The 7% Problem

"HBR Guide to Generative AI for Teams" (Harvard Business Review Press, 2026) by Elisa Farri and Gabriele Rosani (Capgemini Invent) — cross-referenced with the HBR Press Live launch webinar, hosted by Sally Ashworth, 15 September 2026

Half your organization is already using generative AI before its first meeting of the day. Almost none of it survives the walk from the desk to the boardroom. That is the finding Elisa Farri and Gabriele Rosani put in front of a live audience of a few hundred managers on a Tuesday evening, and it is also, more precisely, the finding printed on the first page of their book: a Capgemini Research Institute survey of 500 leaders — director level and above, across sixteen countries — showing individual daily use of generative AI climbing from 15% in 2024 to 49% a little over a year later. Team use, over the same period, sits at 7%. Not “avoid.” Not “distrust.” Nobody thought to.

The webinar launched the book built around that seven per cent — the sequel to Farri and Rosani's earlier guide for individual managers. This recap has since been checked line by line against the finished book, not just the live session, so what follows carries proper chapter and page references throughout — the mechanics the authors demonstrated, the sentiment the room supplied for free, and where both land once you translate them onto a hotel floor.

TL;DR

• The blind spot has an exact number: 49% individual daily use of GenAI, 7% team use (HBR Guide to Generative AI for Teams, Introduction, p.1, citing Capgemini Research Institute, “How AI Is Quietly Reshaping Executive Decisions,” Jan 2026). Hospitality's own leadership meetings — EXCOM calls, pre-opening task forces, revenue strategy sessions — sit almost entirely in that unclaimed 93%.

• Both live case studies the authors walked through map almost exactly onto hotel mechanics — and each maps onto a specific, citable chapter: the Leadership Team Meeting/Business Review chapters (11, 26) for a multi-time zone leadership review, and the Business Model Innovation chapter (19) for a challenger/persona-driven concept workshop.

• The sharpest questions in the room weren't about capability — they were about ownership, data reliability, and where an organization gets an affordable, secure, whole-company AI tool. Nobody on stage had a clean answer. Neither does hospitality yet.

Bottom line: The technology is not the constraint. The design of the meeting is.

The gap has a number now

The book's introduction (pp.1–2) is blunter than the webinar had time to be. Asked why they haven't brought AI into meetings yet, Farri and Rosani write that managers typically give one of two answers: “I never thought about it,” or “I'd like to, but I don't know how.” Elisa Farri's live account of the origin story filled in the texture behind that line — a year on the road presenting the managers' guide, more than a hundred events, thousands of leaders trained, and one question that never once came up: how to use generative AI with other people, in the room, together. When the authors started polling their own audiences directly, the hands stayed down. The Capgemini Research Institute then ran the numbers properly, and the anecdote became a statistic.

For an industry that runs almost entirely on meetings — the morning briefing, the revenue call, the pre-opening war room, the brand standards audit — a 93% blind spot on team AI use is not a rounding error. It is the whole opportunity, sitting untouched.

How the book actually works

The guide is organized around twenty recurring team activities (Chapter 4, p.39), grouped into five clusters across Sections Two to Six that will be instantly recognizable to anyone who runs a hotel leadership calendar: Planning, Decision-Making, Ideating, Problem-Solving, and Reviews and Retrospectives — four activities per cluster. Each activity gets the same three-part treatment — the typical pitfall (a strategy workshop turns inward-looking; a review becomes a status recital instead of a decision), how AI can counter it before, during, and after the session, and a trigger prompt the reader can run directly. The authors are explicit about their method here: every prompt in the book was tested as end users across “Microsoft's Copilot, OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini” (Introduction, p.5) — genuinely vendor-agnostic, not just badged that way. The underlying design intent, stated as the first of the book's own guidelines, is to “be intentional” (Chapter 2, p. 29) and to provoke a back-and-forth between team and model, never to hand over a verdict. AI plays sparring partner, not oracle.

Two live examples, drawn from the authors' broader consulting work, did the actual persuading in the webinar. Both map onto specific chapters a hotel reader can go and use directly.

