Agentic AI Is About to Break Seat-Based ERP Pricing. Ask These Three Questions Before You Sign.
On 1 July, Gartner put a number on something that has been vague for a year. Up to $234 billion of enterprise application spending is exposed to what it calls agentic arbitrage between now and 2030, which works out to roughly a fifth of all enterprise SaaS spend.
Agentic arbitrage is a clumsy name for a simple mechanism. When an agent completes a task across several systems, nobody logs into those systems. And when nobody logs in, nobody needs a seat. Gartner’s blunter phrasing is that “the user interface is no longer a differentiation”.
If you run a 40 person distributor, your instinct is that this is a problem for SAP and Salesforce rather than for you. Mostly correct. But it changes one decision that a lot of smaller companies are making this year.
The contract you sign in 2026 outlives the assumption behind it
ERP contracts run three to five years. Most are priced per named user. That model assumes headcount is a decent proxy for how much value you get out of the software, which has been roughly true for thirty years.
If agents start absorbing the data entry, the reconciliation and the exception triage, the number of humans opening your ERP falls while the volume of work flowing through it stays flat or grows. Under per-seat pricing you would pay less for the same work. Vendors can see that too, which is why pricing is drifting toward consumption, platform fees and outcome-based tiers.
You do not need to guess which model wins. You need to avoid being locked to the losing side of the transition for five years.
Three questions worth putting to any ERP vendor before you sign this year:
- If the number of people using the system drops by a third because automation absorbs the work, what happens to our bill?
- Is API and agent access included, metered, or a separate SKU? Ask for the number, not the roadmap.
- What are the exit terms on our data and configuration if we later move the agent layer somewhere else?
A vendor who cannot answer the second one has not thought about this yet, which is itself useful information.
The cost story has a second half
The other thing that happened this month cuts against the optimism. Reporting in late July found that agentic workloads are driving token consumption and infrastructure spend up faster than expected, and that the savings CIOs had budgeted for have not fully arrived. OpenAI moving to bill for agent usage turns experimentation into an operating line item rather than a pilot budget.
So the honest version is not that agents make your ERP cheaper. It is that agents move ERP cost from a predictable per-head figure to a variable per-work figure. If your volume is seasonal, that is a materially different risk profile, and it is worth modelling before it arrives rather than after.
What actually decides whether any of this works for you
The most important line in the Gartner note is also the least quoted one: better outcomes require systems that retain institutional memory and context over time, and bolting on more AI features tends to add cost rather than outcomes.
Translated for a smaller company, the agent is not the hard part. What matters is whether your supplier records are deduplicated, whether your item master means the same thing in two departments, and whether your roles genuinely reflect who should approve what. An agent working over messy master data produces confident wrong answers faster than any human could.
That is the same argument as designing agent permissions before agent capabilities, and it is why we open ERP engagements with a data and roles audit instead of a tooling conversation.
If you are on an older platform and wondering whether it can host any of this at all, the answer is more often yes than people expect, as an additive layer rather than a replacement.
Our free assessment is one call plus a written note on where your data and permissions would break an agent. For most companies that is the honest first step, and it costs nothing to find out.