Skip to content

Blog

Notes from the studio

Notes from ilOblique on building AI products end to end.

What a buyer already knows about us before the first email — and the five questions we answer in writing↗

The decision is mostly made before anyone writes to us. A reported 95% of deals go to a vendor already on the day-one shortlist, and that list is now assembled by asking a model. So we published the screening questions we get asked, and our answers to each.

Your AI deal passed the demo. The security review is where it dies now↗

In 2026, enterprise AI deals tend to stall in the buyer's security review, not the demo. That review can hold 40 AI questions, and the standard liability form now excludes generative-AI harms. A policy document answers neither. In our view, five engineering artifacts answer both.

Approve-every-output is not a control: what our human gate was actually shown↗

A human approving every AI output is not a control. In Stanford's 51-case sample it is the lower-yield design (30% median gain vs 71% for escalation), and a gate sees only what the system renders. Ours was shown a disabled button, a fix the generator never read, and “sent” as “published”.

Do the hourly math honestly and it says hire. When a studio is still worth buying↗

Per hour, a studio costs more than your employee: $160 × 1,800 hours is $288,000, and KORE1 still counts that above a senior AI hire. Buy a studio only for a bounded build: an acceptance test before the estimate, an end date, and a handover someone on your side owns.

A language app is only as good as the week behind each language↗

One Holt language pair took seven days: 14,484 example sentences and a blind check on 120 dictionary words by a judge that did not write them — 96.7% pass, 0.0% substantive defects. The same report lists what learners in our older languages still see today.

Your language app said “correct.” It scheduled the wrong word.↗

An app that accepts a near-miss answer as correct breaks your review schedule, not the lesson. We found 9 accent minimal pairs in two live Holt languages doing exactly that, named it in our own report, and replaced the hand-kept list with a test that scans the whole word bank.

Rewrite or rescue? Don't buy the verdict — buy the reasons↗

Founders call for help when credits, patience or confidence run out, and ask “rewrite or rescue?” — a verdict they can't check. Don't pay for the verdict. Before any price, get the reasons in writing: repo access, test coverage, what was read, what stays and why.

Sent is not published: eight handovers nothing had read back↗

“Sent” is a fact about your request. “Published” is a fact about someone else’s system. Only a read-back turns one into the other. The first time we checked, eight handovers had none: seven were live, and the eighth had failed authentication on a dead connector that told us nothing.

The “I cannot” your agent is not allowed to say, and the 88 credits it cost us↗

We think an agent’s honest “I cannot” is a product feature, not a model weakness. In one of our pipelines, forbidding it cost 3 restarts, 17 runs and 88 credits for zero accepted results. What fixed it for us was structure, not a friendlier prompt.

A word list is not a dictionary: we recounted Holt's with two keys↗

Counting by the label a human reads understated contamination in Holt's dictionary by two to three times. With a second independent key the honest figure came back at about 18% and 14%. After the merge the learner-visible dictionary was unchanged.

A product that learns from you doesn't break loudly — it goes quiet↗

Kindling's argument is that every edit teaches it. For a stretch of posts it learned nothing: a column from an unapplied migration made the read fail, and nothing said so. Its scoring had gone uniform too — an active account got the same boost as an empty one. Here is what catches that.