SAP AI agents: 73% stalled, 12.6% report ROI

Two years after an AI contract is signed, the CFO asks the same question: what did we buy? A September CIO survey by Futurum puts a distribution on the answer. Of organisations that buy AI from outside vendors, 73.3% sit at “deployed but no ROI” or “ROI only inside a pilot”. The share reporting scaled, sustainable return is 12.6%.

In the same month, on the seller side, SAP moved agent ownership, risk tiering and post-deployment deviation tracking into its process and architecture line. Read the two together and one operational conclusion falls out: what blocks scale is not model capability, it is a missing producer for the evidence chain. The ROI number that can enter a budget process is, in most organisations, nobody’s job to record in production.

1. Between 73.3% and 12.6%, the missing piece is an auditable input

Futurum Group published the analysis on 24 September 2026 around SAP’s 22 September announcement at the Transformation Excellence Summit in Atlanta, and it rests on two surveys the firm ran itself. The first, a September 2026 CIO survey, produces the two headline shares: 73.3% of buyers stuck before scaled ROI, 12.6% reporting it. The second, an enterprise software decision-maker survey from August 2026 (n=833), found expected and actual return on recent software purchases both landing near 14% — a midpoint estimate — while 41.5% of decision-makers said verified ROI research would raise their confidence in allocating budget.

The three numbers only mean something together. The 41.5% describes the buyer’s decision mechanism: budget approval depends on getting hold of return evidence someone else can re-check. The 73.3% and the 12.6% describe the current supply of exactly that evidence: most organisations cannot produce it. The 14% in the middle says something else again — buyers do not hold inflated expectations, they simply keep missing them. That is not an expectation-management problem, it is a bookkeeping problem that starts after delivery.

Provenance matters here and should be stated plainly. Both sets of numbers come from Futurum’s own research, and the same page is how the firm sells custom research; the sample frame is not disclosed, so nobody outside can check who was asked, which industries were covered or whether clients were included. This is an interested-party survey, not a neutral census. The 14% is a midpoint estimate rather than a precise mean, and it must not be quoted as “AI returns are 14%”.

2. Why “deployed but not scaled” is not a model problem

Take an AI project apart and you find three accounts, and only two of them come with paperwork. The first is the purchase: budget released, contract signed, invoice booked. The second is the deployment: licences, compute, integration hours, all recorded. The third is the value: a role that no longer performs one step, a claim that is no longer reworked, a ticket that closes two days earlier. That third account happens inside the business process, and business processes do not usually issue invoices for the step that was saved.

So in most organisations the ROI is not a record, it is a study: sample, interview, estimate the hours saved, write a page. Studies have two structural weaknesses. They return an interval rather than a checkable quantity, so they cannot enter a budget process that approves on numbers. And their cost lands on the team that already has the least slack, which means the people best placed to produce the evidence are the least likely to have time for it.

That is the mechanism behind 73.3%, and it is not about models being unable to do the work. The work gets done, then leaves no auditable trace. Next year’s budget naturally fails to renew, because the decision-maker is holding nothing harder than last year’s deck.

The explanation has a boundary. It covers projects that work but cannot report. For projects that genuinely produced no value, missing evidence is not the problem, missing value is. There is a blunt test for which side you are on. List every AI action your organisation claimed as effective last quarter and ask, one by one: if we stopped doing this, which number would get worse, and by how much? Items with no answer are an evidence problem. Items with an answer and an unchanged number are a value problem.

3. Where SAP moved the evidence chain

The product surfaces in this round can be named. AI Agent Hub maps agents, models and MCP servers to a capability and an owner. Cloud ALM tracks agent actions in production and flags deviation. Signavio brings process mining and AI-assisted process design. LeanIX puts an architecture principles library and an AI enterprise architect assistant into the design stage. WalkMe handles adoption and access. Company Memory is still in preview.

Read as a list, that is a feature catalogue. Read together, it is a structural move: these are parts of the system of record. The answers to “who owns this agent” and “how far has it drifted since go-live” used to live in documents, spreadsheets and consulting reports. They are now being asked to live inside the ERP and process suite itself. Once they do, they become data you can export, slice by time window and drop into an audit pack, rather than questions that require a fresh study to answer.

Put differently, the producer of ROI evidence is shifting from external research to the system of record. For a buyer that shift is measurable. When the vendor’s own suite can emit agent action and deviation records, acceptance can demand an export instead of accepting a conclusion.

