The State of AI in ERP Adoption

AI in ERP adoption has reached near-universality at the experimentation layer and stalled at the scale layer. McKinsey's State of AI 2025 survey finds that 88% of organizations now use AI in at least one business function, yet only about 6% qualify as "high performers" capturing significant EBIT impact, and just 23% report using AI agents in workflows. In ERP specifically, the picture is sharper: vendors have shipped copilots and the first autonomous agents into finance and supply chain, customers are buying the licenses — close to 70% of the Fortune 500 now pay for Microsoft 365 Copilot — but Gartner's 2025 Microsoft 365 Copilot Survey of 215 IT leaders shows 94% reporting measurable benefits alongside only 6% completing global rollouts. The state of AI in ERP in 2026 is best summarized as licensed broadly, deployed narrowly, and agentic only in pilots.

This piece is deliberately about adoption state — how far buyers have actually gone, measured by attach rate, pilot-to-production conversion, and agentic deployment — not a rehash of where AI helps ERP delivery. For the practitioner view of which AI features are worth using inside an ERP project, see our AI in ERP delivery guide.

The headline: adoption is universal, scaling is not

The single most important fact about AI in ERP in 2026 is the gap between the top-line adoption number and the bottom-line scaling number. McKinsey's annual survey, released November 2025, reports that 88% of organizations have adopted AI in at least one function — a number that has been climbing for three years and is now close to a ceiling. But the share of organizations that have embedded AI into core workflows sits around 38%, the share scaling AI enterprise-wide is closer to one-third, and only about 6% meet McKinsey's bar for a "high performer" — an organization capturing significant bottom-line value from AI.

For ERP buyers, the relevant translation is this: the question is no longer "are we using AI?" — almost everyone is, because vendors have shipped AI into the products buyers were already licensing. The question is "has AI changed a core finance, supply chain, or operations workflow that moves a real number?" There, the answer is still mostly no, and the reasons cluster around data, governance, and change management rather than the models themselves.

The McKinsey survey also documents the structural tension that explains the gap: organizations report that AI investment frequently comes at the expense of giving ERP the data and system capabilities AI needs to scale. In other words, the same budget that funds copilot licenses is being pulled from the data-quality, master-data-management, and process-redesign work that determines whether those copilots return value.

Copilot attach: the most-watched number in ERP-adjacent AI

Because Microsoft owns the productivity layer that surrounds most ERP estates, Microsoft 365 Copilot attach is the closest thing the industry has to a real-time AI-adoption barometer. The numbers tell a coherent story:

• Paid seats are large and growing. Microsoft reported more than 15 million paid Microsoft 365 Copilot seats by 2026, with some trackers putting the figure above 20 million, and roughly 218 million active users across Windows, web, and mobile, according to adoption trackers compiling Microsoft disclosures.
• Enterprise is the dominant segment. Microsoft reported 420 million monthly active Copilot users in Q1 2026, up from 230 million in Q1 2025, with enterprise licenses (paid Copilot for Microsoft 365 subscriptions) accounting for roughly 38% of the total.
• The Fortune 500 buys in. Nearly 70% of the Fortune 500 now pays for Microsoft 365 Copilot, making it the fastest-selling add-on in Microsoft's history by the company's own framing.
• Conversion is the bottleneck. One adoption analysis pegs the workplace conversion rate — paid seats as a share of the addressable M365 installed base — at about 35.8%, meaning most of the people who could be on a Copilot seat are not.

The attach number matters for ERP specifically because Copilot is the gateway drug to Dynamics 365's deeper AI. A buyer who has already standardized on M365 Copilot at US$30/user/month is primed to extend into Microsoft Copilot inside Dynamics 365, where the AI is embedded in finance, supply chain, and business-central workflows rather than sitting alongside them in Office.

But attach is not adoption. Gartner's 2025 survey of 215 IT leaders exposed the paradox cleanly: 94% of organizations report measurable benefits from Microsoft 365 Copilot, yet only 6% have completed global rollouts, and 72% remain in pilot or limited deployment. Of the organizations that had finished pilots, just 5% said they were moving to larger deployment in 2025. Nearly half rated the product "some value, shows promise" — a polite way of saying the business case is not yet proven at scale.

