
A YMS shows a trailer at C-17. The driver finds another trailer there, cancels the move, and calls a supervisor. The question is whether dispatch can trust the location record, not which camera to buy. (Fictional example, not a customer result.)
Permission to investigate is not permission to deploy
A sponsor authorizes learning, not an AI purchase or a new decision authority. Name the question, owner, site, shift, records, access, and review date.
The yard lead approves observation and a review of move cancellations at one site. Camera deployment and autonomous dispatch remain out of scope.
Understand reality
Map the work people actually perform. Real-time data is not useful when the organization has no clear response to it.
- Follow a real exception: trigger, normal route, systems, handoffs, wait points, exception, workaround, escalation, and completion. Compare ordinary and difficult shifts.
- Specify the evidence: source → unit identity → location → timestamp → observed event → confidence or exception → receiving system → authorized responder.
- Build a baseline you can defend, and give every alert a named owner and a response authority. Keep observation, system status, prediction, recommendation, and authorization as separate categories.
71% of ABI Research respondents ranked a lack of clearly defined SOPs among their top three blockers to proactive decisions from real-time data.
82% of KPMG-surveyed leaders cited organizational data quality as a critical barrier to GenAI goals.
The driver reports the mismatch by radio; the supervisor checks a handwritten note; the YMS update may lag a physical drop. These are questions to investigate, not assumed customer facts.
Enough understanding for a bounded test
Do not require perfect enterprise data; require honest boundaries and a credible comparison. Keep irreversible moves or customer commitments under human authority.
Measure wrong-location cancellations, then compare AI-assisted event capture with the existing process and a process/YMS correction.
Test and operationalize
Design the smallest test that changes a decision, then measure output and operating outcome separately.
- Start at an appropriate authority level: shadow mode compares without changing work; advisory mode flags or recommends while a person decides; execution requires a separate, explicit authorization.
- Choose the intervention after the workflow is known — the missing input or action determines the capability, not the AI label.
- Do not credit AI for every change. Log what else changed and use a claim the design supports.
Capture observed trailer identity and location, flag YMS conflicts, and let the supervisor verify. Track false alarms, missed mismatches, review time, and changed dispatch decisions.
Choose the intervention after the workflow is known
The missing input or action determines the capability, not the AI label.
SiteTrax.io fit — SiteTrax.io is most relevant when physical-event evidence is missing. Mobile capture, gate cameras, existing cameras, and operational workflows can create unit-level data for people and existing systems. That does not make SiteTrax.io the answer to every AI use case. [8]
Measure output and operating outcome separately
A strong identification rate is not a business result if the event arrives too late or nobody uses it. Count exceptions relative to moves, assets, or decisions exposed.
Gate 3A — What did the test establish?
A clear decision is a better outcome than a favorable but unsupported claim.
Fictional yard result: early detection helps advisory review at one site, but late reads and false alerts create supervisor work. This does not prove autonomous dispatch or network-wide ROI.
Who may rely on it, and under what conditions?
A value conclusion and authorization for routine use are separate decisions. Name exception and support owners, override, outage procedure, manual fallback, incident response, and change control. A software rollback cannot undo a physical move.
NIST’s voluntary AI RMF Playbook includes monitoring, override, recovery, and deactivation considerations. [10]
Revisit
Revisit when something material changes. A second use case is one trigger, not an automatic expansion.
- Common triggers: another site, customer, asset class, YMS or ERP transition, capture device, data source, staffing model, provider, authority level, or material performance change.
- Build, buy, or combine by capability — decide who owns the outcome and lifecycle, not which slogan sounds better.
- Reuse carefully: identity, event definitions, integrations, and controls may transfer; baselines, physical layout, authority, and economics often do not.
Authorize discrepancy flags and source evidence for tested shifts. A supervisor resolves exceptions before dispatch changes a move. No autonomous assignments.
A bounded result, not a perfect success story
Fictional composite: wrong-location records caused cancelled moves and repeated checks.
A credible partial win gives the buyer a better decision than an unsupported claim of full automation.
