What’s new? v4.90 — The assistant stops citing itself and says “I don’t have that”

One in five “sources” were documents the assistant had written itself. Now they arrive marked, and when a project lacks data the assistant says so instead of inventing it. Plus, the Knowledge library can finally be filled from the screen.

v4.90.1 — The library you couldn’t fill

The Knowledge library is where the firm keeps what it works with and that belongs to no case file: the pitch deck, the methodology, the material for a partnership. The assistant consults it when drafting proposals and scripts.

It had a basic problem: to add a document you had to paste its internal identifier, a 36-character code that appears nowhere on screen. Someone wanted to add a partnership document to the “Alliance” collection and couldn’t find how — rightly so.

Now you search for the document by name and add it with one click, or upload the file right there: it lands in Documents, encrypted and versioned like any other, and inside the collection in a single step.

v4.90.0 — The assistant stops citing itself, and learns to say “I don’t have that”

Asked about the deduction cap for a car, a firm’s assistant answered with the article, the section and the exact amounts. Correct, except for one thing: the source it cited was a document it had written itself a month earlier — the result of a web search it had triggered during a test, saved in the case file like any other document.

This release closes that failure, and it was more common than it looks. Measured on real installations: 196 of 902 documents are the assistant’s own output — research it triggered, analyses it drafted — and all of them reached its answers labeled “Source”, looking just like a contract a person had uploaded.

Now they arrive marked, with an explicit rule: that material is for reasoning, never for grounding a fact about the company or the text of a law. The assistant may refer to it for what it is —“an earlier analysis estimated…”— instead of presenting it as evidence. Nothing is hidden: the document still arrives in full.

Saying what’s missing instead of filling it in

Asked “when will this project finish?”, the assistant gave a date; asked “how much is the delay costing us?”, it gave a figure in pesos. Both on data it didn’t have: the project declared not a single task dependency and not a single rate. It deduced the critical path from task names and invented an hourly cost. It wasn’t a slip by the model: nobody had told it that data was missing, and silence looks too much like permission.

Now it also receives what the project does not have: tasks without estimates, undeclared dependencies (and an instruction not to infer them from order or name), missing rates, missing baseline. On the cost question, the assistant went from publishing “between $86,000 and $115,000 in idle hours” to saying that without rates it can’t budget, and talking in hours and days, which are real data.

A task with no owner stops being invisible

The assistant said “Assigned to: So-and-so” when there was someone and said nothing when there wasn’t, so an unowned task read just like an assigned one. There were 153 open tasks like that in real installations. Now it names them, counts them and says who is accountable while they have no owner: the project lead. A task assigned to an external contact now shows their name too.

The same at portfolio scale, for whoever runs several projects: how many have no baseline, how many have no rates, and how many have no open task at all — which isn’t being green, it’s being uninstrumented.

What you’ll notice on upgrade: nothing is switched on, but the way the assistant answers changes. If an input is missing, it says which one instead of estimating it. You get it back by entering the data.

assistant ai case file library release

← Back to the index