MIT’s “GenAI Divide” And What It Means For Supply Chain ROI
$35–40B invested, only 5% at scale. How supply chain leaders can stop the waste and start getting results.
The latest MIT NANDA study is blunt: US companies poured an estimated $35–40 billion into generative AI and 95% have little or nothing to show for it. Only 5% have AI in production at scale. The authors say the core blockers are not infrastructure or talent. The blockers are systems that do not retain context, do not adapt, and do not learn over time in real workflows. Sound familiar?
For supply chain, the message is clear. If your pilots look great in demos but stall in operations, you are not alone. Here is a practical read of the findings for CSCOs, CPOs, and logistics leaders, plus a plan to cross the divide.
What the research actually found
Scale is rare. Only 5% of organizations moved AI into production at scale. Most efforts remain pilots.
Custom tools underperform. Just 5% of custom enterprise tools reach production. Leaders reported “dozens of demos” and only a few that were genuinely useful.
Generic tools win on usage. Employees prefer flexible, familiar tools like ChatGPT, even when enterprise tools use the same underlying models. Adoption beats rigidity.
Impact is uneven by sector. Technology and Media show material effects. Most other sectors, including advanced industries and consumer and retail, report little fundamental change so far.
Workforce impact is narrow. Reductions hit standardized, often outsourced work such as admin processing and basic support. Estimates range from 5 to 20 percent in those non-core areas.
Budgets are mis-aimed. About half of AI spend goes to front-of-house use cases like marketing. The study urges shifting investment to processes that move measurable business outcomes.
Translation for supply chain: pilots that live in slides and sandboxes will not move OTIF, expedite spend, or working capital. Your teams will ignore rigid tools. And the value sits in repeatable, auditable workflows, not one-off demos.
Why supply chain pilots stall
The study’s root causes map directly to common ops failures:
No memory in the flow. An assistant that forgets supplier terms, expedite caps, or lane constraints cannot run a real exception.
Low adaptability. If a model cannot learn from last week’s port delay or quality hold, it repeats mistakes.
Rigid UX. Teams prefer fast iteration. Generic chat tools let planners “steer” the output. Many bespoke apps do not.
Science projects. Nice wrappers around models, no hard integration to your policy, data freshness, or approvals.
Where to point AI spend in supply chain
MIT’s guidance is to fund work that drives business results. For supply chain, that means:
Exception triage that protects revenue. Rank at-risk orders by value, propose three recoveries within policy, draft supplier or carrier outreach for human approval.
Safety stock tune-ups on top families. Recompute with real variability and service targets. Show working capital and OTIF impact before you change policy.
Route and mode decisions on expensive lanes. Offer cost-holding options with service preserved and simple carbon estimates.
Supplier risk briefs. Blend your scorecards with public signals. Trigger pre-agreed mitigations.
S&OP summaries people read. One page, assumptions, data as-of dates, three decisions with P&L impact.
Data quality agents. Fix the inputs that break everything else, like lead times, ASNs, and item masters.
These use cases are measurable. They fit the study’s call for outcomes over demos.
How to cross the GenAI divide in supply chain
The research points to a simple shift: stop buying AI like SaaS, start managing it like process outsourcing with accountability to business metrics.
Adopt a “buy then tailor” approach.
Start with flexible assistants your teams will actually use. Wrap them with your policies, data schemas, and templates. Do not jump to heavy custom builds unless a proven gap remains.
Drive from the front line.
Let planners, buyers, and transport leads co-design prompts, guardrails, and outputs. Adoption follows usefulness.
Enforce production discipline.
Create weekly portfolio gates. Kill ideas that do not move a KPI. Promote only what passes a small pilot with clear lifts.
Insist on auditability.
Every output should show assumptions, data freshness, and confidence. Keep AI in draft. Humans sign off.
A 12-week plan that aligns to the study’s advice
Weeks 1–2: Stand up a live exception view. Replace two status meetings. Start daily triage with human approvals.
Weeks 3–6: Tune safety stock for one family across two DCs. Publish the working capital and OTIF change.
Weeks 7–9: Run routing experiments on three costly lanes. Hold service, cut cost, document carbon impact.
Weeks 10–12: Ship supplier risk briefs for top vendors with triggers. Publish an “AI P&L” that shows service, cost, and cash wins.
If a step does not move a KPI, stop and pivot. That is the discipline the study says most companies lack.
Simple ROI model you can reuse
ROI = (hours saved × blended rate)
+ (expedite reduction)
+ (working capital benefit)
+ (revenue protected)
- (licenses + setup + change costs)
Use the same formula for every use case. Report it monthly. Tie wins to OTIF, cost to serve, and DIO.
Guardrails that keep value ahead of risk
Draft only. No autonomous PO changes or bookings.
Mask PII and sensitive pricing unless strictly required.
Show assumptions and data as-of dates on every output.
Log prompts, inputs, outputs, reviewers, and decisions for audit.
Bottom line
MIT’s NANDA work says the quiet part out loud. The ROI gap is real because most deployments ignore memory, adaptability, and real operations. Supply chain leaders can do better with a business-first portfolio, front-line adoption, and auditable workflows that protect revenue, cut cost, and free cash.
Quick takeaways
Focus spend on exception handling, inventory, routing, and supplier risk where outcomes are measurable.
Start with flexible tools, then tailor to your policies and data.
Install weekly gates and a simple ROI model. Kill fast, scale what works.
What did your last AI pilot actually change in service, cost, or cash? Where will you point the next dollar?
Share your thoughts below and join the conversation in our global supply chain community, Chain.NET. Joining is free and only takes a few minutes: https://mygs.cc/chain