Automating a process from email to ERP when the flow runs on its own and when a person approves

What does automating a process from email to ERP mean?

It means that a document someone copies by hand today becomes data that goes into the management system on its own, after written checks and, where needed, the approval of a person. The process can be a customer's order, a supplier's invoice, a return request; the example in this guide is orders, because they are the most common case and an error in one reaches the customer straight away.

The company in the example has 30 people, sells components to other businesses and receives about 40 orders a day from 200 customers, each in its own format: PDF, spreadsheet, text in the email. Today an employee copies them into the management system and sends the order confirmation. The flow that takes them in has six steps.

Ask Google or an AI assistant how it is done and you almost always find a flow of four to eight steps, and often a traffic light: green goes through on its own, amber a person approves, red is done by hand. On 30 October 2026 we read those answers, and none says how many orders it takes, or how low an error is low enough, to move a type of order from amber to green. That is the question this guide answers.

If you are wondering whether you need an AI agent, a system that chooses on its own which tools to use, the answer is in the guide AI agents for SMEs. Here the flow is written in advance, and AI works only where free text arrives: it recognises the document and reads its data. AI classifies, prepares drafts and finds gaps; it is not the authority that decides.

What do human in the loop, on the loop and out of the loop mean?

They are three ways of keeping a person in an automated flow, and they differ by when the person steps in: before the single case takes effect, after, or never on the single case. The first, human in the loop, is where you start.

The distinction is rarely written down: of the 174 pages cited by Google and the AI assistants that we read, 52 name human in the loop, 7 human on the loop and 3 human out of the loop. And in everyday language “out of the loop” can also just mean “without approval”. It pays to tell them apart, because the risk is different. An order that goes in without anyone approving it, with a person rechecking a sample and watching the indicators, is on the loop; an order nobody will ever look at is out of the loop. For orders we recommend on the loop, never out of the loop: even when nobody approves before, someone looks after and can stop the flow.

The AI Act requires human oversight only of high-risk systems, in article 14, and a flow that enters orders between businesses is not among the high-risk uses: the person in the flow is a design choice, not an obligation. The same article, though, names the risk that applies to every flow where a person approves, “automation bias”, the tendency to over-rely on what the system proposes. Someone who confirms forty drafts a day, right thirty-nine times, ends up confirming the fortieth without looking at it. That is why the draft shows the uncertain fields and the unusual ones at the top, and whoever approves must have the competence and the authority to say no, as article 26 requires for high-risk systems.

Which steps can run on their own, and which does a person have to approve?

It is not decided by step but by type of case. The same entry can go through on its own for the order of a regular customer, with codes and prices already known, and go to a person for a new customer's first order. For each case three questions count: whether the error can be undone, what it costs when it reaches the customer, how often it happens, measured on your orders.

The typical case is the ambiguous code: ABC-450 and ABC-450-R, one at 12 € and one at 18 €. The flow does not choose; it stops the line and shows it with both candidates.

Nearly all the AI assistants' answers use the model's “confidence”, and Google's AI Overview and AI Mode let an order read with confidence above 95% go through on its own. It is a weak criterion: a language model (LLM) does not by itself give a probability calibrated on your orders, and it can read with certainty a code that exists and is the wrong one. Confidence is for deciding which drafts to look at first. An order goes through on its own because it passed every check, not because the model says it is sure.

Which numbers let an order go in without approval?

When, for that type of order, the measured error is below the threshold beyond which the check costs more than the errors it stops, and enough orders have been seen to say so. Three numbers are needed: what a check costs, what an error that reaches the customer costs, how many orders of that type have gone through without corrections.

The type of order, or class, is defined by the rules of the previous step: for example a regular customer, a known format, all codes already matched, list prices, normal quantities. As long as a person approves every order, the flow records for each draft whether it was confirmed as it was or corrected, and what changed. That is the measurement the decision rests on.

How many orders does it take to say so?

If n orders in a row had no correction, with 95% confidence the true error is below 3 in n. It is the rule of three, born in medicine for events a study did not observe (Hanley and Lippman-Hand, JAMA, 1983).

Perplexity, in two answers in Italian, gives a time-based rule: wait for “at least a month of accuracy history”, or 2-3 months. Time alone says little. A month of a type of order that arrives five times a day makes about 110 orders, and even without a single correction they only say the error is below 2.7%.

How low does the error have to be?

Lower than the cost of a check divided by the cost of an error. An example with hypothetical numbers: checking a draft takes a minute, which at 30 € an hour is 0.50 €; an error that reaches the customer, between return, new shipment, credit note and hours spent, costs 150 €. The check pays as long as the error is above 0.33%, that is 0.50 divided by 150. To say with 95% confidence that it is below takes about 900 orders of that type without corrections: at 15 orders a day, three months of work.