A separate Morgan Stanley / RSM AI Adopter Survey from July 2025 put enterprise deployment at 79%, with about half of those having moved past pilot into full rollout — a higher figure than Gartner's, reflecting different sampling and definitions, but the same direction of travel: deployment is broad, deep rollout is narrow. Even the Wall Street Journal, reporting in February 2026, noted that Copilot is "losing ground with users" on engagement despite the seat counts — a reminder that licensed ≠ used ≠ valuable.

Inside Dynamics 365: Copilot credits and the first agentic agent

If M365 Copilot is the productivity-layer story, Dynamics 365 is the ERP-layer story, and it moved decisively in early 2026. The March 2026 Dynamics 365 Licensing Guide introduced a Copilot Credits model that reframes how AI inside Dynamics is sold and metered: Dynamics 365 Finance Premium now includes 1,000 Copilot Credits per user per month, with usage-based billing for additional AI capacity. That is a meaningful shift away from flat per-seat AI pricing toward metered consumption — the same direction the broader AI infrastructure market is moving.

Functionally, the in-product AI has graduated from summarization to action. Dynamics 365 Finance Copilot supports account summaries, workflow analysis, and collections communication. Supply Chain Management Copilot produces AI-generated summaries that help users understand key supply chain data quickly. Business Central offers bank reconciliation with Copilot that checks ledger entries against bank statements and suggests matches.

The most significant 2026 development inside Dynamics is the Account Reconciliation Agent in Dynamics 365 Finance, currently in production-ready preview. It continuously monitors transactions and performs reconciliation tasks — a step toward autonomous operation inside the general ledger. Microsoft is positioning it as deployable with human-in-the-loop controls, which is the responsible-AI framing every ERP buyer should require before letting an agent touch posted entries.

The pricing dimension matters for adoption forecasting. At US$30/user/month on top of a qualifying M365 license, Copilot is a material line item once you scale beyond a handful of power users — and the move to Copilot Credits means the cost curve can bend upward as usage deepens, not just as seats grow. Buyers planning to embed AI in core ERP workflows should model both the seat cost and the credit consumption, not just one.

SAP and Oracle: the agentic ERP arms race

Microsoft is not the only vendor pushing AI deeper into ERP, and the competitive framing has shifted from "copilots" to "agents." SAP's Sapphire 2026 keynote was built around the "Autonomous Enterprise" vision, with Joule Work, the SAP Business AI Platform, and the SAP Autonomous Suite. SAP launched Joule Studio 2.0 for building custom agents and backed it with a €100 million Business AI Partner-Led Adoption Program organized around four service packages — Adoption, Launch, Performance, and Enterprise — each with escalating requirements for deploying custom agents. The existence of a partner-led adoption fund is itself an adoption-state signal: SAP is subsidizing the partner ecosystem to push customers past pilot, which implies SAP believes customers are not getting there on their own.

Oracle NetSuite's October 2025 SuiteCloud platform update added "Ask Oracle" (a natural-language query interface), agentic workflows, and text-enhancement features, with a 2026 roadmap extending agentic capabilities further. Oracle's positioning leans on NetSuite's unified data model as the foundation that makes agentic AI tractable — the implicit argument being that AI inside an ERP with a single source of truth is lower-risk than AI bolted onto a fragmented estate.

The common thread across SAP, Oracle, and Microsoft is that all three have moved from "AI that summarizes" to "AI that acts" — agents that plan and execute multi-step workflows inside finance and operations. That is the genuine frontier of ERP AI in 2026, and it is where the adoption-state data is thinnest, because most customers have not yet deployed agents into production finance workflows.

Odoo: the model-agnostic challenger

The Microsoft / SAP / Oracle tier all embed their own AI models and meter them. Odoo, the open-source ERP that increasingly competes for the mid-market, takes a structurally different position that matters for adoption diversity. In Odoo v19, the built-in AI supports Google Gemini and OpenAI GPT models — Odoo does not natively integrate Microsoft Copilot. On Odoo Online (SaaS), a free AI tier is included with no API key required, capped at usage limits; on Odoo.sh or on-premise, you supply your own OpenAI or Gemini API keys and pay the provider's token rates directly, with no Odoo markup.

