40% of Agentic AI Projects Will Cancel by 2027 (Gartner). Why SAP Joule Adopters Are Especially at Risk — and How to De-Risk.

Gartner's June 2025 prediction landed quietly on most enterprise roadmaps: over 40% of agentic AI projects will be canceled by the end of 2027. The reasons it cited — escalating costs, unclear business value, inadequate risk controls — are the unfashionable middle of the AI conversation. They are also the reasons most SAP-running enterprises will hit a wall.

If you are a CIO or CTO inside an SAP shop, the number that matters more than 40% is a different one. Only 3% of SAP customers are running SAP Joule in production today (CIO.com, March 2026). Joule has shipped, the Joule Studio Agent Builder shipped to general availability in Q1 2026, and SAP's roadmap is the most ambitious it has ever been. But three years into the agentic AI wave, the production adoption rate inside the SAP installed base is in the single digits.

Those two numbers describe the same problem from different angles.

The 40% number, in plain language

Gartner's projection is not a knock on agentic AI. It is a knock on how enterprises are buying it. Most agentic projects today are not projects in the traditional sense — they are proofs-of-concept that escape the lab in a weekend and accumulate sponsor expectations they were never built to meet. By the time the IT steering committee asks for the audit trail, the cost model, the model lineage, and the SoD posture, the answer is usually "we'll add that in v2."

There is no v2. There is a project review meeting. Then a quiet line item on the next quarter's deferred-projects list.

The shops that will survive Gartner's 40% cull have three things in common. They picked agentic targets where the value is provable in weeks, not quarters. They built governance into the deployment from day one, not as a layer on top. And they avoided getting locked into a single model or a single agent runtime before they understood which trade-offs they would actually live with in production.

Why SAP Joule adopters carry extra exposure

SAP Joule is the right answer to a specific question: how do I make my SAP environment conversational and agent-native without leaving the SAP plane? When the question is that one, Joule is excellent. The native skills library is real, the agent builder works, the compliance posture inside SAP's own boundary is tight.

The problem is that most enterprise agentic work isn't that question.

The same March 2026 CIO.com analysis reports that 77% of AI-active SAP shops are using non-SAP AI tools alongside Joule — Claude for long-context reasoning, Gemini for multimodal search, custom orchestration layers for cross-system workflows that touch ServiceNow, Workday, Salesforce, and homegrown middleware. None of that is wrong. It's just that it lives outside Joule's authority, which means it lives outside Joule's governance.

Three patterns surface again and again in failed SAP agentic projects:

1. Governance debt

The agent shipped fast because the team treated approval workflows, financial-authority limits, data-access escalation, and cross-system risk quarantine as "later." In every case, "later" became "after the audit committee asks." Retrofitting governance onto a deployed agent costs three to five times what building it in costs, and the business sponsor never wants to pay for it twice.

2. Single-model lock-in

The team standardized on whichever model the platform shipped with. Six months later, a competing model is dramatically better at the specific reasoning shape the workflow needs — financial reconciliation, supplier-risk scoring, SoD pattern matching — but the agent runtime cannot route to it. The team rebuilds. Or doesn't, and the workflow underperforms quietly.

3. Opaque cross-system orchestration

The agent works inside SAP. It also calls a non-SAP system. Nobody can answer where the trust boundary is, what the rollback semantics are, or which side owns the audit trail when something fails. The first incident review forces a rewrite.

These are not Joule problems. They are problems of running Joule alone, in a stack that needs more than Joule alone provides. SAP knows this, which is exactly why SAP partners exist.

The augmentation pattern: Joule is the conversation, Qubi is the execution

Qubittron is a SAP Silver Partner authorized under SAP PartnerEdge for Sell and Service. Our position on Joule is unambiguous: we extend Joule's reach, we do not replace it. The two are explicitly collaborative.

The division of labor we recommend — and the one we ship to production every week — is straightforward:

Joule is the conversation. Qubi is the execution.

Joule handles the SAP-native interaction surface, the in-context skills, and the agent flows where SAP's own model broker is the right call. Qubi handles the work that lives next to SAP rather than inside it: long-running agents that span hours or days, orchestration across non-SAP systems, custom Segregation of Duties enforcement that goes beyond SAP GRC's native rule library, and multi-model reasoning where you need Claude or Gemini or a domain-tuned model alongside SAP-approved models.

The augmentation pattern doesn't replace Joule's identity, audit, or governance plane. It extends them. The agent flow continues to authenticate against your SAP identity provider. The audit trail continues to land in your existing observability stack. The SoD posture continues to be enforced — extended, in our case, with continuous monitoring rather than the quarterly cycle most GRC programs run on.

