ilOblique

[ilOblique @ilOblique]

Your AI pilot didn't measure your business. It measured who you onboarded first.

In Jeff Wahlberg's teardown, a $447/yr pilot estimate became ~$9,100 on a random sample. Same prices, same code.

Before year two: random sample, then median, p90, p99, max.

#FinOps #AIagents

[ilOblique @ilOblique]

One number to pull on Monday: of the PRs your agents opened this quarter, how many merged, and who owned each one.

If the answer is "nobody in particular", start there.

The full breakdown: https://t.co/T3QX2Mrjbo

#AgenticCoding #CodeReview

[ilOblique @ilOblique]

At the top 10% of engineering orgs, AI writes 45% of merged lines.

Autonomous agents open 4.7% of the pull requests.

Same teams, same year. The gap between those two numbers is the part nobody is pricing.

[ilOblique @ilOblique]

That decision drops a line item. The license goes away. The owner does not.

Someone still has to review, merge, patch and answer for that code for as long as it runs. We think that if nobody is named, you did not save the cost. You moved it.

[ilOblique @ilOblique]

Then weight ratings by the evidence behind them. We think a thumbs-down with a manual edit is the strongest label you have: the user showed you the right answer.

#LLMevals #ProductAnalytics

[ilOblique @ilOblique]

F-Secure's survey (September 2026) shows the same from the other side:

70.1% almost never check AI answers in practice.
47.3% drop their doubts when an answer sounds smooth and confident.
35.3% "check" by asking another chatbot.

[ilOblique @ilOblique]

Our bet: the "nothing" bucket rates highest. If so, your headline satisfaction number is mostly that bucket.

If verified users rate just as high, we're wrong and your ratings are a real signal. That's worth knowing too.

[ilOblique @ilOblique]

A 2026 study in Technology in Society surveyed 544 generative-AI users: the more they trust the system, the more willing they are to give it feedback — and the less likely they are to verify its output.

The authors call it "feedback without verification."

[ilOblique @ilOblique]

We think the users who rate your AI feature most are the ones who checked it least.

A thumbs-up with no edit, no copy, no source click is a weak label. Train on it and your feedback loop learns its own mistakes.

[ilOblique @ilOblique]

The longer version, what became portable and what stayed on you:
https://t.co/brweJmFu3j

#AgentSkills #VendorLockIn

[ilOblique @ilOblique]

In our experience, the vendor-portability plan is: copy the system prompt into another vendor's chat window and hope.

Today one layer of that plan got a real answer. Anthropic is handing the Agent Skills standard to a neutral owner.

[ilOblique @ilOblique]

One check for this week: list the skills installed in your agents. Whose are they? What can each one run? Who on your team actually read them?

What an agent knows now travels between vendors. What it's allowed to do is still yours to review.