Almost all of it is assistive
Six of nine vendors describe functionality that suggests, summarises, or drafts for a person to act on. That is genuinely useful and it is a different category from software that acts on its own authority.
Research · updated August 2026
Every ERP vendor has an AI announcement. Considerably fewer have generally available functionality, and fewer still have anything that acts without a person clicking approve. This tracks the gap, and we include ourselves in it.
Send a vendor's AI section. We will tell you what it commits to versus what it implies.
A bill arrives; an agent codes it and routes it.
The work starts without a person. Requires an authority model, an audit trail, and a policy engine — which is why it cannot be retrofitted onto a screen-driven system.
Where the category sits. Most shipped functionality is at the left of this spectrum regardless of how it is marketed.
What we found
Based on published documentation and generally available functionality rather than on announcements, roadmaps, or conference demonstrations.
Six of nine vendors describe functionality that suggests, summarises, or drafts for a person to act on. That is genuinely useful and it is a different category from software that acts on its own authority.
Two vendors document anything resembling a granted-authority framework. Without one, the question of what an agent may do without approval has no published answer.
Several publish a headline figure. None that we found publish per-workflow rates with an interquartile range, which is the figure a buyer would need to plan against.
Where AI acts, what gets logged ranges from the resulting change alone to the inputs, policy, and confidence behind it. Only the second is reviewable.
Most implementations run the model with the application’s access rather than as a scoped actor. That makes every audit assertion about automated activity weaker than it appears.
Across the nine, functionality announced twelve to eighteen months ago is frequently still in preview or limited release. Roadmap dates in this category should be weighted accordingly.
There are two genuinely different things being described with the same words. One is a model that reads your data and helps a person work faster. The other is software that performs a transaction under a granted authority, escalates when it is outside that authority, and records why.
Both are legitimate. The first is shipping widely and works. The second is early everywhere including here. What is not legitimate is describing the first in language that implies the second, which is currently the norm in this category.
When a model runs with the application’s access rather than as an actor with its own identity and permissions, two things follow. It can reach everything the application can reach, which is a much larger surface than any specific task needs. And the audit trail records the application as the actor, which means the log describes something that did not happen.
For an assistive feature that only reads and suggests, this is a minor concern. The moment anything writes, it becomes the difference between an audit trail an auditor can rely on and one that is technically accurate and substantively misleading.
The most striking finding is that no vendor we surveyed publishes per-workflow accuracy with a distribution. Several publish a single figure with no method, no sample size, and no definition of what counts as success.
That absence makes rational comparison impossible and it makes planning impossible too. A buyer cannot build a business case on a number whose method is undisclosed, so they either build it on hope or they do not build it at all.
We publish per-workflow rates with the interquartile spread, and our own numbers range from 94% down to 61% depending on the workflow. The bottom quartile of our customers sits materially below the median.
That is not a boast about maturity — it is early functionality with honest measurement attached. What we would defend is the architecture rather than the current accuracy: an authority model, a policy engine every actor passes through, and an audit record that includes reasoning. Those make the accuracy improvable safely, and they are considerably harder to retrofit than to build.
Nine mid-market ERP vendors surveyed between April and July 2026: NetSuite, Sage Intacct, Acumatica, Dynamics 365 Business Central, Odoo, SAP Business One, QuickBooks, Xero, and ourselves.
Assessment is based on published product documentation, release notes, and generally available functionality. Announcements, roadmap statements, conference demonstrations, and preview programmes are excluded, which is why several vendors appear less advanced here than their marketing suggests.
We did not have hands-on access to every product. Where documentation was ambiguous we recorded the more generous interpretation, so if anything this survey overstates rather than understates what is shipping.
This page decays faster than anything else on this site. It is dated at the top, reviewed quarterly, and we would expect several of these findings to be out of date within two quarters. Corrections from vendors are welcome and we will note when we make them.
Questions
What can it do unapproved, what is recorded, what blocks a bad release, and what are the published numbers.