A sales back office manager looks back from 2030 at how her job changed over four years. And why the decision behind it is taken today.

If you run a sales back office, you know the morning: inbox open, 60 enquiries, 40 orders as PDFs, three of them illegible, and at ten the buyer from customer X calls to ask where his quote is. This article describes the same morning four years from now, seen through the eyes of a back office manager who lived through the change. The company is invented. The facts the story rests on are not.

15 October 2030, 7:04 a.m.: 1,112 enquiries answered before the first person reached the office

Sabine is 48, has been with the company for twenty years and has run the sales back office for nine of them, at a manufacturer of sealing technology. 320 employees, around 10,000 active customers, an ERP older than some of her colleagues.

Her screen says: 1,140 enquiries since last night. 1,112 of them were answered before she had her coffee. 28 are on her list: custom lengths, a customer over his credit limit, an enquiry in Romanian with the wrong article number.

Her team handles those 28. In 2026 there were 60 enquiries a day; today it is twenty times that. Almost all of last night's 1,140 came from agents. She remembers exactly when it started.

2026: An enquiry at 3:14 in the morning, and everybody found it funny

It was an email, received at 3:14 a.m. Twelve line items, exact quantities, requested date, delivery address, cleaner than any enquiry a human had ever sent. The signature read: “Automated procurement assistant. Please reply in structured form.”

Sabine printed it out and passed it around the team. People laughed. Two days later the quote went out, as always, as a PDF.

What she did not know at the time: the enquiry was one of forty the procurement agent had sent that same night. To forty suppliers. Three had answered within the hour.

Early 2027: “Your quote was the best one. It just arrived too late.”

He had been a customer for fifteen years. A corporate group whose buyer Sabine knew by first name. In early 2027 he wrote to her: “We are moving operational purchasing for C-parts to our agent system. Please make sure your prices and availability can be retrieved by machine.”

She forwarded it to IT. IT asked what that was supposed to mean. By the summer the numbers showed what had happened: the customer had ordered 70 percent less.

The buyer was honest. “Your quote was usually the best one. It just did not arrive when the decision was made. The agent sent out 80 enquiries, had three answers within the hour and placed the order. Yours came two days later.”

Sabine did the maths back then: 10,000 customers. If only three percent of them buy the way this group does, that is 300 customers asking at three in the morning. And none of them waits.

Autumn 2027: At 3:14 another agent asks, and this time the answer takes one second

Management decided that same summer to buy a system rather than build one. Selection ran into the autumn, and the only objection came from their own sales team: “Then everyone can see our prices.”

The answer that carried the day fitted into three sentences. Range and list prices open and machine-readable. Availability live from the ERP, with a date. Contract prices only for agents that identify themselves as a customer.

Six weeks after signing, it was running. The next enquiry at 3:14 a.m. had its answer at 3:14 a.m. Enquiries and orders from emails now landed straight in the ERP, and every agent that asked got an answer: range, price, availability, delivery date. Her team still wrote the quotes. What the company itself had to supply was the decision to want to be found.

2028: Agent quotes agent, binding, without anyone having read it

A year later the system wrote the quotes itself, agent to agent, with no human in between. The trade press had been calling it agent-to-agent commerce for years. For Sabine's team it simply meant quotes nobody types any more.

The system stayed inside the rules the team had set: discount corridors per customer group, minimum quantities, blocked articles, credit limits. Anything inside the rules went out in seconds and was binding. Anything outside them landed with a person.

What surprised her: the agents never stopped asking. A purchase for which a buyer used to collect three quotes now generated dozens of enquiries. Nine out of ten led nowhere. But answering cost nothing any more. And the tenth enquiry brought an order that somebody else would have won before.

And the people? They kept sending emails and PDFs and kept calling. By then some agents were calling too. The same system answered all of it automatically. The channel had stopped mattering.

2029: The PDF is still attached, but nobody reads it

Until then the document came first and the order was derived from it. In 2029 the order came first: as a data record the customer's agent had already checked against the quote. The PDF was still attached, for the legal department and the archive. Nobody read it any more.

What that did to the rest of the company, nobody had planned. Production could now see what agents were asking for before they ordered, and therefore weeks earlier than any order confirmation. And shipping could promise next-day delivery, because orders were confirmed around the clock rather than the following morning.

2030: Send only PDFs and you stop being asked

Sabine looks at her 28 cases. The Romanian customer is a human being, so she calls him. That has become her job: exceptions, relationships, the cases where a person has to talk to a person.

Two of her competitors still send quotes as PDFs, two days after the enquiry. They are still in business. But when somebody asks at three in the morning, they are not on the list. And the customers who still call get fewer every year.

“In 2026 we thought we had time,” she says. “The first customer was gone within a year.”

How much of this is already real?

The headlines about procurement agents in ChatGPT and Google are about trainers. But your customer's buyer uses the same tools. That is how the change reaches you. The story is invented. The starting position is not:

• Coupa, one of the largest procurement platforms for large enterprises, reports more than 450 customers running AI agents in production in purchasing.
• Walmart has an AI agent negotiate terms with smaller suppliers, that is, with the suppliers human buyers never had time for. Three quarters of those suppliers preferred the agent to the human buyer.
• 67 percent of B2B buyers would rather buy without a sales rep; 45 percent used AI in their last purchase, mainly to research vendors and compare solutions.
• Brands that do not appear in AI answers are already invisible to at least 20 percent of B2B buyers. The same mechanism hits order intake: a supplier the agent cannot find gets no enquiry.

How long the transition takes, nobody knows; the forecasts are far apart. When the first agent enquiry lands in your inbox is no longer a forecast: it is probably already there.

What this means for your sales back office

The company in the story invented nothing in 2027. It introduced a system that checks every incoming email and every PDF against customer data, the article master and the terms in the ERP, and involves people only for exceptions. The step to agents was the same job with a new sender.

That is exactly what Workist does today for manufacturers and distributors: translating enquiries and orders from email and PDF as well as orders taken over the phone into the context of the specific customer and into the specific ERP. Agent-ready order intake, of the kind Sabine got in 2027, is the direction we are building in. Hence this series.

Frequently asked questions about agent-to-agent commerce

What is agent-to-agent commerce?

An AI agent on the buyer's side enquires, compares, negotiates and orders from an agent on the supplier's side. No human carries out each individual step, on either side. People set the rules; the individual transactions then run by themselves.

Do we need to build a web shop for this?

No. An agent needs machine-readable answers: range, prices, availability, quote. An interface connected to your ERP can deliver those. A website that a human would have to operate is not required.

Will everyone see our prices?

Only what you release. Many manufacturers publish list prices in a catalogue anyway. Customer-specific terms are visible only to whoever identifies as that customer.

When do we need to react?

The first agent-generated enquiries are already appearing in suppliers' inboxes. You can spot them by their unusually clean structure and by arrival times when no buyer is working. If you are seeing them, the starting gun has already gone off.

What happens to my back office team?

The work shifts. The system answers routine enquiries; people take on exceptions, customer relationships and the rules under which agents may quote. In the story above, Sabine's team calls the Romanian customer themselves.

How agent-ready is your order intake?
Agent readiness checklist, five questions for a self-assessment.

Next in this series: what can AI procurement agents actually do today?

Last updated: September 2026