Delegation Stack Failure

Status: Working note based on two customer-service escalations on August 3, 2026.

Working thesis

I increasingly default to OpenAI directly instead of relying on a specialized AI service to decide which model, reasoning mode, and context I am allowed to use.

The point is not simply that general-purpose AI is more adaptable. Going straight to the source gives me control over the model and reasoning level, lets me build and retain my own context layer, and keeps me from being locked into whatever cost-optimized or constrained model an intermediary selected behind its interface.

Duckbill AI and MedFinder AI are a good real-life example. Both were supposed to remove aggravating administrative work. When they failed, they created a second layer of work that was harder to inspect, redirect, or fix.

Case notes

• Duckbill AI: Returned a polished comparison that still left the call-only facts I needed unresolved. I had to audit the result, identify what was missing, and escalate the work the service was supposed to have completed.
• MedFinder AI: Produced unreliable availability information, directed me to a pharmacy that did not have the medication, restarted a search I had asked it to cancel, and turned the refund into another customer-service fight.

Whether or not either product is literally a thin wrapper around a foundation model is not the important distinction here. From my side, each one is another product layer that controls the interface, available context, model choice, cost tradeoffs, limits, and support system. Those choices become hardest to inspect precisely when the service fails.

What going straight to the source could do

Codex could step outside both products' intended workflows and deal with the failure itself. It:

• Read the histories across Duckbill, MedFinder web chat, and Messages.
• Reconstructed the specific failures and prior commitments.
• Wrote and sent firm escalation messages.
• Kept the requests concrete: owner, deadline, written confirmation, refund reference, and next action.
• Split the follow-up between agents so Duckbill and MedFinder could be handled in parallel.

The specialized services could only operate inside their own systems. Using Codex through OpenAI let me work across the systems, supply the context they were missing, inspect the mismatch between the promise and the result, and change tactics when the happy path failed.

Why the abstraction becomes a liability

The abstraction is useful only while it reliably removes work. Once it fails, I am dealing with:

1. The original task.
2. The product's hidden assumptions and limitations.
3. The customer-service process required to dispute the result.

At that point I would rather go straight to OpenAI. I can choose the model and reasoning level, own the context layer, inspect what the system is doing, redirect it, and use it across whatever interface the problem moves into.

Current boundary

Codex still cannot make phone calls. That sucks. Some of the most important answers are only available by calling a person, so the workflow still stops where a phone call becomes necessary.

Working operating model

• Specialized AI service: Use only when it provides a real execution capability, reliable results, and enough visibility to verify the work.
• TaskRabbit: One-off physical or local tasks.
• Mother's helper: Recurring household support that requires a person to be present.
• Me: Priorities, judgment, approvals, and decisions.

Provisional takeaway

I do not need a service to put its interface, context limits, and model-cost decisions between me and the source. An abstraction has to earn its place by providing a real capability I cannot get directly—and by continuing to reduce work when something goes wrong.

Dataset Preview

• Raw CSV row/table content is available in the source artifact.