This matters for the adoption picture because it lowers the floor. A mid-market company that wants to experiment with AI in its ERP does not need to commit to a US$30/user/month enterprise add-on or to a vendor-locked agent platform. It can run a free SaaS AI tier, or bring its own keys and pay marginal token costs. The trade-off is governance: model-agnostic AI outside a unified identity and governance framework (like Microsoft's Entra/Purview stack) puts more of the responsible-AI burden on the customer and its implementation partner.

Odoo's native AI features — Ask AI, lead scoring, OCR, email drafting, summarization — ship with the platform. Pricing starts around US$24.90/user/month on the Standard plan (first year, billed annually) and renews higher; treat these as benchmarks, not quotes.

The agentic pilot gap: 23% experimenting, roughly 6% scaling

The single sharpest data point in the 2026 adoption picture is the gap between experimenting with agents and running agents in production. McKinsey's State of AI 2025 reports that 23% of organizations are now using AI agents in workflows — systems based on foundation models capable of acting in the real world, planning and executing multiple steps. That is up meaningfully year over year and represents the leading edge of enterprise AI.

But the same survey shows only about 6% of organizations are high performers capturing significant EBIT impact, and the broader scaling numbers (roughly one-third scaling enterprise-wide) confirm that most agent activity is pilot-stage. Applied to ERP, this means: the Account Reconciliation Agent in Dynamics 365 Finance, the Joule-based finance agents in S/4HANA, and the agentic workflows in NetSuite are real products with real capabilities — but the share of ERP estates where an agent is autonomously touching a posted transaction, unattended, is very small.

This is not a reason to wait. The McKinsey data on agentic AI in implementation specifically is more optimistic: McKinsey reports that early-adopter teams have cut ERP implementation time and effort by more than 50% using AI agents, with program timelines roughly halved. That is a delivery-layer result, not a run-the-business result, and it is qualified by methodology — but it is the strongest evidence we have that the agentic layer is already moving real numbers when applied to the right work. The gap between "agents help us build the ERP" and "agents run the ERP" is the central adoption story of 2026.

Why adoption stalls: data, governance, and change

When adoption stalls, the cause is almost never the model. Three failure modes dominate, and they map cleanly onto why ERP is the hardest place to scale AI.

Data readiness. Poor data quality — incomplete, unclassified, inconsistent, or missing-context data — is a primary driver of AI hallucinations in enterprise environments, beyond model limitations alone. ERP is the system where data quality problems accumulate, because ERP is the system of record every other system writes to. If your master data is dirty, your AI copilot will confidently draft wrong answers from it, and an agent acting on those answers can post a wrong entry. Data readiness is upstream of AI readiness, and the McKinsey tension noted above — AI budget cannibalizing data budget — is the structural version of this problem.

Governance and human-in-the-loop. ERP sits at the intersection of finance, operations, and compliance, where AI errors are most expensive. The operating model that makes AI safe here is human-in-the-loop: trained humans retain decision authority over high-risk AI actions and approve them before they commit. Microsoft's responsible-AI principles — fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability — are the canonical reference and map cleanly to ERP. The Account Reconciliation Agent ships with approval gates for exactly this reason. Governance is not a brake on adoption; it is the precondition for adoption at scale.

Change management. The widely cited 55–75% ERP failure range is overwhelmingly people-and-process failure, not software failure, and Gartner's prediction that more than 70% of recently implemented ERP initiatives will fail to fully meet their original business-case goals by 2027 is driven by the same dynamic. No copilot sponsors a steering committee, runs a super-user enablement session, or unblocks a resistant function head. AI adoption inside ERP inherits every change-management problem the underlying ERP project already had, plus a new one: users have to learn to review AI output critically rather than either ignoring it or trusting it blindly.

What the numbers mean for your ERP roadmap

The adoption-state data resolves into a few concrete implications for any organization planning an ERP move or an AI extension in 2026.

