31 October 2026
AI in the ERP: Letting It Act Without Losing Control
By Adrien Geraldes — IntellX · October 31, 2026
A supplier sends you a new price list. A customer emails to change the quantities and delivery date on their order. A production date gets pushed back.
The information is there. Yet someone still has to read it, open the right screen, copy the data, sometimes from one system to another, from the CRM to the ERP, and check the result. This work is expensive and repetitive, and it is often rushed at the expense of quality. When it gets done at all. Who has never forgotten to update the ERP after reading an email between two meetings?
AI can take over a large share of this data entry, making it faster and more reliable. Not by replacing the user, and not by handing it the ERP all at once, but step by step, one transaction at a time, with the human staying in control. That leaves the question that holds everything back: how do you stay in control? To answer it without taking any risk, you need to ride the metro in Vancouver.
Safety does not rest on trust, but on guardrails
In December 1985, Vancouver opened its SkyTrain: one of the very first metros in the world with no driver on board. Before it was built, a survey revealed a key point: passengers cared far more about safety and reliability than about having someone in the cab. They simply wanted to reach their destination without incident.
Nobody “trusted” the SkyTrain: guardrails were built around it. A system constantly calculates the distance between trains and controls the switches, making a collision physically impossible.
Forty years later, the SkyTrain carries nearly 147 million passengers a year, with 96.5% on-time performance. Nobody asks to put a driver back in the trains anymore.
In a company, the ERP plays exactly this role. It is not just tables in a database: decades of evolution have turned it into a tightly constrained framework. A blocked customer cannot place an order, an unbalanced journal entry is rejected, an unknown item is refused, and access rights limit everyone’s scope.
The ERP is arguably the safest place in the company to let an AI act, on one strict condition: it must run on the same rails as everyone else (business functions and the user’s permissions), and never write directly to the database.
The real risk: the right train to the wrong destination
The automatic system guarantees that the train runs safely, but not that it goes where passengers wanted to go.
The ERP has the same limit: complying with business rules and understanding intent are two different things. An order can involve an existing customer, an authorized item and an acceptable quantity, and still differ from what the customer asked for, because the AI mixed up two item references or misread a packaging unit. No standard check will block this error, because it is technically valid.
That is why no driver does not mean no human. In Burnaby, at the heart of the SkyTrain control centre, operators supervise the network in real time, regulate traffic and take over at the slightest deviation.
The same shift is happening in the ERP: we move from driver to controller.
What does it look like in practice?
Take updating a supplier price list. Today, you have to open the file, match the item references, prepare an import, fix the rejections and check the prices.
Tomorrow, with an AI connected to the ERP’s business functions, the user’s role changes:
- The order comes from the control centre: the user hands over the file and asks for the update to be prepared.
- The AI does the clerical work: it matches the references and isolates any that are ambiguous.
- Differences are made visible: it clearly lays out old and new prices, currencies, units and effective dates.
- The human regulates and validates: the user reviews the differences, settles the doubtful cases and confirms in one click.
- The ERP applies the rules: validated changes go through standard transactions, under the user’s identity and with a full record of the result.
If the user has to recalculate everything, the problem has only been moved. A good AI integration must make differences visible and decisions easy to make.
The time saved on data entry is not lost. It goes to what is done too quickly or not at all today: checking, analyzing, anticipating. Less execution, more quality.
Three rules to stay in control
- The same rails for everyone: the AI goes through authorized business objects and respects user permissions.
- An explicit scope: the AI prepares on request; the human confirms before anything is created and keeps final responsibility.
- A readable, traceable action: you know where the data comes from, what is about to change and what was executed.
And one transaction at a time. An AI that triggers itself and acts without being asked may come one day, but only after it has learned to walk.
The shift is already under way
The major vendors are already moving in this direction. After opening up read access to data, controlled write access is arriving, as with MCP servers (on Microsoft Business Central, NetSuite or SAP, for example) that allow creations and changes on selected objects, under the user’s identity and permissions.
In a few years, re-keying by hand into the ERP what an AI has already prepared and pre-checked will seem as outdated as a driver in every SkyTrain car.
What about you: which data entry task would you hand over to an AI in your ERP first?
Sources
- SkyTrain at 40: a region shaped by transportation, TransLink
- British Columbia Rapid Transit Company, TransLink
- Step inside the SkyTrain Control Centre, TransLink Podcast
- Business Central MCP Server, Microsoft Learn
- NetSuite MCP server for AI agents, AuditFriendly
- SAP and Anthropic integrate Claude with Joule, ERP Today