Why we start with one paper
This is the leader track: six sessions for partners, practice leads, and independent consultants who own the P&L, the risk, and the bench. It runs firm-level while the consultant track runs engagement-level - over there, your teams work a six-week Northbridge Advisory engagement for Cadence, an AI note-taker client. Here, session 1 does one thing: it puts the largest field experiment ever run on AI-assisted consulting work in front of you, whole. Almost every claim a vendor, a client, or a junior will make to you this year traces back to it - usually with half the numbers missing.
The discipline for the next 45 minutes: no opinions until the evidence is on the table. Then three moves - what the study found, what the two second-order findings mean for staffing and what the firm sells, and how to turn the frontier from a metaphor into a management object your practice actually governs. The six leader sessions build on each other:
- a1 (today) - the evidence base, and the frontier map as a management object.
- a2 - what AI does to pricing and the leverage business model.
- a3 - the two rails: confidentiality and liability, as firm policy.
- a4 - quality control when drafts are cheap and fast.
- a5 - the firm's AI operating model: build vs buy, knowledge as moat.
- a6 - leading the transition: the bench, the juniors, the client conversation.
The study, properly 9 min live
Dell'Acqua et al., 2023: 758 BCG consultants - about 7% of the firm's individual contributor staff - worked 18 realistic consulting tasks, randomized into a GPT-4 arm and a no-AI control. Inside the frontier of AI capability, the AI arm completed 12.2% more tasks, worked 25.1% faster, and produced quality rated 40% higher than control. Then the researchers added one task deliberately designed to sit outside AI's capability - and on that task, AI users were 19 percentage points less likely to reach the correct answer than consultants working unassisted.
LiveCentaur, cyborg, and the pattern that fails4 min▶
The authors describe two working patterns among consultants who navigated the frontier well. Centaurs divide the work: discrete subtasks go to the AI - draft this, summarize that - while the human keeps every decision point. Cyborgs weave AI through the whole workflow continuously. Both patterns work. What fails is the third, unnamed one: pasting the task in and shipping whatever comes back. On the outside-frontier task, paste-and-ship is exactly how consultants ended up 19 percentage points behind colleagues with no AI at all - the wrong answers read just as confident and just as polished as the right ones.
For a leader, the takeaway is not "train people to prompt better." It is that the returns come from a division of labor, and the losses come from its absence. Your firm's exposure is not the consultant who refuses AI; it is the one who trusts it evenly across tasks whose difficulty is anything but even.
Self-studyWhy "19 percentage points" is the number to memorize2 min▶
Say it precisely: 19 percentage points, not 19 percent - the gap between the two arms' correctness rates. And note what the researchers did: they designed that task on purpose, as a trap, because they knew clients would eventually set one by accident. The honest reading is not "AI fails sometimes." It is: on tasks beyond its capability, AI-assisted professionals did worse than unassisted ones, because a fluent wrong answer is more persuasive than no answer. That asymmetry is the entire reason this track exists.
The two second-order findings leaders miss 7 min live
The headline numbers get quoted everywhere. The two findings that actually change how you staff and what you sell sit a layer down - and most coverage skips them.
LiveThe floor rises, the output converges5 min▶
Finding one: AI is a floor-raiser. Bottom-half performers gained +43% with GPT-4; the top half gained +17%. The gap between your weakest and strongest producers narrows dramatically the moment both have the tool. Consequence for staffing: baseline production - a competent first draft, a clean summary, a serviceable deck - stops differentiating juniors. The qualities that still separate people are the ones the tool does not supply: judgment about what matters, knowing the client, owning a number.
Finding two: output homogenized. AI-assisted answers converged on similar ideas - the study found less variety across consultants who used the tool. Consequence for what the firm sells: if your deliverables and your competitor's deliverables are drafted by the same model, the draft is no longer the product. Differentiation moves to judgment, framing, and proprietary knowledge - the things a client cannot get by asking the model themselves.
Put the two together and the strategic sentence writes itself: the work AI does well is becoming table stakes, so the firm's margin must live in the work it cannot do. Every later session in this track - pricing (a2), quality control (a4), operating model (a5), the bench (a6) - is a consequence of that sentence.
