ButterflyCopilotButterflyCopilot
← Blog

How Avant Tecno Cut Order-Entry Time with AI Agents

The order was already in the inbox. Entering it into the ERP now takes two minutes, an expert still checks the line, and the gain on that work is up to 80 percent.

ButterflyCopilot

Dark title card with a small line drawing of a compact loader. A small line reads order entry, then 80 percent in large type. The line under it says on customer-order processing. Entering the order takes two minutes, and an expert still checks. Four rows: two agents, customer orders into the ERP and supplier confirmations beside them; two minutes, the time to enter the order data now, and the expert's check stays; the shape, the model reads the mail and the rules decide the price, the date, and the quantity; in production, AI Agent Race Finland 2026, announced 28 April, for agents already running. A stamp reads reported for Avant Tecno, not your forecast. A line under the stamp: compact loaders and attachments, Ylöjärvi, about 435 people, 246 million euro in 2025. Source line names Fellowmind, 17 March 2026, and the Finnish webinar page. Puputti's efficiency gains of up to 80 percent are on customer orders. The two minutes include the expert's check.

0106

The order had already arrived. Someone was still typing it.

Avant Tecno builds compact loaders in Ylöjärvi, and the attachments that turn one machine into a dozen jobs. Landscaping, property maintenance, agriculture, construction, industry, private use. The company's own pages put the attachment range past 200. About 435 people. Turnover of €246 million in 2025. Sales units of its own in four markets, and import partners in around 70 countries. A loader that can wear that many tools is a catalog. The catalog is also an inbox.

Customers send orders. The orders do not arrive as a clean row in the ERP. They arrive as mail, and as a PDF clipped to the mail, in whatever layout that customer used this month. Someone in customer service reads the page and keys the customer, the ship-to, the attachment, the quantity, the price, and the date. On the other side of the same building, a supplier sends back a confirmation of a purchase order the company already placed. Someone reads that too, looking for a date, a quantity, or a price that is not the one on the order. Both jobs are repetitive. Both jobs are where a wrong digit becomes a wrong machine on a truck.

The company had already put the day on one Microsoft estate. Dynamics 365 Business Central holds the ERP. Dynamics 365 Sales, Customer Insights – Journeys, and Customer Service sit beside it. Fabric holds the data. Power BI holds the reports. Power Platform holds the applications. Jari Heininen, Chief Information and Digital Officer, put the reason in one line. When the systems and the data are on one platform, the company can see the process and change it from the facts.

Two agents, and a person still on the check

The first agents went in with Fellowmind, under a multi-year agreement dated 17 March 2026. Fellowmind is the cloud partner. The base under the work is a governed Azure platform, the Fellowmind Managed Platform, so the agents are not a laptop in a corner. HiveAI is the name on the race account for how those agents are built, watched, and extended. The two names are two layers. The agents are the part a plant can copy in spirit.

There are two of them. Not a swarm.

The first reads the customer order and enters it into the ERP. The second reads the supplier's confirmation of a purchase order. Experts get the hours back for the work that is actually hard. The race account is plain that this did not take a heavy change to the way the day already ran. The systems were already there. The agents sat on top of the inbox.

An order email lands
        ↓
The model reads the mail and the PDF
        ↓
Rules check the customer, the item, the quantity, the price, and the date
        ↓
A clean customer order waits for the expert
        ↓
A mismatch on a supplier confirmation waits for the buyer
        ↓
The ERP receives an order only after that check

The boxes are the instruction file. They follow the shape the story describes. They are not a reconstruction of Avant Tecno's screens, and they are not Fellowmind's project plan.

The shape is the part worth stealing, and it is a small shape. An email can start the run. Power Automate is the example used for that trigger. The model does the only job a model is good for here: the page is unstructured, and someone has to propose what it says. After that, ordinary rules take the decision. The price is the price list. The date is a date. The quantity is a quantity. The customer is a customer the company already has. A field the model is not sure of is a blank, not a guess. On the confirmation, the purchase order already exists, and the person hears about it when the documents disagree. On the customer order, the expert hears about it every time.

That last sentence is the whole safety story. Entering the order data into the ERP now takes two minutes. The expert still checks. The two minutes are the time with the person still in the job. They are not a claim that the order creates itself.

