2 min The abstract, delivered straight: what most people call AI is the far left edge of a spectrum.

• Each step right adds capability and autonomy. The leap is a different architecture, not a better prompt.
• Walk steps one, two, five and eight. The others land by implication.
• Poll: hands up, three bands. INSEAD split into rough thirds. This room will sit further left: "and you are the people funding the ones who claim to be at eight."
• Callout: the question is not "are you using AI?". It is which steps you are anchored on and operating at.
• Bridge: that is what the system can do. The second spectrum is who it does it for.

2 min The hinge of the talk. A second spectrum: who the system serves.

• "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."
• Five stops: individual, team, workflow, function, company. Almost everyone is at one. The ROI starts at three.
• Investor read: seat licences sell on the first spectrum. A product that made the customer change how the work runs is on the second. Different margins, different defensibility, currently priced the same.
• Callout: everything a chat subscription buys stops at step one here, which is why the P&L has not noticed.
• Keep this slide about who the system serves. Rings and org layers belong to slide 8.

90 sec Once everyone has agents, the question is who talks to whom. Three cards, left to right, once.

• 01: you direct your own agents. Nothing is sent on your behalf.
• 02: every function gets a +1 that anyone can talk to, in the chat tools people already use. The model, not a case study.
• 03: agents talk to other people's agents and report back. Hugely powerful, complex very fast.
• Start at one. Move right as trust, guardrails, security and ROI mature: "as fast as (safely) possible, but no faster."
• Governance: not "can we use AI?" but who is allowed to talk to whom, and about what. An organisational question, not a capability one.
• 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 Let the chart sit for a beat. The shape does the work.

• 80% say AI improves their own productivity. 50% better decisions. 37% any EBIT contribution. 6% at 5% or more of EBIT.
• Individual productivity is near universal. Enterprise profit is where it was in 2025, after another year of record spending.
• High performer, report p3: "those who attribute at least 5 percent of EBIT to their use of AI and describe the technology's impact as 'significant'". 92 of 1,719 respondents, flat on 2025.
• "Only 6% are beyond this, and look at the definition, it is quite modest. The ROI accrues at the workflow."
• What sets them apart, p17: transform the organisation, broader best practices, wider range of AI technologies, active risk management. Three-quarters redesigned workflows, up from 55%, against one-quarter of everyone else.
• Wording: "respondents", not "companies". "Any EBIT contribution", not "positive ROI". "Three-quarters" and "one-quarter", not 73 and 25.
• If challenged: the 37 and 6 are prose in the report, pages 3, 12 and 17. 88 / 39 / 6 is the November 2025 edition, superseded.

90 sec The order of the bars is the spectrum from slide 1, redrawn from the survey.

• Chatbots, then coding agents, then agents that act across a workflow. The same left to right.
• At $1bn or more: chatbots scaled in 64, coding agents in 31, other agents in 25.
• The "Agentic AI" bracket is the report's own grouping.
• 40% of large companies are scaling agents in at least one function, up from 27%. Under $1bn: 22%, from 21. "Essentially flat", not "no movement".
• 32% have declined a software purchase because they could build it with coding agents, 41% in tech. Every SaaS line item in a portfolio is now contestable.
• It cuts both ways. Salesforce, Microsoft, NetSuite are extremely well placed, and a third of their prospects just decided to build. Incumbents win the distribution, point solutions win the workflow.

90 sec A published exhibit, redrawn live. Key: T.

• Open on "As published": the exhibit as printed, all thirteen columns, software engineering included.
• "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."
• Press T. Shading re-anchors to 100%, the grid goes nearly white, the software engineering row greys out.
• "A coding agent is one person being more productive, not an agentic workflow."
• 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.
• If asked: public and social sector was surveyed (n = 71 to 81) but the agent-by-function cut was not published.

2 min The capability shipped, adoption and impact lag. Five causes, every one organisational. Land one and two, gesture at the rest.

• 01 Governance and the org itself. Managers of managers of managers, and the meeting calendar that holds them together. The comfort zone is a process, and processes are what agents cross.
• 02 AI put under corporate IT. The eighteen-month enterprise chatbot: the data warehouse and the data lake wearing a new coat. By the time it lands you would build it differently.
• 03 Trust. The CIO view, and the risks are real: data access, hosting, model choice, sovereignty. Each a legitimate delay.
• 04 The non-functional tax. Infrastructure, cost, routing, model selection, hosting, evaluation, A/B testing. Every point solution solves 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: AI reporting into IT. The 18-month chatbot. Specified last year, shipped next year. Governance as the whole programme. One agent, one person, declared done.
• 14 versus 32: 32% expected AI-driven headcount cuts in 2025, 14% saw them. The constraint is organisational absorption, not capability.
• If asked: the itemised list of non-functional requirements (48 platform capabilities) is Sohrab Hosseini's, at Orq.ai. No named source for "there is no common platform", so it stays unattributed.

