2 min This is the published abstract, sentence by sentence. It is the thing the room bought a ticket for. Deliver it straight.

Run the poll. Three bands, hands up. At INSEAD it split into rough thirds. Expect this room further left and say so out loud when it does: "and you are the people funding the ones who claim to be at eight."

Do not walk all eight steps. Walk one, two, five and eight. The others land by implication.

Bridge into Section 2, say it as one sentence so the room hears one argument and not three talks: "So that is the spectrum. Now the uncomfortable part. Almost everyone is at step one, and it has not moved in a year."

2 min This slide is new and it is yours, not McKinsey's. It is the hinge of the whole talk, so give it the time.

The line to land, in your own words from the plane notes: it is easier to focus on the individual, but the real value comes at the level of workflows across teams and departments, and that is much harder.

The allocator translation: a company selling you seat licences is selling on axis one. A company that required its customer to change how the work runs is on axis two. Those are different businesses with different margins and different defensibility, and they are currently priced the same.

The matrix slide that crossed the two axes has been cut. Do not try to draw it here from the stage; slide 6 makes the sparse-adoption point in McKinsey's own numbers.

Do not use org-layer language here. No rings, no "code out to the industry". That is slide 8, and the two slides only stay distinct if this one stays about who the system serves.

90 sec Ported on 15 September from the hosted INSEAD deck (insead.sambourton.ai, "05 · How they communicate"), at Sam's request, to sit between the second spectrum and the data. Wording carried over with two edits: em dashes removed per the deck rule, and "behelf" corrected to "behalf".

Read the three cards left to right, once, and let the animation do the rest. The Every story (Plus One, the ant death spiral, the reversal) was cut from this deck, so do not lean on it here; card two is the model, not the case study.

The governance line in the callout is the bridge to Section 2: who is allowed to talk to whom is an organisational question, not a capability question, which is exactly why the next four slides show nothing moving.

Bridge: "So that is the spectrum, twice, and the shape of the conversation. Now the uncomfortable part. Almost everyone is at model one, at step one, and it has not moved in a year."

2 min The dashboard slide that used to follow this has been cut, so this carries the section on its own. Let the chart sit.

The shape of the bar chart does the work. Let it sit for a beat before you speak over it.

Accuracy guards. Say "respondents", not "companies". Say "any EBIT contribution", not "positive ROI". The 37 and 6 are prose in the report, not exhibits, so if challenged, cite pages three, twelve and seventeen.

Do not say 88 / 39 / 6. That is the November 2025 edition and it is dead.

The definition of a high performer, and say it out loud because the bar is modest. McKinsey, p3: "AI high performers (those who attribute at least 5 percent of EBIT to their use of AI and describe the technology's impact as 'significant')". That is 92 respondents out of 1,719, flat on 2025. Five percent of EBIT from a technology 90% of them are already using. Your plane-note line: "Only 6% are beyond this, and look at the definition, it is quite modest. The ROI accrues at the workflow."

What the report says distinguishes them, p17, if asked: more likely to use AI to transform the organisation, a broader set of best practices, a wider range of AI technologies, active management of AI risk. Three-quarters report fundamentally redesigning workflows because of AI, up from 55% last year, against one-quarter of everyone else. Say "three-quarters" and "one-quarter", not 73 and 25.

90 sec The order of the bars is the spectrum, redrawn from the survey. Say that explicitly: chatbots, then coding agents, then agentic systems that act across a workflow. That is the same left-to-right you just showed them.

The "Agentic AI" bracket is McKinsey's own grouping, not yours. Worth saying, it makes the point for you.

The 32% is the best single investor fact in the report. Every SaaS line item in the portfolio is now contestable by a customer with coding agents. It cuts both ways and you should say so: the tools already inside the enterprise, Salesforce, Microsoft, NetSuite, are extremely well placed, and at the same time a third of their prospects just decided to build instead. Incumbents win the distribution, point solutions win the workflow.

Careful on the small-company figure. The report says "essentially flat" and the exhibit shows 21 to 22. Do not say "no movement at all".

90 sec Straight after the survey bars on slide 5. It is a real McKinsey exhibit with real numbers in every cell, and redrawing it live is a move nobody else in the room will have made. The point lands in three seconds because the whole grid goes nearly white.

How to run it. Open on "As published", which is the exhibit as printed, all thirteen columns, software engineering included. Say what it is. Then press T (or click the toggle): the shading re-anchors to 100% and the software engineering row greys out. Say why: "a coding agent is one person being more productive, not an agentic workflow".