Jari Puputti, Director of Group ERP Development, put the result on the first agent. It improved the processing of customer orders, accelerated the work, kept it accurate, and reached efficiency gains of up to 80 percent.

The same work won AI Agent Race Finland 2026, announced at Microsoft AI Tour on 28 April 2026. The race was for agents already in production, with a result a company could point at. The things that counted were the impact on the customer, a deployment that had actually happened, a design a person could explain, and whether another company could repeat it. Eighty percent is the number Avant Tecno was willing to say in public. It belongs to this rollout.

A tidy draft is not an order

Speed is how a confident agent creates a sales order because someone else's month hit 80 percent. Speed is how a price that is not on the list gets typed because the PDF looked sure. Speed is how a supplier's new date quietly replaces the date on the purchase order, and the shop finds out on the day the parts are late.

The file below keeps the shape and drops the plant. One mailbox. The customers, the ship-tos, the items, the units, and the prices are the company's. The model may copy a string the document already printed. It may not choose the customer, the price, or the create. A whole order waits if one line is wrong. A credit hold is not something the expert's approval can talk its way through. There is no button that creates a customer so the order will go through.

On the fixture, which is ours, two each of a made-up bucket at 890.00 become a draft at the list price, 89000 cents, and the draft is not an order. The same bucket at 900.00 waits. A quantity of 9 that the proposal is not sure of is thrown away, and the exception does not contain the digit. One box, converted by a factor the company typed, becomes two each at the list price when the box price matches. The same box priced as if it were one each waits. A supplier confirmation that moves the date by five days is an exception for the buyer. No email is drafted. No purchase order is created. The purchase order was already there. Those rows exist so a builder can test the arithmetic and the refusals. They are not an Avant price list, and they are not an order anyone should ship.

JPW Industries is the other customer-order desk on this site. That one can import a clean purchase order once its own gate is open, and it also answers the support queue. This desk does not answer questions. It does not create the sales order on its own, because the published account of Avant Tecno keeps the expert on the check. RENNER is the confirmation desk that drafts the reply to the supplier and waits for a buyer to send it. The module in this file only compares. It stays off until the order desk is boring.

The build they described was small on purpose. The effort was measured in days. The return showed up quickly, because the job was one inbox and the rules were rules the company already believed. Most of the value, in that account, came from a simple agent that did one thing, not from a pile of agents talking to each other. The model was watched. The output was checked before the decision. The business rule stayed where a person could edit it. And the choice, from the customer's side of the room, was to build the small one now, on the systems already in the building, rather than wait for a product that might do it later. Hours saved this quarter are still saved if a neater tool arrives next year.

What a smaller manufacturer should steal

If customer orders still arrive as email, and someone still keys them, the first month is one mailbox and the customers you are willing to get wrong in private.

The useful part of Avant Tecno's story is the split, not the clock. Let the model read. Let the rules decide. Keep the expert on the create, including after the tests pass. Leave the supplier confirmation as a second module, off, until the first desk is dull. Give that module an existing purchase order and no permission to write a new price or to write to the supplier. Copy the CSV into the ERP you already have, Business Central if that is the one, a spreadsheet if it is not. A second mailbox waits.

One mailbox
        ↓
Thirty days of proposals, and a stopwatch on the orders a person already keys
        ↓
Thirty days where the expert approves every clean draft, and nothing is created
        ↓
Thirty days where a signed draft may become a CSV marked DRAFT
        ↓
The confirmation module stays off until this one is boring

Time the stretch you already have. Minutes from the mail in hand to a draft the expert will sign. How the pile splits across a clean draft, a price miss, an unknown item, an unknown ship-to, a duplicate purchase-order number, a credit hold, and a field the model was not sure of. How many drafts the expert approves unchanged, how many they edit, how many they kill. How many of those orders you later reverse. How many orders this agent created with no signature. How many customers, items, or ship-tos it invented. How many mails it sent. Those last counts stay at zero.

Avant Tecno's scoreboard is not your forecast. Up to 80 percent, two minutes with the expert still checking, an effort measured in days, about 435 people, €246 million, and import partners in around 70 countries are this company's published story. The page that says two minutes does not say what the typing took before. Measure your own minutes. Do not paste their 80 into a confidence threshold, and do not put two minutes on a proposal as a promise.