Case one: the CIO's quarterly review

A CIO at a luxury company — Elisa Farri's example from the live session, given the pseudonym “Andreas” there — ran a leadership review spanning three time zones: the U.S. East Coast, Hong Kong and Singapore, roughly forty people dialing in. Before the meeting, every leadership team member worked through a structured prompt individually, using AI as a deliberately skeptical peer to pressure-test the idea they planned to bring into the room. During the session itself, teams split into Microsoft Teams breakout groups, with one member “prompting on behalf of the group” while the AI cycled through roles — facilitator for round-robin introductions, analyst to cluster the pre-work, an external stakeholder to challenge the ideas hard, then analyst again to consolidate. This isn't a one-off invention: the book prints Facilitator and Analyst as named, ready-to-use agent types (Chapter 2, pp.30–31), and its Business Review chapter builds the external-challenge step directly into a trigger prompt — “How would our competitor [specify] evaluate its positioning versus our last quarter performance?” (Chapter 26, p.277). The reported gains were speed — output the team said it could never have produced unprepared — and quality: the adversarial AI role gave the room, in Farri's words, the breath to reopen the discussion rather than settle for the first idea raised.

Case two: the innovation workshop

A consumer products team pressure-testing a new business concept took a different route: two custom agents, prepared in advance, rather than a single evolving prompt. One played Challenger, surfacing weaknesses and objections the team had talked itself past — a role the book introduces as early as Chapter 1: “you can control [AI's tendency to agree with you] by asking AI to play the role of a challenger or stress-test your ideas by taking the perspective of your customer or competitor” (p.12). The other played a Customer Persona, which the Business Model Innovation chapter spells out as its own trigger prompt: “Help us create a persona […] to test this assumption. Guide us, one step at a time, to define the persona's characteristics, goals, and behaviors; then generate a questionnaire we could use and simulate realistic answers from this persona” (Chapter 19, p.210). Both agents were built to ask, wait, and push back — never to hand down a rating — which is precisely what kept the team, not the model, in charge of the concept. The book documents a close real-world parallel in print: a series of workshops with fifty managers from thirteen companies who stress-tested new business models against a custom GPT built as a “digital twin” of Trabucchi and Buganza's Platform Thinking methodology (Chapter 19, pp.204–205) — a citable anchor for any hotel group that wants to see the mechanic working at scale before trying it in-house.

The webinar's detail worth stealing outright — not printed in the book itself, but consistent with its philosophy throughout — was that the same team ran a second, AI-guided session a few days later, not on the concept, but on how they had used AI the first time round. A retrospective on the retrospective's own tooling, treated with the same rigor as the original workshop. It is a discipline most organizations never apply to their own technology rollouts, hospitality included.

The authors' own synthesis of both stories was blunt: none of this survives improvisation. You cannot walk into a meeting, throw a prompt at the room, and expect the magic to arrive unassisted. What has to be designed — in advance, and tested before the room ever sees it — covers three dimensions: the setting (remote versus in-person), the interaction's shape (one chat, sequenced agents, or a skill built for the purpose), and the dynamics — whether the instructions are written to pause and let the humans argue, or to plow ahead and answer for them.

The twin benefit, and the three conditions that make it real

“The Twin Benefits of Integrating AI into Teamwork” is, in fact, the title of the book's third chapter (pp.33–38), and the number behind it is more modest than the one used live in the webinar: qualitative field experiments with more than one hundred managers, working in small groups of three or four on complex business problems — not the three hundred cited from the stage, though the direction of the finding is identical either way. Two-thirds reported higher-quality output and greater engagement using AI as a team; the same two-thirds reported the classic solo failure modes better mitigated — the trust trap of over-reliance, hallucinations mistaken for fact, the conformity trap of settling for bland, average thinking, and the speed trap that turns rapid answers into hasty decisions (pp.36–38). The book also points to independent corroboration: a Procter & Gamble study of its own is cited finding that lone AI users worked faster but converged on average results, while teams using AI together were significantly more likely to land in the highest 10% of output quality (p.35, citing P&G, “The Future of Collaboration,” April 2025). Used together, teams naturally slow down, argue, and check each other. That is the twin benefit: quality up, risk down, simultaneously.

None of it happens automatically. The webinar distilled this into three conditions — intentionality, craft, and collective ownership — that decide whether a team gets the twin benefit or just an expensive chatbot in the corner of the call. Only the first is the book's own named term: “Be intentional” is literally the opening guideline of Chapter 2 (p.29). The other two are the authors' live shorthand for ideas that run throughout the book without a single chapter heading — the warning against becoming “a passive observer simply waiting for the output” (Chapter 1, p.12), and the instruction to “not let the prompter dominate” a group session (Chapter 2, p.27).

