The time AI gives back belongs to someone
Mission: BETTER! on the capacity dividend: AI hands time and capacity back, but unless leadership decides who owns that capacity and where it goes, the div
Mission: BETTER! — I know my AI resilience score. Who gets my Wednesday?
An AI agent can draft a functional spec for three business units, each with its own exception to the same SAP process. Give me a quiet afternoon and I can probably come back with all three versions. Before I do, I’d like to know whether we still need three exceptions. And who can say yes to having just one.
Suppose I finish on Tuesday a piece of work we’d allowed until Friday. If it’s my own project, I know exactly which idea I’d like to try on Wednesday. On a programme, somebody may already have promised my Wednesday to the next project.
Meanwhile, the design authority still meets on Friday. My calendar has opened up.
The business unit waiting for a decision hasn’t gained a day.
What did we actually save?
Finishing on Tuesday tells us when the document was ready. To know how much work we saved, I’d want to know how long comparable specs took and how much time colleagues spent checking this one. To know what the business gained, I’d ask what happened next.
Suppose we spend Wednesday with the three process owners. The first exception protects a commitment made to a customer. Keep it. The second compensates for a limitation in a system we’re retiring. The third has survived two migrations because somebody copied it into the next design.
Nobody in the room can remember the last time it was used.
If the owners check those last two and agree to retire them, Friday’s meeting has a different decision in front of it. Two variants come out of scope before anyone builds them. That’s two fewer variants to test at the next release and two fewer things for someone to understand when a customer calls on a Monday morning.
I would happily spend the whole of Wednesday on that conversation.
A timesheet might show the saving had disappeared. Ask the team responsible for the next upgrade and you might get a different answer.
There’s a useful parallel in a six-month field experiment involving 6,000 knowledge workers. People who used Copilot spent less time on email. The researchers found no significant change in time spent in meetings. I can change how I handle my inbox this afternoon. Moving Friday’s decision takes a few other people. Microsoft Research, Shifting Work Patterns with Generative AI
A mathematician would ask what we had proved
In the conversation you’ll find in Lex Fridman’s mathematics episode, the guests discuss what changes when a proof can be checked in Lean. A machine can verify that a conclusion follows from the formal statement and its assumptions. A mathematician can still want to know why the result holds, which assumption does the work, and whether an interesting special case changes their understanding. Those are different questions, and the ability to settle one gives people room to pursue the others. Lex Fridman Podcast #503, transcript archive
Alex Kontorovich put the point in seven words: “Math is not the fact, it’s the why.”
There’s a narrower version of that distinction in SAP. Give me three approved requirements and I can test whether a design satisfies all three. If one requirement exists only because of an old system limitation, the tests can pass beautifully. We’ve just made an obsolete workaround one of our acceptance criteria. Software engineers call the first check verification: did we build the product right?
Validation asks whether we built the right product.
AI has made the first question much cheaper.
The second still needs someone with the authority to delete a requirement.
After more than 25 years in SAP and enterprise transformation, that’s the part I’d watch. The cost of describing one more exception is coming down. The cost of living with it can land years later, in somebody else’s budget. I’d like the person who will carry that cost in the room while the requirement is still a sentence we can delete.
Where the 88 comes in
RankedIn read my LinkedIn export and gave me 82 out of 100 for AI resilience, a laurel wreath and the title Orchestrator. A day later, with an export three months newer, I scored 88. Both scores sit in the range it calls resilient. I’ll take the wreath. Neither run had access to Wednesday’s calendar.
Apparently this is how RankedIn.app sees me 😄
The score feels close to how I see my own field.
This may be the best time I’ve known to work in software.
Ideas used to sit for months while we found the budget, the skills and a week when the right people were free. Now I can close the door, work with several AI tools and get far enough in a quiet afternoon to see whether an idea is worth pursuing. For me, a good deal of AI’s value is that freedom to try.
How would I recognise the same freedom in an organisation?
