Skip to content
What the field shows · Knowledge systems

Two internal knowledge systems, published in full by the companies that built them.

Both are unusually well documented, including the parts that did not work. These are their numbers, linked to their sources.

Two examples, both halves

The published limits are more instructive than the headline figures.

IBM AskHR

IBM's internal HR assistant handles routine employee and manager requests: letters, vacation, payroll access, compensation changes. Around 80 automated tasks sit behind it.

94%

of common HR questions answered without escalation, IBM's own containment figure. IBM also reports more than 11.5 million employee interactions in 2024 and a roughly 40% lower HR operating cost over four years.

IBM case study, AskHR

What happened next

IBM announced it would triple US entry-level hiring in 2026. An Inc. opinion piece traces this to the remaining 6%: the exceptions and judgement calls the system does not take on.

Inc. opinion piece, July 2026

Uber Genie

Genie is a Slack assistant for on-call engineering questions, built on a retrieval architecture Uber describes in detail. Uber's internal support channels carry around 45,000 questions a month.

13,000h

of engineering time saved since launch in September 2023, across more than 70,000 questions in 154 Slack channels.

Uber Engineering Blog

What happened next

Uber also published the helpfulness rate: 48.9%. Fewer than half the answers were rated helpful, and Genie is described throughout as an on-call copilot rather than a replacement.

Uber Engineering Blog, same post

How we read that

A knowledge system is worth building when it can say what it does not know.

01

The source has to be reachable

Both systems index material a person can open and check. An answer that cannot point at its paragraph is not usable in a regulated setting.
02

Half-useful is still useful

Uber shipped at a 48.9% helpfulness rate because the alternative was a colleague waiting. The bar is the current process, not perfection.
03

Access follows role

IBM's system separates what an employee, a manager and an HR partner may see. Without that split, a knowledge system becomes a data-protection problem.

What we do not promise

  • That retrieval finds everything. Where it finds nothing, the system should say so rather than fill the gap.
  • That the knowledge maintains itself. Retrieval searches; it does not curate.
  • That every answer is helpful. The published rate at Uber was 48.9%, and that was still worth running.

If the knowledge exists but nobody can find it in time, that is a build we know.