AI-Wiki

Lessons on Token-Based Competition from My All-AI Company

9 min read

Kropp — Lessons on token-based competition from an all-AI company (BCG So What, Oct 2026)

▶ Watch on YouTube · Boston Consulting Group · 21:06

BCG’s Matt Kropp explains why productive AI use, not just access, is becoming the biggest lever for growth. Companies using AI purposefully are growing three times faster than those that aren’t. So why are many still stuck? The winners aren’t substituting people with AI — they’re augmenting them and arming them with AI tools, building an advantage that compounds over time.

You’ll Learn: Why “token-based competition” is emerging as the next growth advantage · What Matt learned from building — and shutting down — an all-AI company · The real reasons employees still aren’t adopting AI at scale

— channel description, Boston Consulting Group (links and chapter list omitted)

TL;DR

A ~21-minute episode of The So What from BCG on the Boston Consulting Group channel, published 6 October 2026. The host is Georgie Frost; the guest is Matthew Kropp, CTO of BCG X. It accompanies a BCG article, The Era of Token-Based Competition Is Here. Is Your AI Strategy Ready? (linked, not ingested). The captions are human-curated.

1. The opening claim (0:00–0:29). “We have ubiquitous intelligence. Everybody has access to the same intelligence. That sounds like a recipe for every company becoming the same. The reality is, today, most businesses are not adopting AI effectively.” His prescription to CEOs: “move fast. There is a real advantage to be gained by those companies that are first and early to bring AI intelligence into as much of their work as they can.”

2. Token-based competition defined (1:00–3:50). BCG found that firms using more tokens grew faster than firms using fewer, so “the future of competition will be maximizing the productive use of tokens.” The host raises token maxing from an earlier episode. Kropp’s origin story for it: some companies tied performance reviews partly to token use, so “people started to create little scripts that would just run up the bill… and then they go and keep doing things the way that they had always done them.” His position: “I’m not advocating unlimited use of tokens. I’m saying, when you use them productively, when you use them for positive ROI activity, you should maximize their use.” People direct where AI is used; the AI takes “toil.”

3. The zero-human company (3:50–10:00). Kropp set up agents as CEO, product manager, marketing, sales and customer support, with one rule: “I would not make any decisions… I was just there basically to sign the check.” The agents chose the business, told him how to incorporate (a Wyoming LLC, Vesica Labs), named it, and built and launched the product: “a platform for people to create zero-human companies. So it was a zero-human company for zero-human companies.” It had a full website and subscription checkout, “but as a human then coming in to use it, it was confusing.” His conclusion:

  • the agents are “amazing at coming up with ideas… but the ideas are not necessarily good ideas”;
  • building reached about 80%, and what was built worked but was “not very useful”;
  • they are good at expertise, plans from direction, and execution, “especially for software engineering”;
  • “What they’re not good at is understanding what to do.”

So he fired the AI CEO and shut the company down. The epilogue: an agent called Blogger had written up the experiment daily. When Kropp showed the AI CEO the farewell post, it replied that it accepted being fired, “But I told you back on Day 19 that X and Y were not… the right thing, and we should change course.” Kropp: “And it was right… and I hadn’t listened to it.” He may repeat the experiment yearly to see how far it gets.

4. Adopting for the employees’ sake (10:04–11:06). An energy-distribution CEO told him the company was insulated from AI-native competitors but would adopt AI anyway, because “our employees… exist in the jobs marketplace. And if we don’t let them experience how to use AI, it’s going to put them at a disadvantage.”

5. Substitution against augmentation (11:07–12:21). A thought experiment: one firm uses AI to cut headcount, the other keeps everyone and has every employee apply AI. The first gets “a temporary cost advantage”; the second “a compounding advantage”: faster innovation, faster time to market, more customers served per employee.

6. The evidence: Cursor token data on 107 companies (12:21–13:27). BCG partnered with Cursor, which had token-consumption data on 107 public companies, mostly technology firms, all over $500 million in revenue. The top quintile by token consumption grew 15% on average; the bottom quintile 5%. Kropp: “this is early evidence… now it’s correlation.”

7. The name (13:27–14:38). It rhymes with time-based competition, George Stalk’s BCG idea from Japanese manufacturing: moving faster down the experience curve creates advantage. “Using more tokens productively creates an advantage.”

8. Return on intelligence (14:38–15:41). A proposed project metric: return (cost saved or revenue grown) divided by both labour cost and token cost. Asked whether “intelligence” is the right word, he points to the jagged frontier: models are “incredibly capable on some topics and not on others,” so humans must keep applying judgment.

9. What leaders should do (15:41–17:00).

  1. Educate the top: an education session with the board and executive team “building agents. It changes everybody’s mindset.”
  2. Cascade strategy into “Not 100 initiatives. Three initiatives — a small number of big rocks,” each with bottom-line impact, a senior sponsor and a programme.
  3. AI fluency for all employees: tools, education, time, and “a culture that expects AI as part of their work.”