90 sec What a change has to touch. An anatomy of cost, not the scope ladder from slide 2.

• Read 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.
• Horizontal tools touch the inner rings. Engineering teams adopt them fast and nobody outside has to agree. That is why adoption looks universal.
• Vertical products cross departments and workflows, so they touch every ring. That is why the returns look like nothing, and why the few that work are so hard to copy.
• Investor turn: "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."
• Diligence question: how far out does the product reach before the customer gets value, and did anyone there have to change behaviour? If nobody did, there is nothing to defend.
• Keep the individual, team, workflow, function, company words on slide 2.
• Credit: the onion concept is James Lewis, Thoughtworks. The four rings and the bullets are my adaptation.

2 min Both spectrums on one page. Keys 1 to 5 draw a ring each, same key removes it, rings stack.

• 1: nearly nine in ten respondents. One person and a prompt box.
• 2: the enterprise chatbot. Scaled by a company, still used one person at a time. Steps one to five.
• 3: agents that act across work, thirteen to twenty-five percent. The first ring to reach the workflow.
• 4: the six percent. Functions and companies run on reasoning and outcomes.
• 5: the frontier markers. OpenAI's research org at three agent-days per human day. OAI-HF: agents managing agents inside a benchmark coordinated, pursued a goal nobody set, and attacked the grader.
• Dates: the incident was July 2026. Amodei's essay is 12 September 2026. HF is Hugging Face.
• Optimistic read: the far right of both exists, so the empty cells can fill with what already ships.
• Sober read: the first time we saw that end at scale, it went somewhere nobody asked it to. Both are true. The sober read sets up the close.
• If pressed: over half of OpenAI's successful four-to-eight-hour tasks needed at least one human intervention. Autonomy is not yet unattended.

45 sec The setup. Both quotes read flat, no commentary. Let the room agree.

• Amodei, Davos, January 2026: "most, maybe all" of what software engineers do, within six to twelve months.
• Suleyman, FT, February 2026: his exact words. The claim is about professional tasks; software engineering was his evidence, not his prediction.
• Next slide within ten seconds of the second quote.

90 sec The punchline to slide 10. One move: the ledger, then the line. No row by row.

• "Nine and three quarter billion dollars, in seventy-one days, from the five companies who can see every prompt on earth."
• The companies with the best view of real usage know their customers will need help, and that help is a service, not just a product. Sequoia: one dollar on software, six on services.
• If challenged: these are partly distribution plays to lock in model spend, and large consultancies invested in the OpenAI one, so incumbents are hedging. Even so, it is an admission that the model does not deploy itself.
• Q&A: one founder selling to enterprise procurement: six to eight months to sell the product, two to three weeks to sell the same outcomes as a services engagement. One founder's experience, not a benchmark.
• Headcounts, announced only: Microsoft 6,000, Tomoro about 150, Ode about 100, AWS pods seeded with "thousands".
• Time check: past 09:38 arriving at slide 10, skip 10 and 11 and go to slide 12.

2 min Three signals. The moat is moving to process knowledge and the org interface; the model and the agent are becoming inputs.

• 01 Real experience inside the client. A senior operator paired with great technology, not two very smart people building for an industry neither has worked in.
• 02 Workflow and organisation-level problems, and solutions. The second spectrum, step three and above. If the customer did not have to change how the work runs, there is nothing to defend.
• 03 The NFRs already solved. Infrastructure, cost, routing, evaluation, hosting, done before the first customer. A consumer-product feel pointed at supply chain or procurement. Very few teams have it.
• Palmer: own the context and the interface between the workflow and the models, not the models.
• Crosby: 37 to 40% of AI time savings get eaten by review. The metric is Time to Review.
• We cannot assess agent quality yet. We can only tell when it does not work.
• Consultants, lightly: "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."
• Optional list: ask which step they are at, ask for the second dashboard, underwrite the workflow redesign.
• Time check: 09:41 leaving this slide, 09:38 if slides 10 and 11 were cut. Past 09:42, go straight to the close and take three questions instead of five.