The threshold moves with the cost of the error. For a catalogue component the customer sends back, 150 € may be right; for an order that starts production or a purchase, an error costs far more, the threshold drops and the numbers to pass it never arrive. Those orders stay in the loop.

The measurement has a limit: it counts the corrections of whoever approves, and whoever approves can miss an error. In the first weeks it pays for a second person to recheck a sample of the orders confirmed without changes, and to count also the errors that surface later, from returns and credit notes.

How do you move over, and how do you go back?

The move happens in stages. A type of order moves on the loop with a sample: for example a person rechecks one order in ten, then one in twenty, and every week looks at the indicators, that is orders changed after entry, complaints, returns. It goes back in the loop when something changes that the measurement did not see: a new version of the model or of the instructions, a customer changing the format of their orders or their codes, a new price list, an error found in the sample. Going through without approval is decided class by class, on the pilot's numbers, and revoked when something changes.

In the systems we build you start with a person confirming every draft, and every correction stays on record: that is where these numbers come from. Removing approval from a type of order is your choice, made on the pilot's numbers, and a person stays on changes, cancellations and risky operations.

A case published by a supplier shows how far it can go. For Biodéal, a distributor of organic dairy products, Mirage Metrics' flow created in SAP Business One 2,943 orders received by email from 113 customers, with a median of 48 seconds from the email arriving to the order being ready; about 7% of the orders were flagged for a human check (Mirage Metrics, Biodéal case). The case does not say how many of the orders that went in on their own contained an error: that is the number to ask every supplier for, together with how they measure it.

What arrives, and what reads it?

Not everything needs a model: it depends on how the order arrives. A file with fixed columns is read by a classic program, which always gives the same result; the model is needed where the document is free-form.

If orders already arrive as CSV, a simple reader costs less and is more reliable than a model, and a spreadsheet is worth trying with a classic program before paying for a model. The model earns its keep on free text: “50 pieces of the galvanised bolt from last time” can only be understood by looking at that customer's previous orders.

The model must be able to say “I don't know”. A field it cannot find stays empty and the draft goes to a person, instead of holding a plausible, invented value. And if your largest customers can send orders as data, from their management system or a portal, that is the cheapest route: the AI flow is for all the others. Many processes, though, need a well built form and a clean database, not an agent, and in that case we tell you so.

Which checks are needed before an order goes into the ERP?

The ones that do not depend on the model: every piece of data read is compared with something the model did not write. The risks of a system that reads the mail, from instructions hidden in a PDF to bank detail fraud, are in the guide AI agents for SMEs; here are the checks of the flow.

The duplicate check has a technical name, idempotency: processing the same email again, after an error or a restart, must not create a second order.

Every step leaves a trace: the document, the data read, the checks passed or failed, who approved and when. It serves to reconstruct an error even months later and to measure the flow, as the FAQ on traceability explains. In the flows we build, no data enters the management system without being compared with the customer and item records, the price lists or the history.

What does automating orders from email to ERP cost?

A custom flow costs from 10,000 € for the first process, plus model usage, after a feasibility study that costs up to 2,500 €, deducted from the pilot if you sign. These are Volcanic Minds' entry thresholds: the price of your project is written by the study, with a range and a ceiling. The platforms and services named by the AI assistants have public price lists: the table shows those read on 30 October 2026, excluding VAT.

The subscription is not the whole cost: with a platform, someone builds the flows, the rules and the connection to the management system, you or whoever helps you. For project figures, asking an AI assistant helps little: on 30 October 2026, asked the same question, Google's AI Mode first wrote 1,500-5,000 € to start, then 2,500-6,000 € with a platform; ChatGPT first 15,000-32,000 €, then three bands between 4,000 and over 50,000 €. No answer cited a primary source, and price lists need checking too: in two answers ChatGPT gave Power Automate 15, 150 and 215 € a month, which are the US prices in dollars.

Model usage costs cents per order. An estimate, on our assumptions: an email with a one or two page PDF, plus the instructions and the candidate items taken from the item records, makes about 10,000 input tokens and 1,000 output; Anthropic estimates 1,500 to 3,000 text tokens for each PDF page, plus those of the page image. With Claude Sonnet 5.5, at $2 per million input tokens and $10 output at 30 October 2026 prices, that is 3 cents per pass, 6 with two passes per order. The example's 40 orders a day, about 880 a month, come to about $50 a month, less if part of them arrives as data. The real number is measured in the pilot as a cost per order.

The connection is the item that weighs most. We connect to ERP, CRM, MES, MRP, APS and legacy systems through REST APIs, batch, WebSocket, MQTT, AMQP and other secure channels. If the management system has no API we use the route it allows, a file import or an exchange table; a robot that uses the screens like a person is the last choice, because it breaks when a screen changes. Then there is hosting and, if you want it, maintenance, which is a separate contract paid on the work done, as the guide on maintenance explains. There is no licence fee on the software we develop for you.

