Applied AI research portfolio · 7 cases

Teaching AI agents to decide like their owner

How can an AI agent learn its owner's way of deciding from approvals and corrections, and know when to ask?

Error

The agent decided differently from how the owner would have decided.

Questions

How many times the agent had to stop and ask the owner.

The decision loop: the owner's situation shapes his profile, the agent predicts his decision, acts when confident and asks when not, and every answer or correction updates the profile. OWNER'S SITUATION Cash · time · skills Family · business · reputation Limits and strengths DECISION PROFILE One model per type of decision P(owner approves | context) Confident enough? ACT yes no ASK THE OWNER answer = new label owner's correction, if it was wrong

Seven businesses, one question

Each case is a real operation I run. Each has one main finding.

Why it matters · productivity

What if the agents never stopped to wait for me?

Today every agent stops at my approval, and I can only look at one business at a time. When I spent a week on one project, the others stood still. An agent that decides like me only asks when it is unsure, so the same few minutes of my attention clear many more decisions.

27%

27% is what the per-type profile measured in case 01, at 85% confidence. Each output below is multiplied by 3.7×.

SEO, real decision log: 9 of 24 changes ended the first week waiting for me. Over 12 weeks, days lost waiting fall from 1,260 to 356 with reversibility rules that travel across sites.
Press: a weekly approval batch loses 35% of the reach of pieces I would approve. When the agent publishes the 20% it is sure about, the loss drops to 12%.
Agents: at 27% questions, the agent made 0.24 decisions per 100 that I would have made differently. Mistakes grow when it asks less: move the control to see the cost.

How this is computed: output with the agent = output today × (100 ÷ % of decisions it still asks me), assuming my review time stays the same and the agent's own capacity is not the limit. The question and error rates come from the case 01 simulation. Numbers marked assumption are placeholders until I enter my real weekly figures.

Case 01 · Agents · the framework

One owner, many types of decision, and a specifier in between

My businesses run on LLM agents and a few people. When I gave the agent rules, it followed them with me. When another person used the system, some rules held and others broke. The real object was never the rules. It is how I decide, which depends on my situation and changes with the type of decision.

Fixed rules I wrote once
24.5
errors per 100 decisions
Never asks: 0 questions
One learned profile for everything
1.1
errors per 100 decisions
Asks 57 times per 100
One profile per type of decision
0.3
errors per 100 decisions
Asks 27 times per 100

5 simulation runs × 120 days × 40 decisions a day. The agent asks when its confidence is below 85%.

The same owner weighs the same facts differently

Each row is a type of decision. Each column is a fact about it. Cash is tight, so the amount weighs heavily against spending. Reputation matters, so brand risk weighs against publishing and writing to clients. Production changes must be reversible. A single profile averages these rows and gets every one of them a little wrong.

Weight in the owner's decision: blue pushes toward approve, red toward reject.

Try it: how sure should the agent be before acting?

Move the control. A stricter agent makes fewer errors but asks you more. At every setting, one profile per type gets both numbers lower.

85%
One profile
1.10 errors · 56 questions per 100
Profile per type
0.24 errors · 27 questions per 100

Each point is one confidence setting, 3 simulation runs each. Lower and further left is better. The error axis is logarithmic.

When the owner's life changes, the agent has to follow

On day 60 the owner pays off his debts, and spending stops being so sensitive to the amount. Agents that learned the old owner start asking again. The one that forgets old decisions faster keeps its errors lowest, and pays for it in questions.

Questions to the owner per day, out of 40 decisions. 7-day average.

A specifier answers for the owner. Whose profile does the agent learn?

From day 60, a specifier answers half of the requests. Her situation is different: more relaxed about money and brand risk. If the agent learns from whoever approves, it slowly becomes her. If her answer is kept as a comment, and the owner is asked only when they disagree, it stays the owner's agent.

Decisions different from the owner's, per 100. 7-day average, 5 runs.

Findings

1

One profile per type of decision makes about a third of the errors of a single profile (0.3 vs 1.1 per 100) and asks half as often (27 vs 57 per 100). The single profile fails where the owner's situation weighs most: spending and deploying.

2

An owner who changes has a price. The agent that forgets old decisions keeps errors lowest after the change (0.1 per 100) but asks more (36 vs 27 per 100).

3

If the agent learns from whoever approves, the specifier's profile silently replaces the owner's: 8.5 vs 1.1 wrong decisions per 100. The rules were never lost. The source of the labels changed.

Method, assumptions and limits

Each decision has a type k (sell, spend, publish, client message, deploy) and a context x: amount, reversibility, novelty, brand risk, urgency and quality of evidence. The owner approves when

owner(x, k) = 1[ wk · x + bk + ε > 0 ], ε ~ N(0, 0.35)

where wk, bk is his hidden profile for that type and ε is his own inconsistency. The agent fits logistic models online, refits every day, acts when max(p, 1 − p) is above the threshold, and asks otherwise. Every answer, and every later correction of a wrong action, becomes a training label. Forgetting uses a 15-day half-life on sample weights.

