Company AI brain case studies: what the firms with results did first

A company AI brain works best when one assistant answers from one set of company knowledge that someone wrote down and keeps current. Steno counted its staff's questions before building anything, and Morgan Stanley had advisors grade answers before rollout. Most result numbers in these stories come from the company or its software vendor, so this page says who measured each one.

Seven company AI brain case studies and who measured them

Morgan Stanley¹, Steno², BCI³, Nexio Projects⁴, Super.com⁵ and Wealthsimple⁶ each published results on their own page or their vendor's. Most are vendor customer stories written to sell the product, and few give a baseline or a time period. The UK government's pilot is the one independent evaluation in the set.

None of these stories says what the project cost. Some contradict themselves: Guru's page says its agent handles 100% of Steno's daily questions, though the audit in the same story found about half needed a person².

What the AI brains that worked did first

Side by side, the cases share one order of work. Each step shows up in at least one story, though no single story documents all five.

Steno's audit is the easiest step to copy. Christian George, an associate product manager, scanned the Slack channel where videographers posted about 200 questions a day and found about half had a documented answer². He set 50% as the target for cutting tickets and configured the agent himself in under a week. Tickets then fell about 48%².

Morgan Stanley documents the grading step best. OpenAI's case study says advisors and prompt engineers graded the assistant's answers for accuracy and coherence before deployment⁷. It started out able to answer 7,000 questions and now draws on 100,000 documents, and the firm tests it daily against a fixed set of sample questions⁷.

Both firms started with one group: Steno with one Slack channel², Morgan Stanley with financial advisors¹. BCI gave Microsoft 365 Copilot to 300 employees first, then rotated licences across more than 570 users³.

An AI knowledge base for business only knows what's written down

An AI brain finds the documents that best match a question and writes its answer from them, so each answer is only as good as the page behind it. How AI knowledge assistants work covers the mechanics. Try your own numbers below.

Nexio Projects gave staff one entry point that answers questions on contracts, policies and leave from live HR data, and says about 95% of everyday HR questions now resolve without a colleague⁴. Super.com asked staff to write down what had mostly been shared in meetings before it connected its tools to Glean⁵.

Microsoft Copilot case study: what BCI and the UK government found

BCI, the British Columbia Investment Management Corporation, is the closest Canadian example for a Microsoft 365 firm. In an internal survey, 84% of respondents reported a 10% to 20% productivity gain³. Its clearest result came from a narrower tool: a custom Azure OpenAI assistant that drafts sections of internal audit reports, which BCI says cut the writing effort by 30%³.

The UK Department for Business and Trade ran a three-month Copilot pilot from October to December 2024 and published its own evaluation in August 2025⁸.

Drafting and summarizing saved the most time, while scheduling and image tasks took longer with Copilot⁸. Evaluators found no evidence that the time saved raised productivity, and the diary study had a 32% response rate⁸.

Adoption is easier to publish than results

Morgan Stanley's 98% counts advisor teams that adopted the assistant, and the release gives no figure for time or money saved¹. Thomson Reuters' CIO wrote in August 2026 that 70% of its 26,000 employees used AI tools consistently, and called the next job "narrowing the gap between adoption and value"⁹.

PMP Strategy, a 150-person consultancy, says its consultants built more than 500 agents in six months, and its managing partner expects to throw away 90% of them¹⁰. The firm-wide library held 17 agents against a target of 50¹⁰. Nexio keeps a register of its 100+ agents and checks new ones for duplicates⁴.

We found no small or mid-size North American firm that has published AI brain results with a baseline. The controlled studies in what the productivity research shows are a better guide to what a pilot might return.

How to start an internal AI assistant at a smaller firm

1. Pick one group with steady repeat questions and the one channel where they already ask, as Steno did².
2. Read a sample of those questions and mark which have a written, current answer. That share is your realistic target².
3. Write the missing answers and name an owner for each source document, the way Nexio's AI policy names an owner for every output⁴.
4. Turn real questions with known answers into a test set and rerun it after each change, as Morgan Stanley does daily⁷.
5. Launch to that one group and measure the same thing before and after, such as tickets or repeat questions.
6. Price seats once the answers hold up. Microsoft lists the Microsoft 365 Copilot Business add-on at US$21 per user a month on an annual plan, or US$5,040 a year for 20 people¹¹.

Steno's software took under a week², so plan most of the hours for writing and testing answers and for teaching staff to check what the assistant says. Why staff training decides the result covers that part, and the readiness check shows where your firm stands.