• 88% use AI in a function; ~6% are high performers — Assume your competitors are experimenting; do not assume they are ahead in production. Plan to be in the scaling 6%, not the experimenting 88%.
• 94% report Copilot benefits; only 6% complete global rollouts — Pilot value is real but does not self-scale. Budget for enablement and change management equal to the license spend.
• Copilot Credits model (March 2026) — Model AI cost as metered consumption, not flat per-seat. Stress-test credit usage before scaling.
• 23% experimenting with agents; ~6% high-performing — Treat agentic finance/supply-chain AI as 2027 production, 2026 pilot. Pick one bounded workflow (reconciliation, collections, demand forecasting) and prove it.
• Gartner: >70% of ERP initiatives miss business case by 2027 — AI does not change the failure modes — data, governance, change. It amplifies them. Fix the fundamentals first.
• McKinsey: >50% implementation-effort reduction for early adopters — Apply AI to the delivery layer now (discovery, documentation, testing, scaffolding). Apply it to run-the-business workflows carefully.

The pattern across every row is the same: the technology is ready, the organizations are not. An ERP engagement that ignores this pattern will buy licenses and wonder why the ROI slide never lands.

How to benchmark your own AI-in-ERP maturity

Because the market data is noisy and definitions vary, the most useful exercise is to benchmark your own organization against a simple four-stage maturity model that the adoption data implies.

Stage 0 — Not started. No AI in any ERP workflow, no copilot licenses, no agent pilots. This is now a minority position, but it still exists, especially in smaller mid-market and on-premise estates.

Stage 1 — Licensed, not used. Copilot or equivalent licenses are bought (often as part of a bundle), but active usage is concentrated in a handful of power users and the finance/supply-chain AI features are dormant. Gartner's data suggests this is the most common state for large organizations.

Stage 2 — Piloting in a bounded workflow. One or two AI use cases are running in a real but contained workflow — bank reconciliation with Copilot, collections email drafting, demand-forecast overlays — with human review on every output. This is where McKinsey's 23% agent-experimentation figure lives, and it is where most of the genuine learning happens.

Stage 3 — Scaling in core workflows. AI is embedded in a core finance, supply chain, or operations workflow that moves a real number, with defined governance, monitoring, and a named owner for every AI-touched process. This is McKinsey's ~6% high-performer tier, and it requires the data, governance, and change work that stages 1 and 2 skip.

The honest self-assessment for most organizations in 2026 is Stage 1 trending toward Stage 2. The goal of a 2026 roadmap should be to move one bounded workflow from Stage 2 to Stage 3 — not to boil the ocean with enterprise-wide agent deployments the organization is not ready to govern.

What changes in the next 12 months

Three shifts are already visible in the data and will reshape the adoption picture through 2026 and into 2027.

First, metered AI pricing will replace flat per-seat pricing across the major ERP vendors. Microsoft's Copilot Credits move is the leading indicator; SAP's AI-units model and Oracle's consumption-based additions point the same direction. For buyers, this means AI cost will scale with usage depth, not just seat count — and the organizations that get value will be the ones whose usage justifies the metered bill, not the ones that simply buy the most seats.

Second, the agent layer will move from preview to production in finance. The Dynamics 365 Account Reconciliation Agent is in production-ready preview; SAP's Joule-based finance agents and NetSuite's agentic workflows are on the same trajectory. Expect at least one major ERP vendor to ship a generally available, autonomous finance agent in 2026 — and expect the responsible-AI guardrails (approval gates, audit logs, human-in-the-loop on postings) to be the differentiator that determines which customers can actually turn it on.

Third, the data-readiness bottleneck will become explicit. The McKinsey tension — AI budget cannibalizing the data work AI depends on — is already forcing organizations to choose. The ones that win will be the ones that treat master-data management, data quality, and process documentation as the precondition for AI, not a competing priority. Gartner's prediction that more than 70% of ERP initiatives will miss their business case by 2027 is, at root, a prediction that most organizations will keep making this trade-off badly.

The state of AI in ERP adoption in 2026 is not a story about whether AI works. It works — the 94% benefit figure and the >50% implementation-effort reduction are real. It is a story about whether organizations can turn working AI into scaled AI inside the hardest, most regulated, most change-resistant system they run. The technology is past the demo phase. The organizations are mostly not. For most buyers, the right next step is not more licenses; it is one bounded, governed, human-reviewed workflow in production — and a partner who can tell the difference between the two.