Self-studyHow to read a field experiment without a stats degree3 min▶
Three words carry most of the weight. n - how many people: 758 here, and about 7% of BCG's IC staff, which is why this study outranks every 12-person survey in your inbox. Control - who they were compared against: consultants doing the same tasks without AI, which is what separates "AI users did well" from "AI caused them to do well." Effect size - how big, in plain words: not "significant" but "25.1% faster" and "19 percentage points less likely correct." When a vendor shows you a chart, ask for those three things in that order. If any answer is missing, you are reading marketing, not evidence.
The frontier as a management object 6 min live
The study's authors called the frontier jagged for a reason: AI capability is uneven across tasks in ways that do not follow intuition. It is also invisible - work from beyond the frontier looks exactly as confident as work from inside it - and it is moving, because models change under your feet. A boundary with those three properties cannot be left to individual judgment call by call. It has to be mapped, and the map has to be owned.
LiveFrontier governance: the three-column map4 min▶
A leader's job here is not prompting skill - it is frontier governance: deciding, explicitly and per task class, which work is delegated to AI with light review, which is supervised with a human checking every load-bearing fact, and which is forbidden to delegate at all. The instrument is a three-column map:
| Column | What lives there | The rule |
|---|---|---|
| INSIDE | summaries, restructures, first drafts of routine material | delegate, review before it ships |
| EDGE | analysis, anything needing facts the model must fetch | supervise - a human owns every number |
| OUTSIDE | judgment calls, client politics, commitments, the signature | forbidden to delegate, in writing |
The EDGE column is where the -19 points live: work that looks inside but is not. The consultant track builds this map individually in its first session - your version today is for the practice, and it is a standing document, not a workshop artifact. The frontier moves; the map gets a review date.
Map your practice 🗺 ★ 18 min · everyone works
No tools, no slides - a sheet of paper and honesty. By the end you have a first frontier map of your practice and the note that makes it real.
List the 8 most billed activity types in your practice. Not job titles - activities: competitor scans, interview synthesis, first-draft decks, due-diligence memos, workshop design, the readout itself.
Sort them into the three columns honestly. INSIDE / EDGE / OUTSIDE. If you catch yourself putting everything INSIDE, remember the study built its trap task precisely because that instinct is common - and wrong by 19 percentage points.
Mark every EDGE item with the last time AI failed you there. A date, or "never checked." Items marked "never checked" are not EDGE items - they are unexamined risk sitting in the delegated pile.
Pick ONE forbidden-list item and write why in a single sentence. If the sentence is convincing, it becomes policy language. If you cannot write it, the item may not belong on the forbidden list - move it and say what supervision it needs instead.
Draft the one-paragraph note to your team announcing the map exists. What is delegated, what is supervised, what is forbidden, and who owns the next review. Send it this week - an unannounced policy is an opinion.
Homework 📝 ◐ 20 minutes
- Finish and circulate the frontier map - three columns, review date, owner. Session a2 assumes it exists when we re-price the work that sits in the INSIDE column.
- Read the study's abstract and figure 1 (SSRN, "Navigating the Jagged Technological Frontier") so the four numbers are yours, not mine - you will be quoting them to clients within the month.
- Ask your two strongest juniors where AI last failed them. Their answers are frontier soundings you did not have to take yourself - and a preview of a6's question about what the bench still needs to learn.
Three questions before you go 🎯 ◐ 90 seconds
1 · The four headline numbers from the 758-consultant BCG field study are...
Both directions are the finding. Quoting only the gains is how firms end up trusting AI evenly across an uneven frontier - and the outside-frontier work reads exactly as confident as the inside-frontier work.
2 · Who gained most from GPT-4 in the study?
The floor rises and output converges, so baseline production stops differentiating juniors - and the firm's margin has to move to judgment, framing, and proprietary knowledge.
3 · Operationally, "the frontier is jagged" means...
Uneven, invisible, and moving - which is why it is a management object, not a training topic. The three-column map (INSIDE / EDGE / OUTSIDE) is the governance instrument, and the EDGE column is where the -19 points live.
What this session covered from the official universe
Certificates, videos, and graded assessments stay with the official providers - this track teaches the working core.