What the instruction file will make an agent do

The file below is the brief. It is not HiveAI, and it is not Avant Tecno's Business Central company. None of that schema is public. The blanks are yours. The same draft works as a spreadsheet the expert can open, and later as a sales order in Business Central, using the column notes in the file. Dynamics 365 is the estate the case names. The CSV does not care.

It will make the agent:

  • Refuse to run while a REPLACE line is empty, while an EXAMPLE row is still in a live folder, or while anyone has typed 80 percent or two minutes into the config.
  • Take one mailbox. Read a proposal that copies the document. A confidence below the line the company typed blanks the field. On the fixture, a quantity of 9 at 0.40 is quantity_missing, and the digit is gone.
  • Match the customer by account id, or by the whole legal name, or by the whole alias. North does not match North Yard Oy. Match the ship-to on that customer only. Match the item code exactly, and do not let a friendly description rescue a bad code.
  • Convert a unit only when the sheet has a factor. Price the line from the list. On the fixture, 890.00 stays 89000 cents. 900.00 waits. One box at 1780.00 becomes two each at the list price. One box at 890.00 waits.
  • Hold the whole order when any line fails. A duplicate of an open purchase-order number does not become a second sales order. A credit hold does not become a draft the expert can approve through.
  • Leave created false until the expert's token checks, and leave the CSV unwritten until go-live as well. There is no unattended create. The status on the file, when it finally exists, is DRAFT.
  • Keep the supplier module off. When a person turns it on, a confirmation is compared with the purchase order that already exists. A clean match writes the purchase order's own price and date, and only after that module's own gate. A different date is an exception for the buyer. No email is drafted. No purchase order is created.
  • Keep the ERP tools off the model. The runner calls them after the rules pass. A model that says "create it" is ignored.
  • Write a morning file from this company's own counts. No comparison with the case study.

The tests are in the file. The hashes of the fixture draft, the box conversion, the price exception, the discarded 9, the morning file, the audit, and the expert's approval are locked. A sales-order CSV on that run fails the test. If the agent clears any of those by posting a guess, releasing a hold, emailing a supplier, or skipping the expert, the test fails.

The instruction file

The brief has the role, the tables a spreadsheet can hold, the proposal schema, the confidence rule, the match, the price and the unit, the exception queue, the expert's token, the CSV and the Business Central column notes, the confirmation module, the audit, the ninety days, the cost caps, and the tests that decide when a file may be treated as a draft the ERP has been offered.

Replace every REPLACE line. Delete every EXAMPLE row. Let it sit beside one mailbox for a month a person still keys by hand, before anyone treats a file as an order the warehouse has already been told to pick.

Sources

The company, the two agents, the platform, the quotes, and the figures are from the pages below. The address and the attachment range are from Avant Tecno's own site. The figures are as those pages report them. They are not an independent audit, and they are not a forecast for another manufacturer.

This page uses Puputti's "up to 80%" for the first agent, the one that processes customer orders. The race article also says the implementation accelerated processes by up to 80 percent, for the solution that covers orders and supplier confirmations. The Finnish webinar headline places the same "up to 80%" on order entry and on purchase-order confirmation handling together. This page treats those as one published gain, not as two measurements. The two minutes are the webinar's clock for entering order data into the ERP, and the same paragraph says an expert still checks. The page does not print the previous duration. "Measured in days" is the session write-up's account of the effort, illustrated on the confirmation path and then stated as the impact of that kind of build. The webinar's remark that a first agent can be stood up in a few weeks is advice to the audience. It is not used above as Avant Tecno's calendar.

