How AI Agents Are Changing Business - PwC Chief AI Officer Dan Priest

[Video description] You’ve probably heard it already: AI is going to radically change the way we work. But the details change depending on who’s making the prediction: AI will wipe out the C-suite, or entry-level jobs, or make us all into prompt engineers. Those scenarios are far-fetched, says Dan Priest. In his role as the Chief AI Officer at PwC, Priest sees firsthand how companies across the industrial landscape are utilizing AI–often in ways that clash with those fatalistic prognostications. In conversation with Nicholas Thompson, CEO of The Atlantic, Priest unpacks what he’s learned about AI implementation, agents, and how businesses can adapt in the age of AI.

Produced by Atlantic Re:think, The Atlantic’s creative marketing studio, in collaboration with PwC.

TL;DR

A ~31-minute episode of “The Most Interesting Thing in AI” (Nicholas Thompson, CEO of The Atlantic) — Thompson’s second conversation with Dan Priest, Chief AI Officer at PwC, an 18 months follow-up to a December 2024 episode. Ten load-bearing claims:

1. Two extremes of AI adoption, both irrational. Some clients are “extremely aggressive” — planning double-digit headcount cuts (via layoffs or redeployment) with hard ROI expectations. Others insist their business is “somehow immune from AI.” PwC talks clients back toward the middle, grounded in proven use cases.
2. The billion-dollar one-person company is walked back. Priest revises his own December-2024 prediction: accountability structures, risk controls, and responsible-AI guardrails require distributed human ownership, which structurally limits how far or fast a single-person, all-agent company can scale.
3. Adoption moves at the speed of trust, not technology. Citing a Stanford study on national trust that AI is “on balance better for humanity” — ~80% in China and the UAE, ~40% in the US — Priest frames the adoption gap as a trust gap more than a capability gap.
4. Back office first, then front office, now middle office. Code generation is “the number one most industrialized use case” — squarely back-office. PwC’s own Southwest Airlines case study cut 50% of time/effort from the design phase alone (before code generation), with 30–50% benefits once code generation is in play. AI then moved into front-office customer service, and is now moving into middle office (“probably the least penetrated”). Notably: developer headcount and recruiting are up, even as CS college enrollment falls.
5. Specialists and generalists both matter, with generalists rising. Decades-long specialization trend continues, but Priest sees rising value in generalists who apply first-principles/hypothesis-driven problem-solving to orchestrate teams of highly specialized agents — “we don’t have enough innovators right now.”
6. PwC’s internal knowledge-compounding GPT. Every employee having a personal agent is coming for general-productivity reasons, but PwC’s concrete example is a firm-wide GPT, built from the accumulated learnings of its AI-transformation consulting practice (which does “10X what any other individual team does in PwC” on transformation work), made available to every consultant on every project — turning what would be a small team’s exclusive expertise into a shared resource.
7. Agent performance is bounded by task length, not agent “personality.” PwC uses multiple LLMs deliberately, for checks-and-balances, and tracks accuracy, latency, drift, hallucination, deception rate, and task length per model. Individual agent accuracy clusters at 85–90%. Sustained task length (a proxy for “concentration”) averages ~2.5 hours across models for general analytical/slide-generation work — well short of an 8-hour unattended task, at which point success rate drops below 50%. Coding tasks, being more structured and testable, sustain much longer — roughly 12–13 hours.
8. The best talent gets the most from AI — across domains. Contra the “AI helps beginners most” narrative from some earlier research, Priest reports the opposite pattern at PwC and beyond: the best coding engineers, and (per an anecdote) the best musicians, extract disproportionately more value from AI tools than average performers, despite equal platform access.
9. Hiring is up, and interviews now test agent-management skill. PwC has grown net hiring by ~5,000 people, testing candidates on managing agents and multiple platforms/languages during interviews. The barrier to entry for coding has fallen — non-traditional coders are entering engineering roles via code generators and becoming strong data scientists, testers, or developers.
10. What comes after agents: “agentic applications,” and the hourglass organization. Priest’s headline prediction: bundles of a small number (10–15) of authoritative, specialized expert agents per function (e.g. finance: planning/forecasting, AR/AP, close, reporting) that work together and start to look like an intelligent application — a layer where proprietary, differentiating capability will concentrate (vendors, foundation-model companies, and companies building their own internal stacks will all compete for this layer). Organizationally, Priest describes a shift from the “pyramid” (and the never-fully-realized “diamond” of the outsourcing era) toward an hourglass: expanded entry-level intake at the bottom, a growing leadership layer at the top, and a compressed-but-empowered middle management layer whose members are given AI tools, resources, and a performance-based path up (“up-or-on,” not “up-or-out”).

