The Autonomy Gap: Vendors Sell Autonomous, Operations Deliver Partial
Vendors sell the autonomous enterprise; operations deliver partial automation. Gartner, DSAG and the Lemvigh-Müller case show the real maturity gap.
The slide says “autonomous.” A procurement agent runs the buying process end to end, no human in the loop, a clean arrow from order to confirmation. The board likes the slide. It approves the project.
Then the system goes live. The agent drafts the order confirmation, flags the exception, and waits. A buyer reviews the edge cases. The standard volume flows through untouched; the deviations land on a desk. That is not the slide the board signed off on, but something better that almost nobody budgeted for.
No Mittelstand company built an autonomous enterprise in 2026. They automated stretches of a process, with the proposal from the machine and the sign-off from a person, and that partial automation is the entire deliverable rather than a cheap rehearsal for the real thing.
The vision-versus-reality gap is measured, not felt
This is not a consultant’s hunch but a pattern in the trade press, dated and on the record.
TechTarget put it plainly on 26 June 2026: “the gap between vendor vision and buyer deployment reality remains wide.” An agent is only as good as the foundation it sits on — orchestration, data quality, access rights, and the human handoffs between steps. Strip any of those away and the autonomous demo stays a demo.
The people running these projects say the same thing. At a Computer Weekly CIO roundtable on 23 June 2026, Hayley McKelvey of Deloitte was blunt: “You cannot just drop these agents into an organisation and expect immediate value.” Spending budget on agents is not the same as realising value from them — a distinction the procurement slide rarely makes.
Constellation Research reached the same conclusion from the buyer side on 28 June 2026: what enterprise buyers actually deploy points to “more human oversight, not labour replacement.” The honest deployments add a reviewer rather than removing one.
The maturity stage has a number on it
The autonomous-enterprise pitch breaks against arithmetic.
Gartner predicted on 25 June 2025 that more than 40 percent of agentic-AI projects will be cancelled by the end of 2027. The reasons are not exotic: escalating costs, unclear business value, and “agent washing” — RPA scripts and chatbots relabelled as agents. Of the thousands of vendors claiming agentic capabilities, Gartner counts roughly 130 as genuine. For buyers weighing an “autonomous” purchase today, the headline number is starker still. Agents are projected to make 15 percent of day-to-day work decisions autonomously by 2028, up from zero percent in 2024.
Zero in 2024. That is the baseline a 2026 budget is paying a premium to leapfrog.
The German-speaking market shows the same shape. DSAG (the German-speaking SAP user group, whose investment reports are a standard reference for SAP-running Mittelstand) published its 2026 figures on 26 February 2026 (n=198). Forty-three percent of respondents use AI productively. But only 3 percent of those run it on SAP’s own agent; 77 percent reach for non-SAP tools. AI agents, the report notes, are “not yet deployed across the board.” The native vendor agent, the one on the autonomous slide, is the rarest thing in the room.
Why this lands now — the counter-data point
On 26 June 2026, SAP published a customer case that is the honest version of the autonomous pitch — and reads as its quiet correction.
Danish distributor Lemvigh-Müller deployed three agents to handle order confirmations: one for inbound email, one for data extraction, one for matching. The result: roughly 100,000 of about 175,000 orders processed automatically, near 60 percent of the volume, with three to four full-time roles freed up and ten weeks from start to production. The deviations still go to the buyers.
Read it carefully. This is not a free-floating “autonomous enterprise” but embedded partial automation inside a core process: high standard volume runs automatically, the exceptions route to a human. That is exactly the architecture the vendor slide flattens into a single arrow. The Lemvigh-Müller case succeeded precisely by staying partial — the standard volume ran automatically, the exceptions kept a human.
From “we’ll make the company autonomous” to a gated set
The fix is not to reject agents but to stop buying “the company” and start buying named stretches of named processes. The test I run before any vendor conversation gets to a contract:
Key Takeaways
Why this matters: most agent projects do not die from bad technology. They die because the buyer scoped “autonomous” as a slogan, not as a set of process steps each agent can defensibly own — so the contract promises a maturity stage the floor cannot deliver, and the project is cancelled before the slide becomes real.
Run this before the next vendor call:
- List the individual process steps, not “procurement” as one block. An agent never automates a department; it automates a step.
- Score each step against four fields:
| Field | Pass condition | Fail signal |
|---|---|---|
| Volume / standardisation | High volume, repeatable pattern | Low volume or every case unique |
| Rule clarity | The decision rule is explicit and stable | Rule lives in someone’s head |
| Damage if wrong | Cheap, reversible, caught downstream | Costly, hard to reverse, or compliance-critical |
| Audit trail | Every action is logged and traceable | No record of what the agent decided or why |
- Release for full automation only the steps where all four fields pass. Set everything else to “propose plus human sign-off.”
- Write the vendor contract against the gated set, not against the “autonomous enterprise” headline.
Before/after: Lemvigh-Müller scored its order-confirmation flow this way and released roughly 60 percent of orders to full automation while keeping the deviations on a buyer’s desk — three to four full-time roles freed, ten weeks to production. The “before” was a department described as one block; the “after” was a process split into steps the agent could and could not own.