How much time is saved, and where does the saving come from?

Almost all of it from not copying; removing approval adds little time, and its real advantage is speed. An example with hypothetical numbers on the company with 40 orders a day: copying an order takes 5 minutes. With the flow, 30 drafts out of 40 are right and checking them takes a minute, while the other 10 are fixed in 5 minutes as today. Then one type of order, 15 a day from regular customers, moves on the loop, with one order in ten rechecked.

Not copying is worth 120 minutes a day; removing approval from 15 orders is worth another 13. These are assumptions to show the calculation, but the proportion holds with different numbers too: checking a right draft takes little, copying an order takes a lot. The reason to remove approval is usually another, speed: an order that arrives at seven in the evening goes in at once, and the confirmation leaves within a minute instead of the next morning. In the Biodéal case the median is 48 seconds from email to order ready.

The savings the AI assistants give, up to “from 5 minutes to 15 seconds per order” from Google's AI Mode, have no measurement behind them, and none separates the time of the check from the time of the reading. Your numbers are measured by the pilot, and the time freed becomes a saving only if you use it: to reply sooner, or to handle more orders without hiring.

Where do you start?

From today's numbers, then from a few customers, with a person approving everything. Each step collects the measurements needed to decide the next one.

1. Measure the starting point: orders a day by format and by customer, minutes per order, errors that reached customers in the last year and what they cost. The last number is needed for the threshold.
2. Choose the first customers: the five or ten that send the most orders, with formats that differ from each other.
3. Run in parallel for a few weeks: the flow prepares the drafts without writing anything, and they are compared with what was entered by hand.
4. Move in the loop: the flow writes, a person approves every draft, and every correction stays on record by field and by type of order.
5. Move one type of order at a time on the loop, when its numbers pass the threshold, with a checked sample; the others stay under approval.

The phases and prices are those of the guide AI in business: where to start. The feasibility study, 1-2 weeks and up to 2,500 €, says which customers to start from, how to connect to the management system, what model usage costs and the criterion for stopping. The study stays yours even if you do not sign, and if you sign its cost is deducted from the pilot. The pilot, 4-6 weeks from 10,000 €, covers the parallel run and the approval of every order, steps 3 and 4, and gives the first numbers by type of order. Integration and governance take the process into production: from the first meeting to a process in production takes between 11 and 15 weeks. Moving a type of order on the loop comes later, when the orders seen are enough.

On the agentic AI audit platform the measurement was the same: the system's work compared with a person's. In the measurements made together with the client the agent collects 60% more data than interviews run by hand, with 88% accuracy against an auditor's work, and each audit saves 4 hours. How we build and govern systems with several agents is on the page on AI agent orchestration.

When must a person stay?

When the error costs too much, when the operation cannot be undone, or when the orders are too few to measure.

• Changes and cancellations of orders already entered: they change a commitment you have already made, and often arrive in few words.
• Orders that start production or a purchase from a supplier: an error costs the production, not a shipment.
• New customers and first orders in a new format: there is no history to compare them with.
• Orders above a value threshold, which you set.
• Types of order that arrive rarely: at 5 orders a month, the 300 orders needed to say “below 1%” take five years.
• Anything that changes customer or item records, price lists or bank details, which no flow changes on its own.

In all these cases the flow still does the long work, reading and preparing the draft, and the person decides in a minute instead of copying in five. In the flows we build, a person always decides these cases, with the work already prepared by the flow.

Sources and further reading

• AI Act, article 14, “Human oversight”ai-act-service-desk.ec.europa.eu, accessed 30 October 2026
• AI Act, article 26, “Obligations of deployers of high-risk AI systems”ai-act-service-desk.ec.europa.eu, accessed 30 October 2026
• Hanley and Lippman-Hand, “If nothing goes wrong, is everything all right? Interpreting zero numerators”, JAMA, 1983pubmed.ncbi.nlm.nih.gov, accessed 30 October 2026
• Mirage Metrics, “SAP Order Entry Automation From Email: Biodéal Case”, updated 21 October 2026miragemetrics.com, accessed 30 October 2026
• Microsoft, Power Automate pricingmicrosoft.com, Italian price list, accessed 30 October 2026
• n8n, “Pricing”n8n.io, accessed 30 October 2026
• Microsoft, Azure Document Intelligence pricingazure.microsoft.com, West Europe region, accessed 30 October 2026
• Rossum, “Pricing”rossum.ai, accessed 30 October 2026
• Anthropic, “Pricing”, prices of the Claude modelsplatform.claude.com, accessed 30 October 2026
• Anthropic, “PDF support”, token estimate per pageplatform.claude.com, accessed 30 October 2026