I’d follow one recurring decision through the business. Process exceptions will do. When did the request arrive? When was there enough evidence to decide? Who had the authority to say yes or no? At the next release, did the decision hold, or did somebody have to open it again? SAP can now turn its own AI agents’ execution logs into event logs for process mining. The wait for Friday’s decision won’t appear in them unless someone records the human decisions with the same care.
A few weeks of that would tell me where to spend the next euro. If sound proposals sit for a fortnight waiting for a signature, buying more drafting capacity will give the approver a longer reading list. If decisions come quickly and keep reopening, I’d look at whose knowledge was missing the first time. And if the team spends an extra Wednesday retiring something it would otherwise support for five years, I’d want that Wednesday in the business case.
In Denmark, researchers linked surveys of chatbot adoption to employment records. Users reported saving an average of 2.8% of their working hours.
Over the period studied, the researchers found no significant change in earnings or recorded hours.
Payroll can tell me whether someone was paid more. It can’t tell me whether they finally got an afternoon to fix the problem everyone had been working around. Humlum and Vestergaard, Large Language Models, Small Labor Market Effects
In banking, the question of who answers for an agent’s result already has a legal form. In the UK, the FCA’s Mills Review concluded this year that the Senior Managers Regime still applies as AI acts more autonomously. If a bank can name who answers for an agent’s mistake, it can name who decides what happens to the agent’s savings.
So who gets my Wednesday?
In practice, whoever gets to decide what goes into my calendar.
When that’s me, I know what the freedom is worth.
When I’m the one assigning somebody else’s next piece of work, I owe them more than a fuller calendar.
The next time someone tells me AI saved their team three days, I’d like to ask the team one question.
An earlier version of this article appeared in Unfiltered: SAP Public Cloud. My last Unfiltered issue asked whether to spend the next euro on AI credits or on the Public Cloud move.
This edition takes the same question to people and organisations: what AI changes in how we work, and who decides what happens to the time it gives back.
Mission: BETTER! is about how decisions get made, so it belongs here.
Further reading
Dillon, Jaffe, Immorlica and Stanton, Shifting Work Patterns with Generative AI, NBER Working Paper 33795 (2025). The full paper behind the Copilot field experiment.
https://www.nber.org/papers/w33795
Humlum and Vestergaard, Still Waters, Rapid Currents (RFBerlin, March 2026). The revised version of the Danish study.
https://www.rfberlin.com/wp-content/uploads/2026/03/26078.pdf
METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (2025). Developers took 19% longer with AI and still believed it had made them 20% faster.
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
Agrawal, Gans and Goldfarb, Prediction, Judgment and Complexity: A Theory of Decision Making and Artificial Intelligence, NBER Working Paper 24243. Why judgment and decision rights matter more as prediction gets cheaper.
https://www.nber.org/papers/w24243
Brynjolfsson, Rock and Syverson, The Productivity J-Curve, American Economic Journal: Macroeconomics (2021). Why the gains from a new technology show up only after organisations change how they work.
https://www.aeaweb.org/articles?id=10.1257/mac.20180386
FCA, AI and the Future of Retail Financial Services (the Mills Review), July 2026. The primary source for the banking paragraph.
https://www.fca.org.uk/publication/corporate/the-mills-review.pdf
SAP Help, SAP AI Agent Mining (SAP Signavio Process Intelligence). How agent execution logs become event logs for process mining.
SAP Help, Extensibility KPIs in SAP Cloud ALM. Clean core levels and the Technical Debt Score: the cost of keeping an exception, measured.
Mission: BETTER! offers independent executive perspectives on SAP Cloud ERP, Business AI and enterprise transformation. More of this work: sapperment | Executive Clarity.
Independent publication. Not affiliated with or endorsed by SAP SE. SAP is a registered trademark of SAP SE.Today, an agent can draft a configuration, propose a mapping or write the first version of a functional spec. It cannot decide whether three business units really need three different exceptions, and it does not answer for the result.
First published on LinkedIn. Subscribe to Mission: BETTER! to get new editions as they appear — or browse the sapperment library.