Dynamo for Operations
Agents for Operations.
Your team builds agents that watch lead times, raise purchase orders, sort shipment exceptions, and score suppliers. They run in your cloud, and a person signs anything above a limit you set.
Of supply chain chiefs can't say what AI returned
Gartner, 2026What a purchase order can cost to process by hand
APQC, 2026Runs on the systems operations already uses
Agents in Action
What Operations runs on Dynamo. Pick an agent, then click any node to see what it does.
When a lead time slips, an agent checks carrier rates and either reroutes the shipment or raises a PO.
What it costs today
Three jobs that still run on people and spreadsheets.
Every reorder drafted, chased, and keyed in by hand
A purchase order costs from 14 to more than 54 dollars to process, and the median wait from request to order is 55 hours. About a third of deliveries arrive late, and buyers spend the time chasing confirmations by email instead of managing the risk.
APQC, 2026; Procurify, 2025 (vendor); SourceDay, 2026 (vendor)Three carrier portals, one spreadsheet, and a customer who asks first
Exceptions land in the carrier portals and get tracked in a spreadsheet. 'Where is my order' is 30 to 50 percent of support contacts. One in ten contracted loads was rejected in January, and a retailer's window is 98 percent on time and in full with 3 percent of cost of goods on a miss.
Gorgias, 2024; FreightWaves SONAR, 2026; Red Stag Fulfillment, 2025 (vendor)A supplier moves a date and the whole plan is rebuilt in Excel
82 percent of companies had tariff-driven disruption in 2025, and 72 percent of trade professionals call tariff swings the most impactful change. The new date is keyed into the ERP by hand, and half of large manufacturers still plan in Excel beside it.
McKinsey, 2025; Thomson Reuters, 2026; Qlector, 2025 (vendor)By role
Pick a seat on the operations team.
What each person spends the month on, and the agent that takes it.
The CEO asked what operations is doing with AI. 55 percent of supply chain chiefs can't say what theirs returned, while 67 percent of their digital spend goes to it.
Gartner, 2026Freight and tariffs made every plan provisional. One in ten contracted loads was rejected in January and spot rates hit a record in June.
FreightWaves SONAR, 2026Headcount is flat and the exception queue is not. 32 percent of companies froze hiring last year and 21 percent cut.
Peerless Research Group, 2026
Lead-time watch
When a lead time slips, an agent checks carrier rates and either reroutes the shipment or raises a PO.
See it on the canvas →Build your operations agents on Dynamo
The platform underneath every agent on this page. Connect the ERP, the carriers, and the warehouse once, describe the process in plain English, test it on last month, and run it in your cloud with a person signing where money or a customer is on the line.
Each system is connected with a scoped credential, including the carrier APIs and the outside warehouse's system that reach beyond your walls, and every operations agent inherits the connections. Security reviews the platform once instead of each vendor's agent separately.
The person who owns the process describes it: which orders a slipped date touches, what the buyer's limit is, which exceptions a customer must hear about. Where a step has to be exact, it is. Changes are diffs a reviewer can read, versioned like code.
Regression suites run on your own closed cases on every change: last month's lead-time changes, purchase orders, and exceptions. You see how often the agent would have reordered, rerouted, or flagged, and how often it should have asked a person, before anything runs in production.
AWS, Azure, GCP, Kubernetes, or your own servers. Every run records what the agent read, decided, wrote, and cost. That record is what the retailer's chargeback dispute and the CEO's cost question both ask for.
Research · Whitepaper
Cost per correct answer on enterprise tasks
Accuracy alone doesn't tell a finance team which model to run. This paper measures what a correct answer costs, per task, across the models in our benchmarks.
Questions operations teams ask.
Can an agent place a purchase order or rebook a shipment on its own?
Only under the limit you set. An agent gets the ERP role of the person who started it and a scoped credential on the carrier, and nothing more. A purchase order under the buyer's limit posts; one over it waits for the buyer in Slack. A reroute under the cost threshold books; one over it waits for the logistics lead. A message to a customer is always sent by a person. The approval points are part of the agent's definition, not a setting someone can forget.
Our ERP vendor and our visibility platform are adding agents. What's different?
Those agents see one system each. The decision when a lead time slips runs across the ERP, the carrier, the outside warehouse, the retailer's portal, and Slack. Dynamo connects each once, under scoped credentials, and every operations agent inherits the connections. Security reviews the platform once instead of each vendor's agent separately.
What happens when a carrier or supplier system changes?
The connection lives in one place, not inside every agent. When a carrier changes its API or a supplier moves to a portal, the connection is updated once and every agent that reads it keeps working. Changes to an agent itself are diffs a reviewer can read, promoted through dev, staging, and production with a check at each gate.
How do we know what it cost, and what it saved?
Every run records what the agent read, what it decided, what it wrote, and what it cost. Cost is tracked per agent, per run, and per team, with budget alerts. Set it against the spot premium a reroute avoided or the chargeback a notice prevented, and you have the number the CEO asked for. Traces go to the observability stack IT already runs, over OpenTelemetry.
How do we test it before it touches live orders?
On last month. Regression suites run on every change against your own closed cases: last month's lead-time changes, purchase orders, and exceptions. You see how often the agent would have reordered, rerouted, or flagged, and how often it should have asked a person, before anything runs in production. Our cost-per-correct-answer paper ranks the models on what a right answer costs per task; each task runs on the one the evidence supports.