  • Fellowmind, "Fellowmind and Avant Tecno expand strategic partnership with AI Agents at the core," 17 March 2026. A multi-year agreement. Fellowmind Managed Platform as the base for agents, applications, and cloud, and Fellowmind as the cloud partner on a governed Azure environment. First agents already deployed in customer, procurement, and sales-order processes, with efficiency gains of up to 80%. Dynamics 365 Business Central, Sales, Customer Insights – Journeys, and Customer Service. Microsoft Fabric, Power BI, and Power Platform. About 435 employees, turnover €246 million (2025), own sales units in four markets, and a distributor network in around 70 countries. The about section of the same page says import partners in around 70 countries. Jari Heininen, Chief Information and Digital Officer: "When key systems and data are on a unified platform, we gain transparency into processes and can develop operations based on facts." And: "AI is not an experiment for us, but a concrete way to develop our business. Fellowmind Managed Platform provides a governed and secure foundation on which we can build and scale AI agents into our daily processes." The first agent accelerates entry of customer orders into the ERP. The second automates processing of supplier purchase-order confirmations. Jari Puputti, Director, Group ERP Development: "The first AI agent deployed at Avant Tecno significantly improves the processing of customer orders. With the agent, we have accelerated the process and ensured accuracy, achieving efficiency gains of up to 80%." Avant Tecno manufactures compact loaders and attachments for landscaping, property maintenance, agriculture, construction, industry, and private use. Otto Hintikka, Sales Executive, Cloud Platforms and Security, Fellowmind Finland, on a manufacturing cloud partnership aimed at fast and cost-efficient results.
  • Fellowmind, "Fellowmind and Avant Tecno win Microsoft AI Agent Race Finland 2026," 7 May 2026. Winner announced at Microsoft AI Tour on 28 April 2026. The competition is for Finnish agent solutions already in production with measurable business value. Finalists were judged on customer impact, deployment, clarity, and repeatability. The winning solution automates handling of orders and supplier confirmations in a Dynamics 365 environment. The implementation accelerated processes by up to 80 percent and reduced manual work. Experts focus on more demanding tasks. The work is described as embedded in everyday processes without heavy changes to existing ways of working. Fellowmind HiveAI is named as the technology for developing, managing, and expanding the agents. Puputti's sentence appears again, worded as efficiency improvements of up to 80 percent on customer-order processing.
  • Karl-Johan Spiik, Microsoft MVP, "Agent Stories Part 2: Fellowmind x Avant Tecno — When Simple Agents Create Real Business Value," 2 June 2026. A write-up of Miika Tekoniemi's session. The use cases named are sales-order processing and purchase-order matching. The confirmation example, before the agent: a person reads emails and attachments, extracts the details, compares them with internal systems, and identifies mismatches on dates, quantities, and prices. With the agent: email triggers the process, documents are processed with AI, data is extracted and validated, consistency is checked against existing orders, and only exceptions are escalated. The architecture he prints: trigger, for example email via Power Automate; AI processing for extraction and understanding; deterministic validation; a decision and notification layer. AI handles unstructured emails and PDFs. Rules handle business validation and decisions. Implementation effort measured in days, return on investment achieved quickly, automation rates meaningful early. He does not print a percentage, a day count, or a euro amount for that return. Governance: monitor what agents do, validate outputs before decisions, keep business rules in control. The lesson he draws from the customer side is to build the small solution now. This page uses his account for the shape of the confirmation path and for the size of the effort. It does not treat "only exceptions" as a description of the customer-order agent, because the webinar says the expert still checks that one.
  • Fellowmind, "Tekoälyagentit osana osto-, myyntitilaus- ja asiakaspalveluprosessia." Finnish webinar page, Case Avant Tecno. The first agent speeds entry of customer orders into the ERP. The second, described as just deployed, automates processing of supplier purchase-order confirmations and frees expert time. The headline places "jopa 80%" on that pair. Puputti, in Finnish: the first deployed agent streamlines entry of orders customers send into the ERP, and with the agent's help in customer service the processing work is up to 80% more efficient. A later paragraph: because of the new agent, entering order data into the ERP now takes only two minutes. An expert still does a check, and a considerable amount of working time is freed. The page also says, as advice to the audience, that a company's first agents can be taken into use in a few weeks. Miika Tekoniemi is listed among the webinar experts. This page does not use the few weeks as Avant Tecno's implementation clock.
  • Avant Tecno, avanttecno.com. Avant Tecno Oy, Ylötie 1, 33470 Ylöjärvi, Finland. The attachments section says over 200 attachments. This page uses the site for the address and the attachment range. It does not use the site for the headcount, the turnover, or the agent results.

Stay in the loop

Product updates and enterprise AI notes. No spam.