Vendor math · AI productivity
Six people, six months, one person, one week: breaking down Replit's self-driving company claims
Replit's manifesto says a customer compressed 36 person-months of work into a single person-week. The arithmetic checks out.
The most-shared number in AI this week comes from a sentence in Replit's October 26 manifesto, "The Rise of the Self-Driving Company." Quoting Yoshiyuki Fukuma, general manager of digital transformation at Sumitomo Mitsui Finance and Leasing, it reads:
Coffee Cup AI evaluates claims like this alongside practical AI consulting decisions and the wider evidence on AI job loss.
"A development process that previously required six people working for six months was completed by just one person in one week."Yoshiyuki Fukuma, GM of Digital Transformation, SMFL, quoted in Replit's manifesto
Replit repeated the claim this morning in the press release opening its Tokyo office, where SMFL is the anchor enterprise customer. The number is doing real commercial work.
The multiplication, first
Six people for six months is 36 person-months of labor. One person for one week is a quarter of a person-month. Divided, that is a 144x compression on a four-week month, or roughly 156x if you count real weeks. Call it about 150x, which makes it one of the largest productivity claims ever attached to a named enterprise customer in this wave of AI. The arithmetic is sound; the question is what is inside the number.
Before
After
Three multipliers stacked into one headline
Measured speedups in raw code production top out far below 150x. What makes the number plausible as a statement about elapsed effort is that it multiplies three different effects, and only one of them is AI writing faster.
The agent writes the code faster than a person would type it. Independent measurement puts this multiplier between roughly 0.8x and 3x depending on the task and codebase (Figure 4).
Six people carry 15 communication pairs. One person carries zero. SMFL's own case study describes the old process as "back and forth, siloed development, and repetitive discussions with user departments." Those months were meetings, requirements loops, and waiting. One person with the whole picture spends none of it. This multiplier comes from organizational design.
SMFL's January 2026 case study frames the Replit work as "dynamic prototyping" to make tacit requirements visible. The October retellings call it a "development process" and a "project." A one-week build at a financial lessor still needs security review, compliance, integration, and operations before it runs the business. Nobody has said the week-built artifact is in production.
Credit where due: Replit CEO Amjad Masad is making the coordination argument openly. His manifesto opens with Coase and closes by saying "the biggest economic impact of AI will come from this reinvention of the firm." In that reading, the SMFL story is interesting. It says a company can hand an entire problem to one person who holds the context, and skip the apparatus that used to translate one person's knowledge into six people's work.
The version circulating on stage at TechCrunch Disrupt and in today's press release says "AI made one person 150 times more productive." Those two statements lead a buyer to very different decisions.
The execution multiplier, measured independently
The execution multiplier is the only part of the stack that can be tested, and the testing has been done. The best evidence is a randomized trial from METR published in July 2025: 16 experienced developers, 246 real tasks on mature open-source codebases, tasks randomly assigned to allow or forbid AI tools.
METR's result has limits, and they matter: experienced developers, mature codebases, early-2025 models. The study does not say AI fails on greenfield work. On the contrary, the clearest positive result in the literature, GitHub's Copilot experiment, found a 55% speedup, on a simple task built from scratch by developers with no existing code to navigate. SMFL's build looks exactly like the second case: one strong operator, fresh ground, no legacy code in the way.
The larger-scale picture agrees. The 2025 DORA report, surveying nearly 5,000 technology professionals, found that AI acts as an amplifier: throughput rises, but delivery instability rises with it, and the hoped-for "fail fast, fix fast" compensation never materialized. Nearly 90% of respondents use AI; 46% distrust its accuracy. An organization with weak processes does not get the SMFL story; it gets faster versions of its existing problems.
The fine print on the 150x
The stacked-multiplier structure is not the only problem with the claim.
- It is a testimonial inside a sales document. Fukuma's quote appears in Replit's own launch manifesto, was repeated in Replit's own press release, and now travels with a partnership in which SMFL and Replit sell together across Asia. Nobody independent has seen the underlying data. As The New Stack's Amanda Caswell wrote of Replit's numbers, the company "hasn't shared enough detail to independently verify the gains or separate the impact of agents from other changes to how its teams work."
- It is one anecdote, selected. There is no median SMFL project, no failed case, no count of how many Replit projects at SMFL looked like this, so we are seeing the one project that worked.
- The comparison never defines its units. The manifesto does not say what the six people did for six months (requirements, design, build, test, deploy) or what the one person's week produced (a prototype, a pilot, production code), and that answer decides what the 150x refers to. A prototype that skips the hard 90% of enterprise software is a different deliverable from what those six months bought.
- Replit's supporting numbers have the same shape. Engineers "nearly tripled code output" in the first half of 2026, but "code output" is undefined, and Masad told Platformer his engineers now "only look at the code when it's about to get deployed," meaning the review process changed at the same time as the metric. The 51% token reduction and the 16-minute-to-1-minute task come from Replit's own testing of an early-access feature. Customer spend doubling at six months and quadrupling at a year is attributed as "reportedly," and none of it is audited.
| Claim | Where it appears | Who measured | Verification |
|---|---|---|---|
| 6 people, 6 months done by 1 person in 1 week (~150x) | Replit manifesto; Tokyo press release | SMFL GM, quoted by Replit | Self-reported testimonial, unaudited |
| Engineers nearly tripled code output in H1 2026, quality metrics flat | Manifesto, Disrupt, Platformer interview | Replit, internally | Not independently verified; "code output" undefined; review process changed concurrently |
| 51% drop in token consumption; routine data task from 16 minutes to 1 | The New Stack, from Replit's own testing | Replit, early-access feature | Self-tested; no third-party replication |
| Enterprise share of revenue grew from ~2% (2025) to ~15% (2026) | Crypto Briefing launch coverage | Replit statements | Plausible, unaudited |
| Customer spend doubles after 6 months, quadruples after a year | Crypto Briefing launch coverage | "Reportedly," per coverage | No methodology published |
What a normal company should plan for instead
Strip the 150x down to what the evidence supports, and the planning ranges for an ordinary company look like this. These are our assessment from the studies cited above, offered as ranges because your shop will differ:
- A greenfield, well-scoped build with one strong operator: 2x to 5x compression of build-phase effort is plausible. The SMFL case shows the ceiling can be higher when you also eliminate coordination and narrow the scope. Anything pitched above 10x should be treated as marketing until you can see the units on both sides of the comparison.
- Work on mature systems: low single-digit gains, and possibly a short-term slowdown while people learn the tools. METR's developers were slower precisely because they knew their codebase well enough to spot what the AI got wrong.
- The biggest honest lever is coordination. If your process spends months in requirements loops and handoffs, giving one person the authority and tooling to show the team a working artifact in week one can collapse the calendar even if the typing only gets twice as fast.
Pick the process you would want to compress. Write down how many people touch it, how long it takes, and what the deliverable is, before any AI touches it. Run the tool for 60 to 90 days and measure the same things again. Since you took the baseline yourself, it is the one number in this discussion you can verify.
The arithmetic says to budget from the numbers you measured yourself, and to treat the vendor's as claims until you can check them.