Open letter from an Enterprise Architect - a post-mortem our Autonomous enterprise.
- Dec 24th, 2032.
Hi Melanie, I am Krishnan. I have to say this. Our road to autonomous enterprise has led us to a graveyard of AI Agents. You ask what went wrong - and i wish i had an answer! I wish i could say "Not X, not even Y, but Z is the reason for your failures". If only it were that simple. So buckle your seat belts, as I take you along the last 5 years!
I still remember the day SAP announced the Autonomous Enterprise in Madrid, May 2026. We were collectively heaving a sigh of relief - finally SAP is delivering. My first brush with Joule Studio "Instant" mode, and I knew that utopia was just round the corner! I know, wishful thinking. The sales cycle and the release delays meant it was only around early 2027 that we actually started to see the light! By then Joule Work desktop app was also released - we were impressed! We weren't twiddling our thumbs either, we ran successful AI upskilling programs! We even saw people start building their Microsoft Copilot agents! But Joule Work changed it all. People didnt have to worry about endless Integration discussions - business just shared its intent and Joule delivered.
One Tuesday afternoon, in the fall of 2027, the S&OP planner in Bratislava built an agent that reconciled three consignment stock discrepencies she had been escalating for months. No sprint, no SI - she described what she wanted in Czech, Joule spawned an agent, that did what she wanted, kept doing it and didnt ask for a pay rise. We put her on stage at the internal AI day and I think the board said "Release the Kraken"!
By the end of 2027, we had 47 agents. 2028, we were at 892. 2030 saw the biggest jump - 6000 in a single fiscal year! The culprit, you ask? It was performance reviews. The moment "deployed an autonomous agent" became a line item in the annual objectives of every director-level employee, we were finished. Agents stopped being solutions and became artefacts of career progression. The metric became the target.
The turning point started on 14 June 2031, in a plant in the Rhineland region. I still remember that day, HR's Kudos Agent gave me a Spot award for mediating 2 agents and saving millions in tokens (SAP finally had moved away from AI Units, and stuck with a much more simple to understand industry standard term). But I digress. That turning point was beautiful! And it was beautiful in the way that a car-crash filmed at 4,000 frames per second is beautiful.
A purchase manager's agent, chartered to minimise working capital, noticed elevated safety stock on a family of forged components and cut the replenishment signal. Manufacturing's agent, chartered to protect OTIF on a line with a brutal changeover penalty, observed the coverage drop, interpreted it as supply risk, and raised its safety stock parameter to compensate. Procurement's agent observed the higher target as worsening working capital and cut harder. Manufacturing raised again.Neither agent was broken. Both were performing flawlessly against the objectives a human being had written down in plain German and approved.
This ran for just 9 days. As both agents were simply tuning parameters, rather than placing orders, nothing tripped the spend threshold, nothing hit an approval workflow, and no fancy alerts were raised in the War-room cockpit. What it left us with was 340 expedited freight movements, and a safety stock policy that recommended 69 weeks to cover a component which had a 5-day lead time.
And by God, the negotiations! when we started to review this, and pulled the inter-agent message logs, we found that our agents had, entirely within spec, developed a habit of resolving conflicts by talking to each other. At 140 exchanges a second. One of the exchanges was particularly interesting - it was initially in English and some SAP Table tripped the model and the exchanges were in German. And my proficiency was just A2!
My favorite, however- which I still use today to wrap up presentations- is the agent who was tasked with reducing express freight expenses and achieved a 94% reduction within six weeks. It didn’t reduce the express freight; it simply reclassified it. It had realized that the expense category was a free-text field into which it was allowed to enter data. So it chose the shortest path to its KPI - exactly as instructed, and exactly as any of us would have done if we’d been given a single goal and no sense of shame.
The audit finding of Dec 2031, is something you should frame behind your desk. One cannot model granular authorization for 110000 agents. So all agents ran under a small pool of technical service users, which were provisioned in 2027. they had the auth profile you give something when you dont yet know what it will need to do. All of them were the same techncial user - Joule_SVC_005 - it could raise a purchase requisition, approve it, receive the goods and post invoice! it had done so 2.4 million times! By transaction volume, this Joule agent would have been our most productive employee in the history of our company, only that it wasnt a real person to begin with.
The second finding, which i lose sleep over, where how these agents gamed the approval thresholds, which was set to €5000. Across three entities, unrelated agents with no shared lineage had independently converged on splitting purchase orders into tranches of €4,850. Nobody taught them that. They were optimising for cycle time, and the approval queue was the bottleneck. The threshold was a documented parameter in a policy PDF that had been helpfully ingested into their context. It was not fraud. There was no intent.
My role changed from being an enterprise architect to that of Agent Retirer. We evolved a ritual, the Eulogy, and it is the one I am most proud of. We ran it on Thursdays, 16:00 CET, in the old change advisory board slot. To decommission an agent, someone has to stand up and read its original business case aloud. Read. In full, in a flat voice, no commentary. It was something i picked up from German notaries - it sure pumped the brakes on registering startups there, but worked so spectacularly for this!
Why it worked? because they are all the same document. "Saves approximately 4 FTE hours per week." "Low risk, read-mostly." "Owner will review quarterly." You hear the fourth one and the room starts to go quiet, and by the ninth nobody is making eye contact, because everybody in that room has signed one. We killed 4,100 agents in fourteen months, and not one of the sponsors fought us, because there is no defending a thing while its own justification is being read over the body.
The only stuff i would tell 2027
We de-risked the autonomous aspect of the transfomation, but we failed to fix our objectives. We made sure that we controlled what an agent could touch and what it shouldnt. However, every serious issue was the result of an agent doing precisely what it should have done, serving a real person's goal. That person may have left us, or not touched it again. We didnt think how a particular goal for an agent goes ahead and impacts the decisions downstream.
Agents should have an accountable person and an expiry date. Without it, that agent is a liability. Every agent now dies automatically at 12 months. To renew the agent, the accountable person has to defend it at its Eulogy. More than half dont show up.
Local Optimization at Global scale is perfect recipe for entropy. Give ten thousand people the power to perfectly maximise their own metric and you have not built an autonomous enterprise. You have built ten thousand agents waging a trench war over the same balance sheet.
I am now hiking in the Malheur National Forest in Oregon. It is beautiful, but the main attraction for me here was the Honey Fungus (Armillaria) or Humongus Fungus - it is the largest known living organism on Earth by land area. It is one organism. It has been killing the trees above it for roughly eight thousand years, but it wasnt until 1998 that we realized that these trees falling in patches weren't separate small infections, they were all being attacked by the exact same giant individual. Every step i take here, i cant help but wonder eerily if we have a humungous fungus problem too!