The boundaries belong in the same paragraph. The governance assistant inside this stack includes EU AI Act and NIST compliance intelligence, which is a separate lane and not what we are assessing here. Product names and capability descriptions are vendor statements; we have not verified them, and we do not claim these parts replace a buyer’s own acceptance responsibility. What we read is narrower — which data surfaces have been pulled into the system of record — and that is enough to support the structural claim above.

4. A second number set: what each dollar buys

Another dataset points the same way, but it cannot be merged with the first. In May 2026 Entelligence Research analysed 2,444 organisations and more than a million pull requests and found that of every dollar spent on AI coding tools, 18 cents becomes shipped product while 82 cents goes to fixing, reworking and reviewing AI-generated code. The figures reach us through an ORF analysis published on 19 September; the sampling method and definitions are not public, and the authors sit in the coding-tools market they are measuring.

The reason not to merge the two sets is simple. The 73.3% and 12.6% have organisations as their denominator; the 18 and 82 have dollars. They answer different questions — is there scaled value, and how much does each dollar buy. Standing side by side they corroborate each other; fused into a single “AI efficiency” figure they are wrong.

What we read in the 18 and 82 is a question about the unit of account. An organisation that books only inputs, and never records the closing action of shipping, can never tell which 18 cents were the effective ones. That is also the rule we use in our own agent operations, where cost is booked per successfully closed task rather than per token, for the same reason: tokens are input, closed tasks are output. We have broken down how that calculation works in a separate analysis, linked at the end of this article.

5. Our position

First, ROI evidence should be produced by the runtime, not by a study. Research answers “what is typical for projects like this”, which is the right input for a go decision. Runtime records answer “what did this quarter’s spend buy”, which is the right input for a renewal. Handing the second question to the first means re-running a study for every budget round: expensive to produce, soft to defend.

Second, the clause to change is the requirement, not the price. Do not ask a vendor what their ROI is. Require an export of the last 30 days of task-level records before the pilot ends: what the task was, what it cost, what the outcome was, who verified it. Drop any one of those four fields and the record cannot enter an audit pack. This is where the 41.5% becomes operational — if buyers want verified ROI research, the way to get it is to write verifiability into the contract rather than into expectations.

Third, whoever holds the system of record defines the unit of acceptance. SAP’s move is one sample of that on the seller side, and it holds symmetrically for buyers: hand the unit of acceptance to a vendor and you have handed over the bookkeeping. The unit can be a task, a ticket, a claim or a quote, but it has to be defined by the business, and it has to be defined before the project starts.

Fourth, a product existing is not evidence. We are not claiming that buying a governance product or a process suite produces ROI. Evidence exists only once a runtime records the action. Tools lower the cost of recording; they do not replace the record. For teams writing their next budget, we would put the record before the platform.

6. Limits and uncertainty

• Both Futurum surveys are the firm’s own, method undisclosed, on a page that sells its research services: an interested-party survey.
• SAP’s product capabilities are vendor statements. We have not verified them independently.
• The 73.3% and 12.6% denominator is buyers of external AI, not all enterprises.
• The 18 and 82 cents come from a May 2026 third-party analysis with unpublished sampling and definitions; it is not a census of the industry.
• We do not assess the compliance intelligence inside SAP’s governance assistant; that belongs to a different lane.

7. Three things to do before the next budget review

One: give every running AI project a closed unit — a claim, a quote, a ticket. Without a closed unit, every number downstream can only be written as an interval.

Two: write exportability into acceptance. Require 30 days of task-level records before the pilot closes; if the vendor cannot export them, the pilot did not pass. This is easier to verify than an ROI promise and much easier to phrase in a contract.

Three: look at the distribution of a quarter, not at a percentage. For the same closed unit, mean cost plus standard deviation surfaces trouble earlier than an ROI figure. A flat mean with a narrowing spread means the process is becoming predictable, and that is material you can put in front of a budget committee.

Do those three and what you hold is a record that can survive a budget process. The remaining question stops being whether AI is worth it, and becomes which way to push the slope of that record next quarter.

References: Futurum Group, “SAP Links AI Agents to Business Processes and Oversight” (2026-09-24, including its two proprietary surveys) futurumgroup.com/insights/sap-links-ai-agents-to-business-processes-and-oversight/ ; ORF, “From Tokens to Weights: The Changing Value of AI” (2026-09-19, citing Entelligence Research, May 2026) www.orfonline.org/expert-speak/from-tokens-to-weights-the-changing-value-of-ai