If your team is just watching the screen, that's not a good sign — AI is taking the lead, not the team. Gabriele Rosani, on the tell that a session has gone wrong

The room told its own story

The chat box provided its own dataset, and it is worth reading in order. The early minutes brought the usual scatter of arrivals — Sheffield, Chichester and Derby in the UK, Illinois and Florida in the US — including one attendee who had started the book the previous day and was, unprompted, already inspired. That geographic spread, on a UK-hosted evening broadcast, is itself a small data point about how far this particular anxiety travels.

The two sharpest questions of the night came from the same attendee and drew visible peer approval — reaction counts and a “great question” from another attendee in the room. The first asked how confident teams really were about owning an output generated so quickly. The second asked how reliable the underlying data for the workshop's “customer persona” agent was — whether a synthetic respondent is a fair stand-in for a real one. Both are exactly the questions a hotel EXCOM should be asking before it lets an AI-run guest persona anywhere near a concept review, and neither was rhetorical: they came from practitioners who had clearly already tried this and hit the edge of their own comfort with it.

The closing Q&A produced the most consequential exchange of the session, and it wasn't really a question about the book at all. One attendee laid out a familiar shape of problem plainly: individuals already paying personally for Claude or ChatGPT, plus Gemini and a Zoom AI note-taker for meeting minutes, and no affordable, secure, whole-company option they could trust with confidential information. The host had to concede, fairly, that the authors couldn't referee AI vendors — the book is deliberately platform-agnostic, and capability shifts too fast to pin to one model. That's not a gap in the book. It's the governance work every organization, hospitality included, still owes itself before any of this reaches a meeting with real numbers.

One more question landed in the Q&A box and never received an answer before the session closed: whether the recent public discussion of AI systems behaving in unintended, “rogue” ways concerned the authors personally. The session ended on that unresolved note. Given that the previous forty minutes had been spent teaching a room full of managers to widen AI's role in their own decision-making, that particular silence is worth sitting with rather than filing away.

Translating the mechanics onto the hotel floor

None of the book's mechanics require a translation layer for hospitality — they map onto the existing leadership calendar almost without adjustment.

The three conditions travel just as directly. Intentionality sits with the GM or department head running the session, not with IT — a recurring theme in this practice's own governance work: AI adoption in hospitality is a leadership decision, not a procurement line item. Craft means resisting the passive version everyone defaults to first — uploading a monthly P&L deck and asking “any concerns?” is the same flat, one-shot pattern the authors warned against, not a genuine sparring session. And collective ownership has an exact hotel-floor translation of its own: if your department heads are sitting quietly while a corporate AI note-taker summarizes the call, that is the spectator-mode warning sign, transplanted intact — the meeting has an AI in it, but the team is no longer the one making the decision.

THE PERTLINK VIEW

This webinar answered its own headline question — how do teams use AI together — with genuine craft and two well-chosen case studies. It could not answer the question its own audience actually cared about most: where does an organization get a secure, affordable, whole-company environment it can trust with confidential data before it invites AI into a room with real guest and rate information in it? That is not a criticism of two authors who were candid enough to say, correctly, that it isn't their book's job to referee vendors.

It is, however, exactly the boundary this practice keeps returning to under other names — the Decision Envelope, the AI Override Rate, the governance layer that has to be designed rather than assumed before a Human Experience Orchestrator can responsibly widen AI's role in a hotel's own decision-making. Bringing AI into the meeting is the easy seven per cent to fix. Governing what it's allowed to say once it's there is the harder, and still largely unfinished, work.

Sources

• Farri, E. and Rosani, G., HBR Guide to Generative AI for Teams (Boston: Harvard Business Review Press, 2026) — Introduction, pp.1–6; Chapter 1, pp.9–18; Chapter 2, pp.19–32; Chapter 3, pp.33–38; Chapter 4, p.39; Chapter 19, pp.199–211; Chapter 26, pp.271–281.

• “HBR Guide to Generative AI for Teams” — HBR Press Live launch webinar, hosted by Sally Ashworth (Harvard Business Review Press), with Elisa Farri and Gabriele Rosani (Capgemini Invent), 15 September 2026 — live transcript, slides and chat.

• Capgemini Research Institute, “How AI Is Quietly Reshaping Executive Decisions,” January 2026, N = 500 C-level executives across sixteen countries, as cited in the book's Introduction, p.1.

• P&G, “The Future of Collaboration: How AI Is Supplementing Teamwork and Innovation at P&G,” April 4, 2025, as cited in Chapter 3, p.35.

BOOK AVAILABLE VIA: https://hbr.org

The intelligence may be artificial. But the experience is human.

Made with the help of various AI tools – but always with a HITL