Danfoss put AI agents on its order inbox. They make 80% of decisions

Danfoss gave the email orders its EDI links never caught to AI agents. Its digital chief says they make over 80% of the decisions. Cost: not published.

Danfoss, the Danish maker of drives, valves and refrigeration controls with EUR 9.4 billion in 2025 sales, has handed the customer orders that arrive by email to AI agents. Its Head of Digital Business, Carlos García, says agents now make more than 80% of the decisions on those orders, saving about five minutes per order by the company's own estimate. The system comes from Go Autonomous, a Copenhagen start-up, and was live in 53 markets by May 2026. Danfoss has not published what it cost.

This week the argument in tech circles is about who will sell agents to companies, and whether a general assistant from a big platform can become the one every employee uses. The Danfoss case points the other way. The agent that is actually running inside a large manufacturer does one dull job, on one inbox, against one ERP, and the executive in charge says the general-purpose tools Danfoss tried first could not do it.

What Danfoss actually automated

Danfoss already had system-to-system ordering with its largest partners. In its 2025 results it says 75% of all customer orders are received and handled digitally. The problem was the rest: distributors and smaller customers who email an order, often as a PDF, a spreadsheet, a photo or free text in the message body. García told Google Cloud, for its customer story, that Danfoss's own web ordering tool for these customers often needed as much manual input as email.

Before the agents, a customer service person read each email, checked prices and part numbers, and keyed the order into the ERP, moving between as many as five systems. García puts the turnaround for those manual orders at around 42 hours.

The Go Autonomous agent reads the email and its attachments, extracts the order lines, validates them against Danfoss's ERP, and either creates the order or hands it to a person with the open questions marked. Staff now work in one screen instead of five. Danfoss and Go Autonomous built the ERP and CRM connectors together.

The agent running inside a large manufacturer does one dull job, on one inbox, against one ERP.

What changed, by the company's own numbers

The figures come from Google Cloud's customer story, in which García is quoted, from Go Autonomous's releases, and from Danfoss's own results release.

Decisions: above 80% of the decisions on these orders are made by an AI agent, García says. Google's summary words this as up to 80% of transactional decision-making and communication.

Time per order: the Danfoss team estimates about five minutes saved per order, and says the number of orders processed fully by hand is falling steadily. Google's summary lists processing time down by 50%.

Turnaround: García describes cutting the 42-hour manual turnaround dramatically and aiming at an acknowledgement within a minute for a significant part of incoming orders. The vendor's own case study headline says under one minute. Note the tense in García's quote, which talks about where Danfoss is heading.

Rollout: Go Autonomous announced in January 2025 that Danfoss had gone live in Spain, France and Italy. The Google story quotes García saying the system is live in over 20 countries, and the vendor says 26 countries went live in one day. By 25 May 2026, the Danish business daily Børsen reported, Danfoss had rolled the tool out in 53 markets and cut time on these B2B customer interactions by 90%. Go Autonomous's release about that article puts it at roughly nine minutes down to one.

Staff: García says onboarding new hires is much faster, the work is more enjoyable and churn has fallen, while adding that this was not something Danfoss tracked at the start.

What did it cost?

Danfoss has not said. Neither the Google story nor the vendor's pages give a price, a payback period, order volumes or error rates. The Google story says the system is expected to save Danfoss millions of euros by shortening the cash conversion cycle, which is a forecast, not a result. A $15 million annual saving with a six-month payback circulates on blogs and AI-statistics sites; we found no Danfoss, Google or Go Autonomous document that states it, so it is not in this piece.

The time figures also do not line up exactly. Five minutes saved per order is a Danfoss team estimate in the Google story. Nine minutes down to one is how the vendor summarised a newspaper feature. They may measure different things over different periods. Treat both as directional.

What still needs a person is stated plainly: complex transactions such as pricing exceptions. Danfoss says it wants to connect the platform to its data warehouse so agents can handle some of those later.

Does this apply to a 200-person manufacturer?

Yes, if a meaningful share of your orders still arrive by email and get keyed in by hand. The useful lesson is the scope: Danfoss automated only the orders its existing integrations could not reach, used a tool built for B2B order entry rather than a general chatbot, and kept people on exceptions. Count your email orders before buying anything.

The arithmetic is easy to run on your own numbers. At Danfoss's estimate of five minutes saved per order, a business keying 100 email orders a day would free a little over eight hours of staff time a day, roughly one full-time role. That is our calculation, not Danfoss's, and it assumes your orders look like theirs. At 20 orders a day it is under two hours, and a better order form or a customer portal may be the cheaper fix.

Two things transfer whatever your size. First, the agent can only check what your ERP already holds. If price lists, customer-specific terms and part numbers are out of date in the system, an agent will either stop and ask or confirm the wrong order. Second, measure before you start: orders per day by channel, minutes per order, turnaround from email to confirmation, and how many orders are corrected later. Danfoss could quote a 42-hour starting point because someone had measured it.

What it does not mean

It does not mean agents have replaced Danfoss's customer service team. The company says staff were freed for other work and still handle exceptions, and it has published no headcount change. It does not mean a general-purpose assistant can do this out of the box; García says the general platforms Danfoss looked at could not scale to its needs. And it does not mean the hard part was the AI. Danfoss had already moved three-quarters of its orders onto digital channels and cleaned its data before the agents took the remainder.