Take it to your operation
A two-page Use-Case Decision Record and a 15-question customer & vendor checklist. Use them with an internal team, an incumbent provider, new vendor, or SiteTrax.io.
Use-Case Record — Decision & investigation
- Use case / record ID
- Current decision, status, date
- Sponsor / owner / lead
- Problem and consequence
- Site, shift, workflow, exclusions
- Evidence, freshness, gaps
- Baseline, measure, denominator
- Investigation approval / date
- Simpler alternative
Use-Case Record — Test, reliance & change
- Test commitment / approval date
- Intervention, comparison, window
- Success criteria / stop condition
- Result, limits, added effort
- Other changes during test
- Value conclusion / evidence
- Routine use: allowed / prohibited
- Fallback, support, review date
- Change note / reason / approver
Write the decision and its evidence. Do not overwrite prior decisions.
Vendor checklist — Outcome to authority
- 1.Which decision or outcome changes?
- 2.What happens in real exceptions?
- 3.Which physical or digital evidence supports it?
- 4.How fresh and inspectable is that evidence?
- 5.What baseline and simpler alternative apply?
- 6.What changed besides the technology?
- 7.What new review or support work was added?
- 8.What may the system recommend or execute?
Vendor checklist — Operating fit
- 9.Who owns exceptions and overrides?
- 10.Which systems connect and who owns integration?
- 11.Can we access and export unit-level events?
- 12.Which capabilities should we build, buy, or combine?
- 13.What is the full operating burden?
- 14.What happens when the system is wrong or unavailable?
- 15.What permits expansion, and what makes us stop?
Why interoperability belongs here — more than seven in ten respondents to ABI Research’s mid-2025 supply-chain survey considered interoperable, open, standardized APIs important or very important in vendor selection. This is a vendor-selection preference, not proof that a particular integration works. [6]
Frequently asked questions
Make the next operational decision better
A useful first AI effort leaves a clearer problem, better evidence, a bounded result, and a defensible next move.
The standard: if the solution cannot show what happened, who owns the response, and what improved under normal conditions, the operating decision is not complete.
Where SiteTrax.io fits — when the missing input is a physical event in a yard, gate, or asset handoff, SiteTrax.io can help create usable unit-level data through capture methods that adapt to existing workflows. That evidence can support visibility, proof of activity, analytics, automation, and AI. [8]
Educational resource, not a substitute for safety, security, legal, labor, compliance, or procurement review. No IANA or ASCM endorsement is implied.
Sources & claim boundaries
Each figure is dated and describes its own population and measure. These studies cannot be averaged or treated as one AI success rate.
McKinsey, August 2026 — The state of AI in 2026. 1,719 respondents across 97 nations. 80% improved individual productivity; 37% some enterprise EBIT impact. Source
IBM Institute for Business Value, May 2025 — 2,000 CEOs. 25% of AI initiatives delivered expected ROI; 16% scaled enterprise-wide. Source
Project NANDA, July 2025 — The GenAI Divide. Preliminary findings; the 95% figure concerns the report’s GenAI research, not all AI projects. Source
SiteTrax.io authored the Four Pillars; IANA developed the Intelligent Container Journey in partnership with SiteTrax.io. Source
ABI Research, 2025 — Supply Chain Management & Logistics Survey, 490 respondents. 71% ranked unclear SOPs a top-three blocker; 7-in-10 valued open standardized APIs. Source
KPMG U.S., September 2025 — Q3 AI Quarterly Pulse, 130 U.S. leaders at $1B+ organizations. 82% cited organizational data quality as critical to GenAI goals. Source
Internal SiteTrax.io grounding — Brand Playbook v4.1 informs product fit and operator language; does not substantiate external performance claims.
KPMG International, September 2026 — Global AI Pulse Q3 2026, 2,131 leaders. 61% review AI costs at approval; 59% monitor in operation; 12% assess value vs cost enterprise-wide. Source
NIST AI RMF Playbook — voluntary Manage guidance covers monitoring, override, incident response, recovery, change management, and deactivation. Source