Interview with: Lalit Das, Founder & CEO, Agentra AI
Why the next phase of supply chain technology will be about AI that acts, not just advises?
Supply chains no longer have a visibility problem. They have a decision-latency problem. It takes about 8.7 hours to detect a disruption and more than 40 more to understand its impact, all before anyone acts. Meanwhile, disruption costs the global economy an estimated $184 billion a year.
Enterprises already own excellent systems of record. What they lack is a system of action. Copilots don't fix that, they just give faster answers about a broken process.
The next phase is agents that act. They spot a missed pickup, cost out each re-route option, and stage the new booking in the ERP for one-click approval. The value was never in the insight. It is in closing the gap between knowing and doing.
Can Agentic AI Make Supply Chains Truly Autonomous?
Not yet, and anyone claiming full autonomy today is overselling. Autonomy has to be earned one level at a time: observe, recommend, act with approval, then act autonomously.
Here is what that looks like in practice. An agent watching carrier feeds flags a missed pickup. It then proposes a re-route with the cost difference. Next, it prepares the booking for one-click approval. Finally, once its accuracy is proven, it clears freight invoices within tolerance on its own. Crucially, the agent doesn't promote itself; the customer grants full autonomy per agent and per threshold, and our agents run in production at the "act with approval" level today.
Reasoning through trade-offs is the easy part. The hard parts are clean data, a replayable audit trail, and clear accountability when something goes wrong — and the failure modes are specific: bad master data, prompt injection through supplier emails, and bullwhip effects when many agents react to the same signal. That's why the approach that holds up is bounded autonomy, not full autonomy: low-stakes, reversible actions run on their own, medium-stakes actions need approval, and strategic calls stay with people.
Why supply chain operations could become one of the biggest use cases for Agentic AI?
Three reasons. The first is volume. Supply chains make thousands of repetitive but judgement-heavy decisions every day, from exception triage to invoice matching high decision density, much of it buried in messy email, PDF and EDI workflows.
The second is fragmentation. A single order touches an ERP, a transport system, a warehouse system, carrier portals and dozens of partners, and no two of them agree. Agents that hold context across all of them replace the person reconciling screens.
The third is that the cost of delay is measurable. Organisations lose around 2.4 per cent of annual revenue to disruption, so every hour saved has a number attached. Supply chains also come with clear KPIs already in place — OTIF, expedite cost — and built-in constraints like contracts, MOQs and lead times that give agents clean guardrails to operate within.
Scale, complexity and clear ROI rarely come together like this. That combination is why our Fortune 500 deployments show 35 to 50 per cent less time spent on exceptions and 6 per cent lower freight costs.
How agentic systems are changing the way enterprises approach planning, procurement and execution?
Today, each function works from its own version of the truth, and every handoff is an email and a re-key. When a demand plan changes, someone has to trace the knock-on effects by hand: purchase orders, inventory, transport, and delivery dates already promised to customers. Some of those effects are never traced at all.
Agentic systems put everything on one shared record — an event-driven layer that feeds a shared supply chain graph, sitting above the ERP, which stays the system of record. When demand moves, orchestrator agents spread the impact automatically. Procurement re-decides quantities and suppliers, transport rebalances loads, and only what falls outside policy reaches a human. The heavy trade-off math itself runs through dedicated solvers, forecasting and simulation tools that the orchestrator calls on, rather than being guessed at.
Planners stop reconciling spreadsheets and spend their time on the calls that need judgement. And every resolved exception becomes part of the system's playbook, so hard-won knowledge stays when people leave.
Who Controls the Autonomous Supply Chain?
The answer is simple: people remain in control. Humans set the objectives, boundaries and policies, and agents operate within them. Accountability always stays with the organisation, never with the AI.
In practice, that requires a few things. First, clear decision boundaries. Businesses should define exactly what agents can decide alone — for example spending limits, approved suppliers or customer priorities — and anything beyond that is escalated to a person.
Second, transparency. Every action an agent takes should be recorded along with the reasoning behind it, so teams can see not just what was decided but why. That is essential for audits, compliance and building trust — every action logged for audit and replay.
Third, meaningful human oversight. Rather than approving every small decision, people should focus on high-impact, irreversible or unusual situations where their judgement matters most. They should also be able to override or pause agents at any time, backed by blast-radius caps and a kill switch for when something needs to stop fast.
Fourth, clear ownership. Just as every team has a manager, every agent's area of responsibility should have a named owner who reviews its performance — tracking decision quality itself, like override and reversal rates, not just how much got automated.
Governance shouldn't be an afterthought — it's what allows autonomy to scale safely. At Agentra, we build this in through audit trails, approval workflows, configurable autonomy levels and role-based permissions. The goal isn't to remove people from the supply chain. It's to put them where their judgement makes the biggest difference.
