Writing about the problems that come up when building AI applications and running supply chains. Originally published on LinkedIn.
Browse by topic or start with the latest post. Read the full text here, or join the discussion on LinkedIn.
17 posts · Newest first
Can we accept this order? Inventory balances alone won't tell you. Committed stock, production capacity and substitution rules all matter, and the gaps often surface only when you test a real business question.
The carrier calls: tomorrow's extra capacity is 150 units, not 300. A shipping demo shows how a planner can add that temporary constraint and recalculate the plan without changing the underlying knowledge graph.
An AI can analyze the table in front of it and still miss what's happening in the business. Ontology connects the products, customers, events and rules it needs to investigate beyond that table.
Sales fell. Was demand weaker, or was the product out of stock? Two polished reports can hide a crucial difference: whether AI followed an analyst's instructions or formed and tested the explanation itself.
After we connected CoCo Desktop to a knowledge graph, it began helping us decide what to investigate, not just what to put on a dashboard. A brief note on what changed.
Add enough prompts, rules and workflow branches, and an AI project can start to look like the system it was meant to improve. The useful question is which tasks belong with software, agents or people.
A customer starts ordering less often, then changes the SKU mix. Total volume still looks fine. An agent needs operating knowledge to spot those early signals and give the planner something worth investigating.
A forecast isn't an order, and a safety stock target isn't inventory. Unless those distinctions are clear in the data model, an agent can use the right number in the wrong context.
A planning system can tell you what to replenish without telling you which products are worth keeping. Those decisions need cost, demand, capacity and commercial trade-offs on the same table.
Adjusting a forecast, allocating stock and approving an expedite are different decisions. Before automating any of them, someone has to define the financial stakes, the limits and who can approve an exception.
The plan goes into the system. Something changes, and the planner opens Excel. If exceptions and final decisions still sit outside the software, the manual process hasn't really gone away.
Sales wants to protect the account. Marketing wants to keep the SKU. Production wants fewer changeovers. Simulation won't settle the politics, but it can put a cost against each choice.
One product contributes $30 but takes three hours to make. Another contributes $20 in one hour. When production time is scarce, the higher unit margin can point you toward the wrong product.
Drop a loss-making SKU and the factory's fixed costs may simply move to the remaining products. Before discontinuing it, you need to know which costs will actually disappear.
What happens to profit if we drop a SKU, change the production mix or mark down stock today? A report of last month's results can't answer those questions on its own.
Stock gets more expensive to hold just as it gets harder to sell. Whether to discount now or wait another month depends on the product, the channel and how much shelf life is left.
A product that barely sells needs a different fix from one ruined by discounting or high costs. Calling them all 'unprofitable' hides the cause, and knowing the cause still doesn't mean anyone will act.
No posts found. Try a different search or topic.
Tell us where you're getting stuck. We'll help work out what needs to change and where to start.
Talk to us →