Limits: a linear owner, synthetic decisions, a fixed threshold for asking (a learned rejector is the next step), and an owner who is always available to answer.

References
  1. Hadfield-Menell, Dragan, Abbeel & Russell (2016). Cooperative Inverse Reinforcement Learning. NeurIPS.
  2. Poddar, Wan, Ivison, Gupta & Jaques (2024). Personalizing RLHF with Variational Preference Learning. NeurIPS.
  3. Kirk et al. (2024). The PRISM Alignment Dataset. NeurIPS Datasets and Benchmarks.
  4. Park et al. (2024). Generative Agent Simulations of 1,000 People. arXiv:2411.10109.
  5. Mozannar & Sontag (2020). Consistent Estimators for Learning to Defer to an Expert. ICML.
  6. Madras, Pitassi & Zemel (2018). Predict Responsibly: Learning to Defer. NeurIPS.
  7. Simon (1955). A behavioral model of rational choice. Quarterly Journal of Economics.
  8. Mullainathan & Shafir (2013). Scarcity: Why Having Too Little Means So Much.
  9. Gama et al. (2014). A survey on concept drift adaptation. ACM Computing Surveys.

Code and full notebook: 01_owner_decision_profile.ipynb

Case 02 · Equilibra

Copying the owner's money moves fails for people in other situations

$8,800
misplaced per user when the agent copies the owner's moves

Equilibra helps users study their portfolio with their own AI. The owner's rule is simple: pay off every card first, keep a stable cushion, and put at risk only what a fixed deposit would pay in a year. Copying his actions breaks for users with different debts and income. Learning the rule over the situation, with 13 well-chosen questions to the owner, decides for 1,000 different users as he would.

Code and full notebook: 02_equilibra_situation_rules.ipynb

Case 03 · Press

A correction with a reason teaches more than many approvals

4× faster
learning when the owner explains why, instead of only yes or no

Press is the content agent for social networks. Waiting for weekly approval batches loses about 35% of the reach of pieces the owner would have approved. Letting the agent publish alone when it is confident cuts the loss to 12%, with 13 wrong posts out of 3,000. In our real posts, the Spanish version holds viewers about twice as long at 2 seconds in 8 of 8 pairs.

Code and full notebook: 03_press_corrections_vs_approvals.ipynb

Case 04 · SEO

Autonomy for reversible changes, and rules that travel across sites

360 → 126
questions to the owner in 12 weeks

Built on the SEO agent's real decision log: 7 of 10 documents were done alone, while 3 of 5 staging changes and 2 of 3 production changes waited for the owner. Letting only reversible changes run alone cuts the questions. A correction learned on one site travels to all of them, with one question about its scope, so repeated rules fall from 70 to 6 to 0. Treating every correction as global would wrongly block about 21 changes.

Code and full notebook: 04_seo_reversibility_rule_transfer.ipynb

Case 05 · Vending

The owner's limits beat the textbook optimum

$83 → $155
weekly margin after travel

Restocking vending machines looks like a forecasting problem. The textbook plan ignored the owner's time and driving distance. His own rules include them and nearly double the weekly margin after travel. Learning each correction once takes 12 corrections instead of 150. One rule alone, skipping the clinic on low-sales weeks, adds $47 a week and saves about $77 in driving.

Code and full notebook: 05_vending_owner_rules_restock.ipynb

Case 06 · GrabGo

Cheaper checkout by changing the shelf, not the camera

−90%
money lost to recognition errors

GrabGo turns any space into a store that charges from images: the shelf camera, or a photo the customer sends by WhatsApp. Giving look-alike products the same price cuts the money lost to recognition errors by about 90% with no new hardware. Photo-only checkout misses about $160 per 1,000 carts. The owner's privacy and cost weights choose a different setup per location: photo-only at the dealership and the hotel, camera plus photo at a 200-person factory.

Code and full notebook: 06_grabgo_images_owner_tradeoffs.ipynb

Case 07 · Suppliers

Phone photos plus a scale, judged with the owner's tolerance

15% → 2.5%
of lots where the agent and the owner disagree

Samples and orders from China are inspected by a person with a phone: photos, videos and the weight of each box. Photos alone miss swapped components, and about 13 lots a year would ship. A scale catches them on the first lot. One fixed tolerance rejects about 7 good lots a year; the owner's tolerance learned per product cuts disagreement to 2.5%, with about 6 questions to him a year.

Code and full notebook: 07_suppliers_phone_inspection_weight.ipynb