The risk, and it is real. You are telling a room containing McKinsey that a QuantumBlack-authored chart is drawn on a flattering scale, while billed as the founder of QuantumBlack and days from being their Senior Advisor. Do it generously or not at all: "this is a good chart and I am going to redraw it, because the published version normalises to the highest number in the grid, which is thirty-one percent, and that makes a sparse map look dense."

The QuantumBlack analogy is the payoff: we found the intersection points in every value chain for machine learning. The same will happen here, faster. This chart is the before picture.

Public and social sector is not in this exhibit. McKinsey surveyed it (n = 71 to 81 in Exhibits 4, 8 and 9) but did not publish the agent-by-function cut for it. If asked, say so rather than guess.

2 min New section, not in any earlier draft. It is the honest answer to the previous three slides and the room will be waiting for it.

The five in full, moved off the slide on 15 September so you say them rather than the room reading them:

01 Governance and the org itself. Managers of managers of managers, and the meeting calendar that holds them together. No pick-it-up-and-go culture. The comfort zone is a process, and processes are what agents cross.

02 AI put under corporate IT. The eighteen-month enterprise chatbot project, which is the data warehouse and the data lake wearing a new coat. By the time it lands you would build it differently anyway.

03 Trust. The CIO view, and the risks are real. Data access, hosting, model choice, sovereignty. Every one of those is a legitimate objection and a legitimate delay.

04 The non-functional tax. Infrastructure, cost, routing, model selection, hosting, evaluation, A/B testing. Every point solution is solving all of it again, for a handful of users, because there is no common platform.

05 Pace. The industry moves faster than a procurement cycle. Anything specified twelve months ago is now the wrong design, and everybody knows it while they finish building it.

Anti-patterns, also off the slide: AI reporting into IT. The 18-month enterprise chatbot. Specified last year, shipped next year. Governance as the whole programme. One agent, one person, declared done.

Do not read all five. Land two and gesture at the rest. Cause one and cause two are the ones this room has personally funded.

The 14 versus 32 is the best-constructed number in the report. Use it to defuse the doom reflex before it starts, and to make the argument: the constraint is organisational absorption, not capability.

Do not attribute the non-functional point to Jeremy Palmer from the stage. That material was not cleared for slides. It is yours as a general principle, and the forty-eight-item version belongs to Sohrab Hosseini at Orq.ai if you want to credit it in conversation.

Also not on the slide, and not to be asserted: "there is no common platform, Satya". No source was found for that quote. The sentence works without it.

90 sec Ported from the Global Week keynote. It was cut from the first build of this deck for time and because it looked like a repeat of slide 2. It is not, and the difference is worth thirty seconds of your preparation.

Slide 2 is who the system serves. This is what a change has to touch. One is a ladder of scope, this is an anatomy of cost. Do not reuse the individual, team, workflow, function, company words here or the room will hear the same slide twice.

Read it from the middle outwards, once: "Code. Architecture. Teams. The organisation. Every AI tool you have been sold this year lives in the inner two. Every number you are waiting for lives in the outer two." Then stop. The picture does the rest.

The allocator turn is spoken, not on the slide (v4 removed the callout). Say it: "What you are underwriting is ring depth. A product that stops at the two inner rings has a fast sales cycle, a small contract and no defence. One that reaches the outer two has a slow sales cycle, a large contract, and a moat made of somebody else's process. Right now the market prices them the same." Ring depth is the diligence question: how far out does this product have to reach before the customer gets value, and did anyone in the customer's organisation have to change behaviour? If nobody did, there is nothing to defend.

Attribution guard: credit the onion idea to James Lewis at Thoughtworks if asked, and be clear the four rings and the bullets beside them are your version. Do not claim his diagram.

Redrawn 15 September to follow the IB talk version: bottom-aligned rings, one bullet list per ring on the right. The outer "External and Industry Context" ring is gone. The two side cards became the tags on the far right; their longer wording, in case you want it back: "Horizontal AI tools are the easy part. They only touch the inner rings, so engineering teams adopt them quickly and nobody outside the team has to agree to anything. That is why adoption looks universal." and "The value sits in vertical AI products. Those cross departments and workflows, which means touching every ring above. That is why the returns look like nothing, and why the few that work are so hard to copy."

90 sec Rebuilt on 15 September from Sohrab Hosseini's Orq.ai capability map, replacing the sixteen-box Global Week grid. Rewritten vendor-neutral on 15 September: the four-by-twelve frame is his, the items are generalised. Credit him for the frame if asked: "the shape of this comes from Sohrab Hosseini at Orq.ai; the items are what any of these platforms has to do."