What was actually ingested

The full ~30:47 episode via its auto-generated (ASR) English caption track (a manual English (US) track was also available but not used) — 225 transcript segments, with 10 YouTube-provided chapter markers. Light ASR cleanup applied at ingest time: (a) removed a pervasive DOM-scraping artifact where YouTube’s screen-reader-oriented spoken-form timestamps (“1 minute, 3 seconds”) and inline “Chapter N: …” chapter-marker text were captured as if they were spoken dialogue (213 occurrences of the former, 10 of the latter — a new failure mode not previously documented in the youtube-transcript-skill’s known issues; worth adding to the skill’s failure-modes list if it recurs); (b) corrected “PWC” → “PwC” and one stray typo (“for .ots of reasons” → “for lots of reasons”).

Why this source matters to the wiki

This is the wiki’s first source authored from inside a Big Four professional-services firm’s own AI transformation practice (as opposed to consulting-firm research/marketing content about other companies, or academic/vendor material). It supplies:

• A concrete, quantified task-length ceiling for agent autonomy (~2.5 hrs general work vs. ~12–13 hrs coding) — a specific, falsifiable operational metric that most of the wiki’s agentic-AI material discusses qualitatively.
• A named, working instance of digital-transforming/improving-digital-maturity: a firm-wide GPT compounding a specialist team’s learnings and distributing them to the whole organization, functioning as an internal knowledge-democratization platform.
• A second independent voice (after Allen’s AWS talk) converging on the hourglass organization shape for the AI era, from a materially different vantage point (Big Four internal transformation practice vs. a cloud vendor’s executive-forum talk) — worth watching as a possible emerging consensus frame, see organizational-frameworks-for-ai-adoption.
• A practitioner-level restatement of the jagged frontier’s unevenness claim, independently observed across coding and music-making, not just knowledge work generally.

Linked entities and concepts

• dynamic-capabilities — hourglass restructuring and the internal-GPT knowledge-compounding case are concrete digital-transforming instances.
• warner-wager-process-model — tagging vocabulary source; see dynamic_capabilities: above.
• enterprise-ai-adoption — back-office-first deployment sequencing (code gen → customer service → middle office) is a concrete instance of the wiki’s broader adoption-pattern material.
• jagged-frontier — “the best talent gets the most from AI” is a direct practitioner corroboration.
• organizational-frameworks-for-ai-adoption — the hourglass-organization model is a second independent data point for this synthesis.
• PwC — first wiki mention of this organization as a central subject (not a passing mention); promoted to an entity page on this ingest given the depth of firm-specific detail (internal GPT, task-registration “operating system” for agents, hiring figures, Southwest Airlines case study).
• Dangling (single-source mention, deferred per Author-entity promotion): Dan Priest (Chief AI Officer, PwC; interviewee) — first wiki mention; promote on second-source mention. Nicholas Thompson (CEO, The Atlantic; host/channel author) — first wiki mention; promote on second-source mention.

Source quality

Auto-generated (ASR) captions. Content is branded/sponsored content: produced by Atlantic Re:think (The Atlantic’s creative marketing studio) in collaboration with PwC — Priest is speaking about his own employer’s practice on a platform co-produced with that employer. Per §Lifecycle confidence rules, treat firm-specific claims here (the Southwest Airlines case study numbers, the ~5,000-person hiring figure, the 85–90% agent accuracy rate) as directionally credible but not independently verified — no linked underlying methodology or published report was cited in the interview itself. The trust-rate Stanford study Priest cites is referenced by description only, not linked or titled — a data gap if the wiki wants to verify or cite that claim directly in the future.