Internal AI Knowledge Assistant for a Services Firm

A B2B professional-services firm had two decades of project knowledge locked in scattered drives and a handful of senior engineers’ heads. New hires took months to ramp, and the same questions reached the same three experts every week.

After a deliberate data-cleanup effort, I stood up a retrieval-augmented assistant connected to the document repository, so any employee could ask a natural-language question and get sourced answers from past projects, policies, and standards. New-hire ramp shortened and escalations to senior staff dropped. The magnitudes below are independent research benchmarks for comparable assistants, not this client’s measured results.

The challenge

The firm’s most valuable asset, its accumulated experience, was also its least accessible.

• Specs, lessons learned, and scoping templates lived in scattered drives and in senior engineers’ heads.
• New hires took months to become productive, with no fast way to learn from past work.
• The same questions hit the same handful of experts every week, taxing exactly the people whose time was scarcest.

Knowledge that lives in people’s heads is a single point of failure. It is unavailable when those people are busy, and it leaves the building when they do.

The approach

I treated trust as the real requirement: the assistant would only get used if its answers were reliable, which meant fixing the data before building the search.

I: Implementation Planning

The knowledge was real but messy, so I planned a deliberate data-cleanup effort first. Then I designed a retrieval-augmented assistant over the document repository that answers in natural language and attaches its sources, so employees could both get an answer and see where it came from.

M: Migration & Execution

After the cleanup, I stood up the retrieval-augmented assistant connected to the repository, letting any employee ask natural-language questions and get sourced answers from past projects, policies, and standards. The data-hygiene groundwork made the answers reliable enough that the team actually relied on them.

The results

Faster new-hire ramp. New employees could draw on two decades of project knowledge from their first week, instead of waiting months to absorb it.

Fewer escalations to senior experts. Routine questions resolved through the assistant, freeing the firm’s most experienced people, and institutional knowledge became searchable instead of tribal.

For context, independent research on AI knowledge assistants finds new hires reaching the performance of staff with roughly 3× their tenure, and about 20–25% fewer escalations to senior experts (Brynjolfsson, Li & Raymond, Quarterly Journal of Economics, 2025), against a baseline where knowledge workers spend roughly a day a week searching for information (McKinsey Global Institute). Those figures are industry benchmarks, not this client’s measured results.

Why this matters

Institutional knowledge that lives in people’s heads is a liability dressed up as an asset. It walks out at the end of the day, it is unavailable when the expert is busy, and it retires for good when they do.

Making that knowledge searchable, with sources attached, turns tribal know-how into something every new hire can draw on from day one. The unglamorous prerequisite is data hygiene: clean inputs are what make the answers trustworthy enough that people actually use them instead of interrupting the expert down the hall.