Matched cohorts
We picked 14 projects shipped in 2025–26 and matched each one against a 2022 project of comparable scope: same product type (web app, mobile app, AI integration), similar team size, and equivalent functional surface area.
Speed Report · Q1 2026
How we measured the 2.1× speed claim
Every agency claims to ship faster with AI. We're the only one who publishes the methodology behind the number. Here's how we got to 2.1×: what we measured, what we didn't, and where the data is honestly thin.
Last updated 2026-04-15Cohort 14 matched projectsReviewer M. Carter, ex-Stripe EM
0.0×
faster than our 2022 pre-AI baseline, normalized for scope
The headline numbers
0.0×
faster than our 2022 pre-AI baseline
0 days
median time to MVP launch (2025–26)
0 days
median time to MVP launch (2022)
0
matched projects in the comparison
What we measured
The definition of 'shipping faster'. Without the marketing.
A 2.1× claim only means something if you know what was being timed. Here's the boring, useful definition. Speed = days from signed scope to a production deploy that the client uses, divided by normalized scope units.
Time tracking from day one
Every hour is logged in Harvest against a phase: discovery, design, frontend, backend, AI integration, QA, deploy. We do not count internal review meetings against client time. Nothing is back-filled. Entries land within 24h or are dropped.
Normalization
Raw hours are normalized by functional surface area. We count screens, models, integrations, and AI capabilities as scoped units. A 12-screen app is not compared 1:1 with a 30-screen app, even within the same cohort.
Reproducibility
The full anonymized dataset (project IDs, scope units, hours per phase) is available on request under NDA. Methodology is reviewed by a third-party engineering manager outside the company once per quarter.
The raw numbers
Phase-by-phase, 2022 vs 2025–26
Median calendar days per phase across the 14 matched cohort pairs. Normalized to a 'standard' MVP scope of 18 screens, 6 data models, 2 integrations.
| Phase | 2022 baseline | 2025–26 (with AI) | Speedup |
|---|---|---|---|
| Discovery & scope | 11 days | 6 days | 1.8× |
| Design & prototyping | 14 days | 5 days | 2.8× |
| Frontend implementation | 21 days | 9 days | 2.3× |
| Backend & data | 18 days | 8 days | 2.3× |
| AI integration | — | 4 days | n/a |
| QA & hardening | 12 days | 5 days | 2.4× |
| Launch & handover | 6 days | 3 days | 2.0× |
The 2.1× headline is the geometric mean of the six phases that exist in both cohorts. AI integration is excluded from the comparison because the 2022 cohort had nothing equivalent. Adding it to the average would inflate the number without being honest about it.
Where AI helps, and where it doesn't
The honest split: 5 wins, 5 limits
If we tried to use AI for everything, we'd ship faster and worse. The number above only holds because we know which phases AI accelerates and which it actively harms.
Where AI accelerated the 2.1×
- Boilerplate scaffolding (routes, types, forms, schemas) collapses from hours to minutes.
- Design exploration: 8–12 viable directions in a day instead of two over a week.
- Test generation: 70%+ unit coverage emerges in parallel with the feature, not after.
- Migration scripts and data transforms: drafted, reviewed, run in a single sitting.
- Documentation stays current because it's regenerated from the diff, not written by hand.
Where we kept humans in the loop
- Architecture decisions and trade-off conversations: humans only, no AI in the room.
- Production deploys, secrets, infra changes: AI can suggest, only an engineer executes.
- Anything user-facing that names you, your customers, or your numbers: written by humans.
- Critical path debugging where being wrong is expensive: AI is a second opinion, not a first.
- Code review on the PR that ships: always a senior engineer, never an automated gate alone.
What we don't claim
Caveats, in plain language
Every speed claim has weak spots. Here are ours, written out before you have to ask.
Sample size is honest, not large
14 matched pairs is enough to show a real trend, not enough to publish a paper. We update this report every quarter as more projects close.
Selection bias exists
Projects we ship are projects we accepted. Teams that came to us in 2025–26 may have been better-prepared than 2022 clients. We control for scope but not for client readiness.
AI integration projects skew the average
The 'AI integration' phase has no 2022 equivalent. We excluded it from the 2.1× headline. The number is built from the six phases that exist in both cohorts.
We don't count time we didn't bill
If a senior engineer spent a Saturday thinking about your problem, that hour is not in the dataset. The number reflects billable delivery time, not total cognitive load.
Want the dataset?
We'll send the anonymized cohort data under NDA
Project IDs, scope units, raw hours per phase, and the normalization formula. If you're evaluating us against another agency, this is what you compare with.