Meet the engagement

The study's authors describe two patterns among consultants who navigated the frontier well. Centaurs divide the work: discrete subtasks go to the AI (draft this, summarize that), and the human keeps every decision point. Cyborgs weave AI through the whole workflow continuously. Both work; what fails is the third, unnamed pattern - pasting the task in and shipping whatever comes back.

Two second-order findings matter as much as the headline. First, the bottom half gained most (+43% vs +17%): AI is a floor-raiser, which means baseline production stops differentiating anyone. Second, output homogenized - AI-assisted answers converged on similar ideas. Put those together and the conclusion writes itself: the parts of consulting AI does well are becoming table stakes, and the parts it cannot do - judgment, framing, client specificity, accountability - are where careers now compound. The scorer you meet in session 5 measures exactly those.

The outside-frontier task was chosen precisely because AI would get it subtly wrong - the researchers built a trap on purpose, the way session 5's scorer builds one. The honest reading is not "AI fails 19% of the time"; it is "on tasks beyond its capability, AI-assisted professionals did WORSE than unassisted ones, because fluent wrong answers are more persuasive than no answer." Every task you delegate this course, you will first ask: is this inside the frontier, and how would I know if it were not? That question, asked habitually, is worth more than any prompt technique.

In spring 2023, Samsung permitted ChatGPT internally; within about twenty days, employees had pasted confidential source code and an internal meeting transcript into it, and by May the company had banned generative AI on work devices. The consulting version of that incident is worse - the leaked material belongs to your CLIENT, and the trust you burn is the product you sell.

The working rule has an architectural half and a behavioral half. Architectural: consumer AI tiers may train on your conversations unless you opt out - the vendors say so themselves - while enterprise and API tiers of the major vendors do not train on business data by default, contractually. Firms buy enterprise tools precisely to make the guarantee contractual instead of behavioral. Behavioral: even inside approved tools, client-identifying material gets anonymized before it enters a prompt. Session b8 makes you practice both until they are reflexes.

In October 2025, Deloitte Australia partially refunded the Australian government for a A$439,000 report after an academic found fabricated references and a misattributed court quote - errors Deloitte attributed to generative AI use it later disclosed. The refund was modest; the headline was not. Every consulting firm on earth now has a version of the same clause: a human is accountable for every citation, every quote, every number that reaches a client.

The consulting cousin of the fake legal case (which learn-ai-law covers in depth) is the invented statistic - plausible, well-formatted, and nonexistent. Session b3 teaches the source-log discipline that catches it; session b9 makes the citation audit a standing QA gate. For now, one habit: an AI-supplied fact is a lead, not a fact, until you have seen its source with your own eyes.

Cadence's data room, distilled to the numbers the whole course reuses:

One more thing came with the data room: a note about a large financial-services prospect in late-stage talks, with names and a contract figure attached. You will meet that note again in session b8 - it is the confidentiality drill, and the lab's gate already knows its contents.