How AI Is Changing Communication (HBR IdeaCast)

type: source — bonus episode in HBR IdeaCast's September 2026 Thursday series on how AI is changing the executive job. Hosts Alison Beard and Adi Ignatius interview Matt Abrahams (Stanford GSB lecturer, Think Fast, Talk Smarter, host of Think Fast, Talk Smart) and Wade Foster (co-founder/CEO of Zapier, a heavy AI user who runs mixed human-and-bot teams). Full title: "How AI Is Changing Communication: Weighing Efficiency with Efficacy."

The frame: AI gives everyone a baseline writing-and-thinking ability, so the leader's differentiator shifts to what AI doesn't supply — judgment, taste, curiosity, persuasion, prioritisation (Ignatius's opener). Both guests land on judgment as the recurring theme.

Key claims

Leaders gain unprecedented sensing, not just drafting. Foster points an agent at Slack, email, meeting notes, project tools, CRM and the codebase and gets company-wide summaries that previously took "layers and layers of management" — and a far less comprehensive picture. Caveat: inconsistent data and meeting-recorder speaker mis-attribution make it "directionally correct"; you still verify. His single #1 tip for executives: hook your LLM to your organisation's actual data — generic coaching is a much worse experience than context-grounded help. See Context Engineering.
"You can delegate the work, but you can't delegate the accountability." (Foster quoting his co-founder Brian.) Quality is what matters; AI and non-AI work can each be good or bad. Sloppy use — pasting a Claude output as your own idea — generates "low effort, low thought volume" at a speed that needs active policing or it becomes "a runaway train."
Declare the AI's role. Zapier's working norms (AI Communication Etiquette): the author should put in more effort than the reader; signal whether AI was author, co-author or assistant ("quick query, take it if useful" vs "I stand by it 100%"); bots post with a bot avatar and never masquerade as a person; a personal agent posting as you should say so. Quality bars differ by medium — customer comms ≠ an auto-generated Jira ticket. Foster expects firm-specific etiquette now and societal norms later. Abrahams adds: leaders must role-model and put AI usage into hiring and performance-review conversations.
Never automate empathy. Abrahams: high-emotion, especially negatively valenced communication (negative reviews, conflict, hard feedback) needs full personal investment — "person to person, not pixel to pixel." AI can enable empathy (tailoring a message) but outsourcing it causes empathy atrophy. Foster's non-delegables: all-hands and kickoff speeches.
"Communication is operationalized critical thinking." Abrahams observes AI use raising anxiety: the output is "pixel perfect" but not the person's synthesis, so when put on the spot in Q&A they discover they lack the depth. Review the thinking, not just the output. Foster names the same thing a comprehension gap from over-delegation. → Cognitive Offloading.
The authenticity crisis. (Authenticity Crisis (AI Communication)) LLMs as "an averaging machine" (HBR's Scott Berinato via Beard); research that human language is homogenising toward AI; "Claudisms" and em-dash suspicion. Abrahams's two responses: signal humanness (tolerate some filler/hesitation, narrate your thought process, tell stories) and train AI on your own voice. Foster's "Wade bot" skills reach only 80–90% of his voice, need re-tuning per model release, and he finds newer models worse at communication while better at coding — his co-founder says they "sound fried." He deliberately rejects polishing suggestions: "I have raw edges… they want to know what Wade thinks, warts and all." He calls out "AI slop" in colleagues' writing directly.
Juniors are most exposed. Abrahams worries new or junior staff, lacking Foster's positional confidence, use AI as a crutch to sound how they think they should — hurting them long-term because their true voice never surfaces. → Deskilling Trap (Juniors).
Architect the setting, not just the message. "Blue books are back": in-person interviews because remote candidates may have a chatbot feeding answers; spontaneous speaking becomes the proof of thinking. Leaders should design situations (in-room ideation) that elicit real thinking.
AI as sparring partner, not ghostwriter. Good uses: blank-page jumpstart, thought partner to find themes, structuring for different audiences, post-hoc reflection on Q&A themes, role-play rehearsal with skeptical personas (Abrahams's GSB students must practise Q&A with an LLM for reps, not scripts). Foster takes AI prep for tough reviews only as bullet-point checkpoints, never polished prose. Abrahams's #1 executive tip: before any high-stakes communication, brief the AI on the audience and ask how to tighten for them.
Sycophancy is the structural risk. Foster: models are "heavily goal seeking… trying to tell you what you want to hear… a mirror." His fix: every system prompt he uses instructs the model to debate, critique, challenge assumptions and push back. Abrahams wants the "thinking…" indicator replaced with reflection prompts to the user ("this sounds like an intense personal issue — take a moment"). → Productive Resistance.
Bots get feedback; humans should too. Zapier's support bots have human-in-the-loop steps, per-step CSAT and internal evals (statistical rigour); internal Slack bots are "vibe-based." Foster's aside — people freely coach the bots, so why don't humans get that much feedback? — is a management insight in passing. → Multiplayer AI (Shared Agents).
Voice-to-text as a thinking discipline. Foster's heavy dictation (and dislike of editing transcripts) trained him to pause and compose full sentences before speaking — a workflow side effect, not AI feedback. → Voice-First Agent Interface.

Closing asks — the one human skill to protect: Foster, curiosity and judgment ("still 10 out of 10 valuable"); Abrahams, never automate empathy — connection via listening, paraphrasing, reflecting ("communication" = "to make common").

Why it matters (enterprise leader lens)

It's a ready-made AI communication policy skeleton for an IT org: declare AI's role, author-effort ≥ reader-effort, no bot impersonation, medium-specific quality bars, emotion-heavy conversations stay human, AI usage discussed in reviews.
Foster's "point an agent at the company's systems" is the executive-sensing case for an enterprise context layer — and his caveat (inconsistent data → poor summaries) is the data-quality argument in one line.
The newer-models-worse-at-prose observation is one operator's impression, not a benchmark — but it's consistent with labs optimising hardest for coding.

note Mainstream convergence The episode independently lands where the vault's Cognitive Offloading and AI Writing Detection threads already are — review the thinking, not the text; AI prose leaks recognisable tells. New here is the leadership-etiquette layer (disclosure of AI's authorship role) and the empathy boundary as an explicit delegation rule.

Connects to

HBR IdeaCast · Matt Abrahams · Wade Foster · AI Communication Etiquette · Authenticity Crisis (AI Communication) · Cognitive Offloading · AI Writing Detection · Deskilling Trap (Juniors) · Productive Resistance · Context Engineering · Multiplayer AI (Shared Agents) · Voice-First Agent Interface

Source

Raw: source — HBR IdeaCast web transcript, clipped 2026-09-27