There is one metric I trust more than any enthusiastic reviews of AI — retention. Everyone shows up to try a new toy on day one; the whole question is who comes back in week two, when the novelty wears off and only real usefulness remains. Well, our digital assistant’s weekly-to-monthly audience ratio holds at about 90%: nine out of ten people who used it this month come back every week. In all my years of watching internal corporate tools, I have never seen a number like that.
For a sense of scale, look at Microsoft Copilot — the most widely deployed enterprise AI tool in the world. Of all the employees whose companies bought them access, roughly one in three uses it regularly, and that is not a Microsoft failure but the norm of enterprise rollouts: access is issued by decree, and retention cannot be. So when an internal tool holds 90% WAU/MAU without a single “please use the bot,” I know we have hit something real — and below are the numbers from two months and the main conclusion about why the assistant took root where classical automation is powerless.
The numbers, unvarnished
We launched on May 29, straight to the whole company, with no focus group and no gradual rollout. In two months 185 employees used the assistant; in July the actives ran at about 170 a month with 140–156 a week, and dialogues since launch have accumulated to 10,500. The median for a full answer is about 13 seconds, and those seconds include document search and tool calls.
The most telling curve is calls per week: 1,090 at the start of the month against 2,545 at the end — 2.3× growth. Let me stress the part I am genuinely proud of: without a single “please use the bot,” without usage KPIs, without HR emails. We connected it once — after that the audience grew on its own, because colleagues showed each other their screens.
And here is the curve I did not expect at all — autopilot. 138 employees have accumulated 165 active schedules: morning briefings, mail digests, reminders — more than a hundred background runs a day. The assistant started working for people before they asked it for anything, and that shift from “asked — answered” to “brought it unprompted” seems to me more important than any audience figures.
What people actually do
The top operations of the month will sober up anyone expecting miracles from enterprise AI: document feed and search — 7.5 thousand calls, day plan — 3.5 thousand, emails awaiting reply — 3.2 thousand, mail digest — 3.1 thousand, followed by knowledge base search, meeting rooms, calendar events and finding colleagues in the org structure.
Sounds boring? Exactly — and that is the whole point. Classical automation is powerless before such tasks: you cannot write a workflow for the request “find that contract Ivanov wrote about in March.” Yet every employee racks up a hundred or so of these micro-tasks a week, and together they devour hours — it is just that nobody ever got around to automating them, because each one taken separately is too small. The assistant closes precisely this long tail of small context switches between systems that automation never had hands for.
If we conservatively assume the assistant saves an employee 15 minutes a day — honestly, an expert estimate, we ran no time study, but the structure of the requests supports it — then across 170 active users that is about 42 person-hours a day. Five full positions returned to the company, running on inference on our own hardware.
Why it works — three decisions
Behind the retention stand three architectural decisions, and I am convinced the story would have gone differently without any one of them.
First: the assistant’s rights equal the employee’s rights. The agent enters the system on the person’s behalf, every ACL is enforced on every call, and no service superuser exists — even the agent cannot see a document that is “not yours.” Without this, a corporate assistant in a regulated company is impossible in principle, whatever the presentations promise.
Second: two phases for everything that changes the world. Creating a document, booking a meeting room, sending an email — first a draft, then the person’s explicit “yes,” and only then the action. The assistant does nothing irreversible without confirmation.
And third: when it can’t find something, it says “not found.” Answers from documents and the knowledge base come with links to the source, and hallucinating in place of an answer is forbidden by the architecture. A boring rule — but it is what trust rests on, and without trust there is no 90% retention.
All of this runs in a closed loop on our own cluster — around one and a half billion tokens a month, and not a single request leaves the perimeter.
Key takeaways
- Watch retention, not first-week enthusiasm: WAU/MAU is the detector of real usefulness.
- The value of a corporate assistant lies in the long tail of small routine, not in showcase scenarios.
- Rights equal to the user’s, two-phase actions, “not found” instead of invention — three conditions you cannot do without.
- The best adoption driver is a colleague showing their screen. Not a decree.
The data is from our own operations, May–July 2026. The personal assistant is also the first thing that greets you in the live d8n demo: ask it anything.