That continuity is the entire point. The reason 40% of agentic projects will cancel is that they introduced a new identity plane, a new audit plane, a new procurement contract, and a new vendor concentration risk — all in service of a workflow that the business could not yet quantify. Augmentation reverses every one of those bets.

An eight-point de-risk checklist

If you are a CIO or CTO scoping agentic work that involves SAP, run any project — yours, your partner's, or your incumbent vendor's — through these eight checks before you fund the next phase.

1. Identity continuity. The agent authenticates against your existing SAP identity provider (and, if you have one, your enterprise SSO). No new identity plane. No service accounts that bypass IT.
2. Audit continuity. Every agent action — reads, writes, model calls, downstream system calls — emits an event to your existing observability stack (Splunk, Datadog, your SIEM of record). No new audit plane that requires its own dashboard.
3. Reversibility. Every write the agent performs has a documented rollback path. SAP transactions can be reversed; non-SAP API calls have idempotency keys or compensating actions.
4. Model portability. The agent is not locked to a single foundation model. The runtime supports BYOK across at least OpenAI, Anthropic, Google, and Mistral, and a model swap is a configuration change, not a rewrite.
5. Cross-system trust boundaries. For any agent flow that touches a system outside SAP, the trust boundary is documented and the failure semantics are explicit. Who owns the audit trail when the third-party API returns 500?
6. Continuous SoD posture. SoD enforcement is continuous, not quarterly. Violations are detected at the agent-action level, before they post to the SAP system of record, not at the next quarter's GRC cycle.
7. Provable ROI in 90 days. The first agent ships against a workflow where the savings (or revenue, or risk reduction) can be measured against a baseline you already have. If the business case can't be defended to the CFO in 90 days, scope smaller.
8. Exit cost <= 90 days. If the project gets canceled, the business can walk away in 90 days or less. Data is portable. Workflows can be reconstituted in another runtime. No multi-year contracts that survive a project death.

Score honestly. Anything below 6 of 8 is a Gartner statistic in waiting.

What this looks like in production

We had this exact conversation with a US-based manufacturing CTO last year. His team had spent thousands trying to make AI work in their SAP production planning environment. Pilot after pilot stalled. The agent that did ship was opaque, single-model, and required a manual approval queue that the planning team eventually stopped using.

When we picked the work back up under the augmentation pattern — Joule conversational on top, Qubi handling the execution and the cross-system orchestration into the planning systems — the calculus changed:

"We spent thousands trying to make AI work in our SAP production planning. It didn't ship. Qubi delivered in three months at one-tenth the cost — governed inside our own infrastructure." — CTO, Manufacturing, United States

The unglamorous parts of that quote are the operative ones. Three months, not nine. Inside our own infrastructure, not on a vendor-owned plane. One-tenth the cost, because the augmentation pattern reused the existing identity, audit, and rollback infrastructure rather than introducing a parallel one.

That's not a Qubi-only outcome. It's the outcome any augmentation-first agentic project should target. You can run the eight checks above against any vendor or SI partner you're considering, including ours.

Where to start

If your SAP shop is in the 97% that has not yet put Joule in production, the first move is not to pick a runtime. It is to pick the right first workflow. The strongest first workflows we see in our pipeline today are:

• Continuous SoD monitoring that replaces the quarterly GRC cycle with weekly or daily reports landing in the SAP CoE Director's inbox
• Production planning agents that handle the routine planning decisions and surface only the exceptions to human planners
• Procure-to-Pay augmentation where Joule handles supplier-facing conversations and Qubi handles the audited Ariba write-back and exception routing
• Requirements capture in SAP services engagements, where every requirements meeting is captured, gaps are flagged at week three, and delivery risk is visible before go-live rather than after

Whichever you pick, run it through the eight-point checklist before you fund it. And anchor the architecture in the augmentation pattern, not in a parallel-stack replacement bet that the business won't fund through its second year.

If you'd like to walk through your specific landscape — which SAP modules, which deployment mode (self-hosted, BTP, or SaaS), which non-SAP systems are in scope — book a 30-minute call or reach out via the contact form. We'll map your candidate workflows against the eight-point checklist and give you a straight read on which ones are worth funding now and which ones aren't.

Joule's adoption curve will steepen. The 3% will become 30%. But the SAP shops that get there first will be the ones that built the augmentation pattern in from day one — and treated agentic AI as an extension of the SAP discipline they already run on, not a new vendor-managed silo.

---

SAP, S/4HANA, ECC, SuccessFactors, Joule, Ariba, BTP, GRC, and PartnerEdge are trademarks of SAP SE. Qubittron is an independent SAP Silver Partner with PartnerEdge Sell and Service authorizations.