Enterprise AI: what shipped

A GPT-4 assistant that answers advisor questions against the firm's own research library, rolled out to roughly 16,000 advisors from 2023. A second tool, Debrief, sits in client meetings with consent and drafts the notes, action items, and follow-up emails into the CRM.

An evaluation suite that gated every rollout, built with the model vendor before launch. Per the vendor's case study, retrieval coverage of the document corpus went from 20 to 80 percent, and adoption reached 98 percent of advisor teams. Those figures are the company's and the vendor's own. Treat them as their claim.

Nothing on the public record so far. The design choice worth copying: the assistant retrieves and drafts, the advisor sends. The human stayed on every output that touches a client.

An in-house portal giving employees access to large language models inside the bank's own security and compliance controls, for drafting, research, and analysis. From zero to 200,000 onboarded employees in eight months across 2024 and 2025.

Building where the data controls already lived, and treating adoption as something to earn: the rollout was optional first and woven into existing workflows rather than mandated. It also took a bank-sized technology organization, which is the honest footnote. MIT's review found externally bought tools reach production about twice as often as internal builds. JPMorgan is what the exception costs.

Nothing on the public record so far. The bet is scale over specialization: one governed doorway to models instead of hundreds of scattered pilots.

An OpenAI-built customer service assistant that handled 2.3 million conversations in its first month of 2024, two thirds of all support chats, and cut resolution time from 11 minutes to under 2. The company projected a $40 million profit improvement for the year. Those are Klarna's own figures, from its own press release.

Quality, slowly. By May 2025 Klarna was rehiring human agents for disputes, complex refunds, and hardship cases, with the CEO conceding publicly that cost-driven automation had produced lower quality service. The model that survived is hybrid: the assistant handles the routine volume, a human takes over where the cost of an error rises.

The launch press release and the correction are the same case study, fourteen months apart. Any vendor citing Klarna's month-one numbers without the 2025 sequel is telling you half a story, which is a useful thing to know about the vendor.

Automated order taking, built with IBM and tested from 2021 at more than 100 U.S. restaurants. A real production deployment, at scale, facing customers in the open air.

Accuracy in uncontrolled conditions. Background noise, accents, and overlapping voices produced order errors that staff had to redo, and that customers filmed. The test ended in June 2024 and the technology came out of every restaurant by the end of July.

Ambient, customer-facing audio is a different problem from a demo booth, and two years of production testing is what it took to establish that. The failure is more informative than most success stories: it puts a boundary on the map.

A website chatbot answering policy questions. In 2022 it invented a bereavement refund policy that contradicted the airline's actual terms, and a customer booked on the strength of the answer.

The legal theory that the bot's words were not the company's. British Columbia's Civil Resolution Tribunal ruled in February 2024 that Air Canada was responsible for everything on its website, chatbot included, rejecting the argument that the bot was a separate entity. The award was small, CA$812. The precedent was not.

Your AI's output is your company speaking. Every customer-facing deployment on this page now operates under that ruling, which is why the question of what happens when the system is wrong belongs in the first vendor meeting, not the post-mortem.