Dinand Tinholt (@tinholt) header
Dinand Tinholt (@tinholt) profile picture

Vice President @Capgemini. | Building at the intersection of AI, data, and business transformation | Views my own

Chicago, IL
Joined February 2009
Spoke at Microsoft Ignite today with Caroline Matthews, at the Anthropic booth. The subject: Claude agents working through a 40,000-deduction backlog in CPG, investigating each dispute and resolving it, and the process learning to fix itself as they went.
1
15
Exception-based oversight beat universal human approval by $21 to $29 per transaction across 9,000 agent decisions. Accuracy was 98 to 100% on complete records.
1
10
The Dog That Didn't Bark: when one decisive record was silently missing, accuracy fell to 30 to 38%. Agents approved 72 to 95% of defect cases without a flag when the missing record would have ruled out fraud. Full write-up: medium.com/@tinholt/the-dog-…
8
9,000 real agent decisions, three Claude models, five back-office workflows, four oversight regimes. Pre-registered. Full human approval never came out on top.
2
1
31
What happens when you send more agent work to review but keep the review team the same size? I ran 90,568 simulated configurations to find out.
1
25
Fixed team: the delegation margin falls, slope −0.593. Grow the team in step with the load: it rises, +1.148. Same routing decision, opposite sign.
1
18
90,880 configurations of agent organisations. Delegation depth weighed more against quality than anything else in the model. The first version had one flaw: its reviewers never got tired.
1
22
Let them tire and the lever drops from +1.46 to +0.93, turning negative past 80% sustained utilisation. Scaling staff with the escalations scored +1.15. Fixed headcount scored -0.59.
1
11
I rebuilt the game that got me into computers, with a model under my desk and a day of agents. A small 3-bit open-weights model rebuilt a 1988 engine from a spec: 503 tests, 78 hours, 26 kWh.
1
2
44
Then a blind gauntlet against a commercial bar built Mercy County in a day. Spec and tests first, cheap local build second. A ratio that applies far beyond games.
1
18
A fine-tuned encoder first reached its workflow ceiling at 1,000 labelled cases: 0.825 vs 0.822 on challenge cases. The from-scratch classifier never got there, even at 3,000. And the shipped eight-vote mode bought nothing. Claim withdrawn. medium.com/@tinholt/how-many…
2
62
At $11.65 in video-generation charges, a creative idea becomes something you can afford to explore. Chicago in Five Breaths: 51 seconds of watercolor Chicago, Belmont Harbor to Navy Pier after dark. dinand.com/side-projects/chi…
1
52
An architecture diagram starts a debate about the architecture. My second Maison Agent experiment presents the AI control plane like a fragrance ad: one gold ring gathering black ribbons, one sentence beneath it.
1
1
36