According to Gartner, enterprise spend on supply chain management software featuring agentic artificial intelligence (AI) capabilities is expected to reach around $53 billion by 2030, up from less than $2 billion in 2025, with adoption among enterprises using supply chain management (SCM) software projected to rise from around 5% in 2025 to roughly 60% by 2030. Boston Consulting Group's complementary tracking suggests that agentic systems accounted for around 17% of total AI value in 2025 and are projected to reach about 29% by 2028.
Forecasts of that kind measure investment intent. They do not yet measure operational outcomes. And across all-mile deployments, the pattern visible to anyone close to this category is that AI is being bought faster than the operating models around it are being redesigned to absorb it. For any Chief Supply Chain Officer (CSCO) running a retail, Consumer Packaged Goods (CPG), logistics and manufacturing supply chain, the consequence is predictable.
A new agentic capability gets implemented — whichever leg of the all-mile stack it covers, whether routing, predictive Estimated Time of Arrival (ETA), control-tower visibility, exception management or post-purchase orchestration. Pilot Key Performance Indicators (KPIs) look great. Then twelve months in, the savings curve flattens, exception backlogs grow, and the operations team is back to making the same manual calls it made before — only now with a more expensive dashboard. The pattern is independent of which capability was adopted.
The problem is not the algorithm. The problem is what sits around it.
The real bottleneck is not cost. It is coverage.
Most all-mile AI conversations open with cost — and the last-mile leg is usually where the conversation lands. Over the past five years, industry analyses consistently point to last-mile costs taking a materially larger share of total logistics spend, particularly in B2C supply chains where e-commerce volumes have surged. The number matters less than what it hides: cost is downstream of coverage. The first mile, middle mile and last mile only optimize when AI sees and orchestrates across all of them. You cannot optimize what your AI does not see.
In a typical CSCO’s stack, three or four AI tools live in three or four functional silos — one for routing, one for forecasting, one for control-tower visibility, one for exception management. Each is intelligent in isolation. Together, they do not talk. Every handover between them is still a human ticket, an email, a Teams ping. That is where the Return on Investment (ROI) leaks.
The shift that matters is not “more AI.” It is moving from point AI tools to agentic orchestration mapped to the value chain.
HITL is not a percentage. It is an operating layer.
Most all-mile conversations frame Human-in-the-Loop (HITL) as a ratio — what percentage of AI decisions need human review. That is a trap. As the agents get better, the ratio shrinks. If your value proposition is sized by it, you are selling something that is structurally getting smaller every quarter.
The HITL question that actually matters is not what percentage of decisions AI handles. That number will keep climbing, and it should. The real question is what operating layer wraps the agents — exception operations, governance, integration, change management, customer service, carrier and Service Level Agreement (SLA) management, master data, compliance — and whether your partner can run it as a managed service. That is where most of the work sits, and that is what does not get smaller as autonomy increases. In several scopes, it actually grows: more agents mean more governance, not less.
In production, HITL at scale looks like this:
- Agents execute the routine decisions across the value chain. This spans order intake and orchestration, transportation planning and route optimization across hundreds of operational constraints, multi-carrier sourcing and rate selection, hub and warehouse handling, real-time dispatch and execution control, in-flight tracking and dynamic re-planning, Estimated Time of Arrival (ETA) prediction, driver guidance and on-route navigation, post-purchase customer experience and delivery orchestration, freight settlement and reconciliation, and the analytics and governance feedback loops that close the loop end-to-end across the all-mile value chain.
- Human specialists own a layer around the agents — supervising exceptions, governing model behavior, managing carriers and SLAs, handling decisions that touch revenue, compliance or customer trust, and re-engineering the operating model as the agents evolve.
- A control layer ties them together — knowing which agent owns what, which decisions escalate, how human specialists are inserted at the right points, and how the agents learn from human overrides.
That is orchestration, not integration. And it is a bigger job than HITL is usually scoped for, not a smaller one.
None of this is an argument against autonomy. Autonomy will keep increasing, and the agents themselves will keep getting better at the routine decisions. The point is the opposite: increasing autonomy makes the operating layer around the agents more important, not less. The more decisions agents take independently, the more weight rests on the governance, the contextualization, the exception design, the change management and the supervisory architecture that surround them. The two trajectories — agents becoming more autonomous, and the operating layer becoming more central — are not in tension. They are the same trajectory, read from two different points in the value chain.
Why this is an alliance problem, not a software problem
No single vendor builds all of this. The AI specialists — routing, predictive ETA, control-tower platforms — are best-in-class at what they do, and deliberately narrow. The managed services layer brings the operating model, the value-chain coverage, and the human expertise. Neither side scales without the other.
There is also a layer of contextualization that no horizontal platform provides out of the box. Every deployment has to be tuned to the specific industry — the operational rhythms of grocery are not the operational rhythms of pharma, and neither matches the constraints of industrial distribution or omnichannel retail. Within an industry, every client brings its own network topology, carrier mix, customer promises, service-level economics and exception patterns. Translating a horizontally capable AI platform into a client-specific operating model is where alliances earn their value.
This is where the Sales & Fulfillment (S&F) practice at Infosys BPM is investing. Building on Infosys Topaz as the underlying AI foundation, we are mapping agentic AI capabilities — both home-grown and partner-led — directly onto the S&F value chain, so every node from order-to-cash through all-mile fulfillment has a clear answer to three questions: which agent runs this, what is the human supervision model, and how does it connect to adjacent agents upstream and downstream.
That mapping is the deliverable. Not a slide deck. An operating architecture that a CSCO can actually deploy against.
What changes for CSCOs
Three practical shifts for anyone re-evaluating their all-mile AI strategy:
- Shift the procurement lens from AI tools to agentic coverage of the value chain. A localized optimization in a single function tends to deliver less compounding value than a smaller, broader gain measured across the full chain. A useful test for any vendor conversation is whether they can place their agent on a value chain map and show how it connects to what sits upstream and downstream of it.
- Rewrite the operating model before the procurement specification. AI deployments fail in operations, not in Information Technology (IT). The operations organization chart, the exception thresholds, the supervisor span-of-control — these have to be redesigned around the agents before the platform goes in.
- Treat alliances as architecture, not procurement. The right question is not “who has the best logistics platform.” It is “who can stitch best-in-class agents into a coherent operating model, contextualize it to my industry and network, and stand behind the outcome.” That is a managed-services-led alliance, not a software Request for Proposal (RFP).
The honest close
A lot of what is being sold as “AI transformation” in all-mile logistics right now is still point automation wearing a new label. The CSCOs who outrun the rest in the next 18 months will not be the ones with the most AI logos on their architecture slide. They will be the ones who redesigned the operating model around an agentic value chain — and chose partners who could orchestrate and contextualize it end-to-end.
If you are mid-RFP on an all-mile AI platform and the conversation has not moved past tools, it is worth a pause. Happy to compare notes on how we are mapping this for clients in retail, CPG, logistics and manufacturing — and where the failure modes usually hide.
