Newsletter
6 min read · Issue 01 · July 2026 · Yaroslav Stus, Leebry
Welcome to the first issue of Work in Progress.
A biweekly newsletter for IT leaders who've been handed an AI mandate, no budget for it, and a deadline that somehow pre-dates the meeting where it was assigned.
We're the team building Leebry at MacPaw 🇺🇦
Most writing about AI in the workplace is either convinced the hard part is already over, or it's written by people who have never had to revoke access for someone who left two weeks ago but somehow still owns half the shared drive.
We wanted something in between: the view from the build side, written for the people on the buying side, with no one pretending the work is finished.
And yes, we're a vendor. That should probably disqualify us from writing this newsletter, but we think it's the reason to write it.
Starting with lessons we learned this sprint from what we shipped.
by Yaroslav Stus, Associate Director of Engineering @ Leebry
Most of what we shipped this sprint comes back to one question: when an employee asks an AI tool something, who decides what the model is allowed to see?
Get that decision wrong, and the AI confidently surfaces things people should not see. That's a buyer's biggest fear, and one of the biggest responsibilities on the build side. We have been working through this on Leebry, so that is the example I will walk through, but the principle applies to anything you might be evaluating.
The decision we made early:
we are not expecting the LLM to decide who has access and who does not.
From there, the implementation depends on what kind of system we are integrating with. There are two buckets.
Company-wide knowledge
Things like Confluence and Google Drive. We treat the content there as information that is generally available to everyone in the company, the way a public library works. If something needs to be restricted, the controls live in the source system, not in the AI layer.
Per-user systems, where access is the entire point
Jira is the clearest example. We are not indexing Jira data. We are querying it in the user's context every time. The audit log stays in the destination. You can see who made the change, who left the comment, and who created the ticket.
The principle behind both: the AI tool should not invent its own access model. It either treats a source as shared, or it inherits the per-user rules already set in the source system.
A couple of things worth doing if you are evaluating any AI tool right now.
Use enterprise-level access for whichever LLM is underneath the product
OpenAI, Anthropic, and Gemini all offer this, and it is the only way to guarantee your data is not used for training or exposed to other users.
Do not put PII into LLMs
Mistakes happen everywhere, and any incident there generates a lot of effect.
Permissions are not the interesting part of an AI tool. They are the part that determines whether everything else works.
Last month, OpenAI launched a $4 billion deployment company with 150 forward-deployed engineers, backed by McKinsey, Bain, and most of Wall Street, designed to "help organizations redesign their critical workflows" around AI. The pitch: enterprises are ready; the only problem is execution speed.
Read that pitch from the IT side of the building, and it falls apart.
The reason why enterprise AI moves slowly is rarely that there are too few engineers in the room. It's that the executive who set the strategy and the team executing it are working from different briefings.
We surveyed 100 IT directors for our AI at Work Report.
61%
of IT directors surveyed say leadership expectations are already running ahead of what their teams can realistically deliver
100
IT directors surveyed for the Leebry AI at Work 2026 report
No number of forward-deployed engineers fixes that, because forward-deployed engineers do not sit in the boardroom meeting where the AI roadmap got oversold.
We built a quiz off the back of our "AI at Work" report. Seven short scenarios which give you a maturity score and a sense of where your organization sits against the 100 IT directors we surveyed.
The next time your CEO asks why the AI initiative is not moving faster, the answer is in the breakdown. Or share it with your IT peers to see who actually has it together.
Take the quiz: leebry.com/ai-assessment/
That's the first edition.
If something here resonated (or if you think we're completely wrong) reply and tell us. Half the reason we're writing this is to compare notes with people doing the same work.
— The Leebry Team
Subscribe to get upcoming episodes
A biweekly field report on enterprise AI deployment, written for IT leaders by the people doing the work.
More editions
In One Document Apples are Vegetables. In Another, Fruits.
Issue 02 · 6 min read
Personalization Without Transparency is Just Surveillance
Issue 03 · 5 min read