Where it does not apply: this test fails if the process has no clean master-data foundation. Without it, the agent automates the error along with the work — it confirms the wrong order, faster. A weak audit trail is also a hard stop, because an authorisation profile before the agent acts is the precondition for letting it act at all, and without an ERP connection AI stays an expensive toy no matter how it is gated.
What the market numbers don’t show
The vendor and the hype-loyal colleague have a fair reply: “This is a snapshot. Partial automation today, end to end in two years — anyone not building on ‘autonomous’ now misses the curve.”
The objection is not wrong about its direction, only about the timing of the purchase decision. Budgeting a 2028 maturity stage with 2026 money buys a slide, not a capability, and the cancellation statistics fall on exactly those projects. You can build toward more autonomy without paying today for a level the floor does not yet deliver.
And here is the stronger counter, the one that keeps this from being vendor-bashing. In tightly regulated or high-volume standard processes, more autonomy is already defensible today. Lemvigh-Müller runs 100,000 confirmations with only the deviations escalated. The line does not run along “autonomous yes or no” but along a different question: which stretch is standardised volume, and which stretch needs judgement. The market numbers (Gartner’s cancellation rate, DSAG’s 3 percent) show the average. They do not show that the gated, high-volume cases are already working. The slide hides both halves: it oversells the average and undersells the cases where partial automation already pays. The job is to draw the line the deck refuses to draw. That is also the argument against treating AI as a reason to delay ERP decisions, because the gated case is buildable now, on the foundation you already run.
A regulatory footnote that closes the “invisible” option
Even if an agent did run fully autonomously, one door stays shut. From 2 August 2026, Article 50 of Regulation (EU) 2024/1689, the EU AI Act, requires that AI systems interacting with people, and AI-generated content, be identifiable as such. The obligation applies regardless of risk class, which means it reaches internal assistants and agents too. “Fully autonomous and invisible to everyone” is therefore not only technically immature; it is legally closed off, with penalties up to 15 million euros or 3 percent of group turnover for breaches. The transparency rule and the maturity curve point the same way: the human stays in the loop, by design and by law. For where this fits a broader strategy, see where AI actually helps in the Mittelstand.
Frequently asked questions about the autonomy gap
What is the difference between “autonomous” and “partially automated” AI — and why should it matter before a vendor call?
Autonomous means the agent decides and acts without human sign-off. Partially automated means the agent proposes, or clears the standard case, while a person approves it or handles the exception. In 2026, Mittelstand deployments deliver the second almost without exception. Gartner (25 June 2025) does not see autonomous day-to-day decisions reaching 15 percent until 2028. Buying “autonomous” today budgets a maturity stage that operations cannot yet deliver.
Does “over 40 percent will be cancelled” mean AI agents aren’t worth it?
No. Gartner (25 June 2025) reads that figure as an early market shakeout, not a technology failure. What gets cancelled are projects without clear value, with runaway costs, or relabelled RPA dressed up as agents. Projects with a clear, narrowly scoped use case, such as a high-volume confirmation match, survive and pay.
We want to automate procurement. How do we decide what the agent may do on its own?
Do not assess “procurement” as a block. Score each process step against four fields: high volume with a clear rule, low damage if wrong, and a working audit trail. Full automation is defensible only where all four hold; everything else runs as “propose plus sign-off.” Lemvigh-Müller (SAP News, 26 June 2026) automated roughly 60 percent of its orders this way and kept exactly the deviations with a person.
Does the EU AI Act apply even if our agent runs internally?
Yes. From 2 August 2026, Article 50 of Regulation (EU) 2024/1689 requires that AI systems interacting with people and AI-generated content be identifiable as such. That holds independent of risk class, so it covers internal assistants and agents. “Fully autonomous and invisible” is not only immature; it is legally blocked, with fines up to 15 million euros or 3 percent of group turnover.
If everything today is partial automation, why not wait for the agents to mature?
Because partial automation already delivers the value — full autonomy is not the prerequisite for it. Three to four full-time roles freed in ten weeks (Lemvigh-Müller, 26 June 2026) is a result, not a stepping stone. Waiting means leaving the real lever on the table to hold out for a slide that, on Gartner’s maturity curve, is still years away.
Next step
Which of your process steps can an agent actually run alone — and which ones keep a human in the loop?
I’m happy to walk through your specific processes against the four-field test and mark, step by step, what is defensibly automatable today and what stays on “propose plus sign-off” — no cost, no pitch.
→ Or read first: Assess processes for AI automation · Where AI actually helps in the Mittelstand
Sources and links: Gartner: Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 June 2025) · DSAG Investment Report 2026 (26 February 2026) · TechTarget: Partial automation is the real promise of AI agents (26 June 2026) · Constellation Research: What we learned about AI projects from enterprise buyers (28 June 2026) · SAP News: AI Agents to Take Over 100,000 Order Confirmations at Lemvigh-Müller (26 June 2026) · Computer Weekly: UK tech chiefs on agentic AI (23 June 2026) · EU AI Act — Article 50, Regulation (EU) 2024/1689
Read more on pfisterer.xyz: An authorisation profile before the agent acts · Without an ERP connection, AI stays an expensive toy · AI as an ERP exclusion criterion