Do not read forty-eight boxes aloud. Say: "Four columns. Forty-eight things. Every company you meet has built most of them, badly, for about forty users." Then go straight to the diligence question.

This is the slide that earns the words "under the hood" in the title. It is also the most cuttable slide in Section 2 if you are running long, because slide 7 cause four and slide 13 filter three both survive without it.

It started as a vendor's list and has been generalised. Whoever you buy it from, or if you build it, this is the plumbing that has to exist before the workflow does: reliability (fallbacks, retries, caching, regression tests), robustness (guardrails, validation, red teaming), resilience (monitoring, incident response, kill switch, recovery) and control (identity, permissions, audit, budgets).

Do not attribute the non-functional argument to Jeremy Palmer from the stage. Same guard as slide 7. Harness is not on this list; do not add it.

2 min The two spectrums come back on one page and you talk to the clusters. Press 1, 2, 3, 4 in turn (or click the pills) to draw a sketched ring round each cluster; press the key again to remove it, or Clear. Rings can stack.

Walk it as four rings. 1: nearly nine in ten respondents, one person and a prompt box. 2: the enterprise chatbot, scaled by a company, still used by one person at a time, steps one to five. 3: agents that act across work, thirteen to twenty-five percent, the first time the ring reaches the workflow. 4: the six percent, functions and companies run on reasoning and outcomes. Then point at the two frontier markers: the OpenAI research org at three agent-days per human day, and OAI-HF, where agents managing agents inside a benchmark coordinated, pursued a goal nobody set and attacked the grader.

Get the dates right. The incident was July 2026. Amodei's essay is 12 September 2026. HF is Hugging Face. Do not conflate them.

Choose your emphasis before you walk on. Optimistic: the far right of both exists, so the empty cells can fill with what already ships. Sober: the first time we saw that end at scale it went somewhere nobody asked it to. Both are true. The close is the three thoughts, so the sober read is the natural fit.

Carry OpenAI's caveat if pressed: over half of successful four-to-eight-hour tasks needed at least one human intervention. Autonomy is not yet unattended. It was on slide 13 until 15 September; the Crosby time-to-review quote replaced it, so say it rather than point at it.

45 sec Split from the old slide 12 (v3 numbering) on 15 September to mirror the Global Week structure: what they said, then what their behaviour says. This is the setup. Read both quotes flat, no commentary, and let the room agree with them.

Use Suleyman's own words. The Global Week version paraphrased it as "software engineering will be automated in 18 months", which is stronger than what he said and not defensible. His claim is broader and softer-edged, and software engineering was his supporting evidence rather than his prediction.

Then click. The next slide is the punchline and it should arrive within ten seconds of the second quote.

90 sec The punchline to slide 11. Deliver it as one move: the ledger, then the line. Do not explain the ledger row by row. Say: "Nine and three quarter billion dollars, in seventy-one days, from the five companies who can see every prompt on earth."

The honest counter, have it ready because someone will make it. These vehicles are partly distribution plays to lock in model spend, and McKinsey and Capgemini invested in the OpenAI one, so the incumbents are hedging rather than surrendering. Concede it and keep the point: even as a distribution play, it is an admission that the model does not deploy itself.

Q&A ammunition, not on the slide because it has no published source: a founder selling an AI product to enterprise procurement teams told you it takes six to eight months to sell the product and two to three weeks to sell the same outcomes as a services engagement. Attribute it as one founder's experience, not a benchmark.

Only announced headcounts are on the slide: Microsoft 6,000, Tomoro about 150, Ode about 100, AWS pods seeded with "thousands". The Global Week estimates are removed; do not quote them.

Time check. If you are past 09:38 arriving at slide 11, skip 11 and 12 together and go to slide 13. They are the first cut, as a pair.

2 min Your own filter, from the plane notes, verbatim in substance. This replaces the five diligence questions as the leave-behind. Do not do both, it is the same job twice.

The Jeremy quote is cleared. Everything else he said on that call is not, so do not improvise around it. Nothing about PhysicsX numbers, customers, hiring, his transition, or his view of Visium.

The consultants line is spoken, not on the slide (v4 removed the green callout): "One consequence I did not expect: consultants can make very good founders now, because they know where the work actually happens inside an organisation, and they already have the trust to get near it. I have become suspicious of founders who have never been inside one." It will get quoted back to you at coffee and it is slightly provocative in a room of venture investors who fund the opposite. Deliver it lightly and let them argue.

Time check. You should be at 09:41 leaving this slide, and at 09:38 if you cut as planned. If you are past 09:42, go straight to the close and take three questions instead of five.