Agentic supply chain planning is only as good as the data your systems agree on

Board’s new pitch is a headline: “The future of planning isn’t another chatbot.” They’re right, and the line gives away more than they meant it to. Agentic supply chain planning doesn’t fail because the agent can’t reason. It fails because the agent reasons over numbers your own systems don’t agree on.

The launch is real and the direction is correct. The operator read is the part the keynote skips.

TL;DR: On June 24, 2026, Board introduced a Supply Chain Agent and a Merchandiser Agent for what it calls Agentic Continuous Planning — AI agents that continuously evaluate demand, inventory, pricing, and logistics signals and recommend actions, instead of producing a plan a human revisits on a cycle. The agent is the easy part. Underneath it sits the same fragmented ERP, WMS, and TMS data, where “on-hand,” “committed,” and “in-transit” mean different things in different systems and update on different clocks. Continuous replanning amplifies whatever that data already is. Faster planning on signals that reconcile once a night is just a faster route to the wrong commitment. Write the cross-system data contract first; turn the agent on second.

What Board actually shipped

On June 24, 2026, Board announced its Supply Chain Agent and Merchandiser Agent, extending a portfolio that already included an FP&A Agent and a Controller Agent. The Supply Chain Agent is meant to anticipate demand shifts, align decisions across service, cost, cash, and margin objectives, and improve supply resilience through continuous evaluation. The Merchandiser Agent connects demand, inventory, pricing, and assortment, flags inventory risk, and optimizes open-to-buy and stock-out decisions.

The framing is sharper than most. Board’s headline — “the future of planning isn’t another chatbot” — is an explicit shot at the idea that you can bolt a language model onto an enterprise and call it planning. In its own words, the agents work inside a “purpose-built planning environment that combines business context, forecasting, scenario planning, governed workflows, and enterprise-scale planning models.”

Read that qualifier twice. It’s not marketing filler. It’s the admission. The agent is only useful to the extent that the business context, the governed workflows, and the planning data are already trustworthy. The vendor is telling you, in the value proposition, that the work lives below the agent.

The thing that compresses is the thing that breaks

The promise of agentic continuous planning is cycle time. A planner used to rebalance a category weekly; the agent rebalances it whenever conditions move. That’s genuinely valuable — when the inputs are sound.

But speed has no opinion about whether the inputs are sound. Compress the replanning loop and you compress the loop on good data and bad data equally. This is the same pattern that shows up everywhere automation lands on an un-fixed substrate: in autonomous field-service dispatch, autonomy amplifies whatever the field data already is, and in AI manufacturing quoting, faster quotes on stale cost data just lose money quicker. Supply chain planning is the same shape at a bigger blast radius. A wrong weekly plan is a bad week. A wrong plan recomputed every hour is a bad week you commit to sixty times before anyone checks.

So the question isn’t whether the agent can replan fast. It’s whether the numbers it replans on mean the same thing across the systems it reads.

The broken layer: your systems don’t agree on a number

Here’s where it actually goes wrong, and it’s boring, which is why it gets skipped.

A retailer’s planning agent reads “on-hand inventory” to decide a replenishment. The WMS reports 1,400 units. But the WMS counts stock already allocated to open orders as on-hand, while the ERP nets that allocation out and reports 900 available. Same SKU, same instant, two numbers, and nobody wrote down which one “available” means. The agent picks one — silently — and recommends against a replenishment that was actually needed.

Now layer in time. The TMS updates in-transit positions on a nightly batch, not live. The agent kicks off a replan at 9 a.m. and confidently optimizes against where the freight was at midnight. The logic is flawless. The picture is eight hours old. Nobody gets an error, because from the agent’s point of view nothing failed — it planned exactly as instructed, on the data it was handed.

That’s the failure mode of agentic planning: not a crash, a confident wrong answer built on a definition mismatch the agent has no way to see.

Planning signal the agent reads	Where it lives	Why two systems disagree
On-hand inventory	WMS	WMS may count allocated/committed stock as on-hand; ERP nets it out — same SKU, two quantities
Committed / reserved	ERP, OMS	A sales order “commits” stock the WMS still shows as available until the next sync
In-transit position	TMS	Positions update on a batch, not live; a 9 a.m. replan runs against last night’s locations
Available-to-promise	Planning system	Each system computes ATP from a different on-hand value and lead-time assumption
”Complete” / delivered	TMS vs ERP	Proof-of-delivery scanned in the field vs invoice posted — hours to days apart

None of these are exotic. Every one is the kind of mismatch that’s tolerable when a human planner eyeballs the numbers once a week and quietly corrects for it in their head. Hand the same numbers to an agent that plans continuously and takes the human’s correction out of the loop, and the tolerance disappears.

Autonomy is rising. The prerequisite isn’t being built.

The direction of travel is not in doubt. Gartner predicted in March 2026 that 60% of supply chain disruptions will be resolved without human intervention by 2031. The agents are coming, and they’ll act.

The gap is underneath. Gartner’s Future of Supply Chains 2026 research found that 95% of supply chains must quickly react to change, but only 7% can execute decisions in real time. That 7% is not a planning-logic problem — the math to rebalance a network has existed for decades. It’s that the signals aren’t trustworthy in real time, because they live in separate systems on separate clocks.

The fresh data says the divide is widening between firms that fixed this and firms that didn’t. Blue Yonder’s Supply Chain Compass 2026, published March 31, 2026, surveyed 678 senior supply chain professionals at companies above $500 million in revenue across North America and Europe. The less-optimistic respondents were 2.5x more likely to operate in silos, 2.3x more likely to struggle with slow data sharing, and 4x more likely to run disjointed supply chains. The optimistic ones were investing in one thing first: unified data, not more agents.

This is the same lesson the broader market keeps relearning. The connectivity was never the wall — integration and data contracts were, and remain, the layer that decides whether AI works.

The working version: write the data contract, then turn it on

The fix isn’t to wait for the agents. They’re fine. The fix is to stop treating the data contract as something the vendor’s “business context” already gave you.

Reconcile the definitions before the agent reads them. Sit the ERP, WMS, TMS, and OMS owners at one table and answer the unglamorous questions: does “on-hand” include allocated stock? When does an order commit inventory? Which system is the source of truth for in-transit, and how fresh is it? Write the answers down. That document is the contract the agent assumes exists.

Make freshness explicit and make drift fail loud. Every signal the agent plans on should carry how current it is, and a batch-fed field flowing into a continuous replan should raise a flag, not quietly pass. When an upstream system renames a status or changes how it computes a quantity, the pipeline should stop, not let the agent keep planning on a silent mismatch.

Start the agent narrow. One category, one region, one decision, with a human checkpoint, before you let it run continuously across the network. Earn the autonomy on data you’ve reconciled; don’t grant it because the demo made replanning look free.

Do that, and an agentic planner is exactly what Board says it is — continuous, fast, connected to a real plan. Skip it, and you’ve automated the part that was never the bottleneck while leaving the part that always was. A planning agent that replans every hour on data that reconciles once a night isn’t faster planning. It’s a faster way to commit to the wrong number. Reconciling those systems is the work worth doing first, and it’s the conversation worth having before anything goes continuous.