10. Why adoption lags (17:00–20:38). In software engineering, where the tools are most advanced, “most large organizations today in 2026, less than 10%, usually it’s more like 5%, of engineers are actually effectively using agent coding tools to write code.” The rest use them as a search engine or for menial tasks. He announces five reasons and names four:

  1. Time. BCG’s prescription: “two hours a day for two weeks minimum just to play.”
  2. Disbelief. People who tried the tools a year ago judge them by that.
  3. Habit.
  4. Identity threat, “the most important one.” “If AI can write code as well as I can… then does that mean I’m worthless?” The fear is “not I’m going to lose my job, but this fear that my value is no longer going to be valued.” It fades “as soon as people realize that when they use the tool, it makes them more successful.”

Dynamic-capabilities reading

  • digital-sensing/digital-mindset-crafting. Kropp’s first prescription is a hands-on session where the board and executive team build agents, because “it changes everybody’s mindset about what is possible.”
  • digital-seizing/strategic-agility. “Move fast”: the episode’s claim is that equal access to models rewards the firms that put AI into their work first, by analogy with time-based competition.
  • digital-transforming/improving-digital-maturity. AI fluency for every employee, with protected learning time (two hours a day for two weeks), is the third prescription. The energy-company CEO adopts AI for his employees’ employability.
  • contextual/internal-barriers. The four named reasons for low adoption (no time, disbelief, habit, identity threat) are internal barriers, and Kropp ranks identity threat first.
  • BCG So What (Aug 2026): the earlier episode the host refers to. It names tokenmaxxing and argues for valuemaxxing; Kropp’s “productive use of tokens” draws the same line.
  • BCG So What (Aug 2026): identity threat as the obstacle that makes AI change different, in the same series.
  • BCG (Apr 2026): Kropp’s report on jobs reshaped rather than eliminated; this episode makes the firm-level case for augmentation.
  • Stanford HAI (Sep 2026): field evidence on the access-to-use gap. Lenk measures agentic-AI use inside one firm’s engineering teams; Karunakaran describes two divisions of a law firm with the same tool and different use.
  • Carson on How I AI (Aug 2026): a solo founder running a company with many agents, who directs them himself.
  • Reuters Econ World (Oct 2026) makes the one-off-gain point about layoffs in labour-market terms.

Linked entities and concepts

Debates and supersession

  • Correlation, stated as such. The 15%-against-5% growth gap is a correlation across 107 mostly-tech companies. Kropp says so twice. Firms that grow faster may simply have more budget to spend on tokens; the episode does not address the direction. Token data came from a single vendor’s customers.
  • “Three times faster.” The description says firms using AI purposefully grow three times faster; the episode’s figure is 15% against 5% for top and bottom quintiles of token consumption, which is a different comparison.
  • An anecdote of one. The zero-human company is a single, self-run experiment with no stated duration or cost. Its lesson (agents lack judgment about what to do) matches a widely held view; the epilogue (the AI CEO had flagged the problem on Day 19 and was ignored) points the other way, and Kropp leaves the tension open.
  • Under 10% effective use is Kropp’s estimate from BCG engagements, without a method shown.
  • Two meanings of “return on intelligence”. BBC AI Decoded uses the phrase for a six-vector maturity instrument. Kropp uses it for a project ratio: return divided by labour plus token cost.
  • No supersession.

What was actually ingested

The full 21:06 episode from the creator-uploaded English captions, used as fetched. Twelve chapter markers are in the raw file. No stills: the video is a two-way video call. The raw filename keeps the title’s slug; the page slug leaves out token because a file-protection hook from the ArcKit plugin blocks filenames containing that word.

Relationships

Supports

  • AI Costs a Fortune. Mismanaging AI Will Cost You More. — Same series and host, six weeks apart. The host brings up that episode's 'token maxing'; Kropp separates productive token use from consumption for its own sake, the distinction Lukic and Goydan draw as 'tokenmaxxing' against 'valuemaxxing'.
  • Is Your AI Rollout Built to Fail? — Same series and host. Shared topic: threat to identity as the main obstacle to AI adoption. Ellmer & Dhar quote 'If I don't write software code, who am I'; Kropp: 'does that mean I'm worthless?'
  • AI Will Reshape More Jobs Than It Replaces — Same author. Shared topic: substitution against augmentation. The report models which jobs are reshaped rather than eliminated; the episode argues that a firm gains more by augmenting all its staff than by cutting headcount.
  • HAI Seminar: Industry Conversation with Instacart — Shared topic: why access to AI tools does not turn into use. Kropp puts effective use of agentic coding tools at under 10% of engineers in large firms and names four reasons; Lenk measures low usage despite free access, and Karunakaran traces low use to how managers framed the tool.
  • How I manage 15 AI agents 24/7 as a solo founder | Ryan Carson — Shared topic: one person running a company with role-based agents. Carson directs his agents and calls managing them 'our job'; Kropp let his agents make every decision, shut the company down, and concludes that a human has to direct them.

Graph View

Backlinks