The shift
How AI rewrites the economics of building software. Five sections of evidence, not vibes, and a note wherever the evidence runs out.
The argument runs in five steps. Teams collapse into soloists. Sprints stretch into streams. Meetings are replaced by messages. The cache makes a long context affordable. And the work compounds, because every project leaves tools behind for the next one. Each section makes one of those claims, shows the figure behind it, and says what that figure cannot carry.
01From teams to soloists
10 roles worth of surface area, covered by 1.
AI doesn’t just make you faster at what you already know. It makes adjacent domains accessible: security, DevOps, systems programming. Domains that used to need a dedicated hire become reachable with the right co-pilot.
The ratings below are keyword coverage over project notes, plan files and the technologies named in about the last 30 days of session logs, so they show breadth of work touched rather than depth of expertise. A high score says the month was spent in that vocabulary. It does not certify a specialist.
lowerhigher
ai-pilot-data.json, instrumentRatings, 99 session logs sampled, about the last 30 days, generated 2026-10-02
The seats the cost model prices for the same ground:
- Full-Stack Developer x2
- Data Engineer x2
- DevOps Engineer
- Frontend Developer
- ML/AI Engineer
- QA Engineer
- Project Manager
- Biz/Workflow SME + UI Builder, half time
02From sprints to streams
The sprint model assumes work arrives in discrete batches. AI-assisted development is continuous: deploy when ready, iterate in real time, no ceremony between idea and production.
The pilot record draws that as a calendar. Sessions ran on 187 of the 209 days it covers, and on every day since 28 May 2026. All 22 blank days fall before 7 July 2026, when the record began to be kept as it happened. Before then it holds only the logs still on disk that day, so a blank there can be a pruned log as easily as a day off.
One cell is one UTC day, 2026-03-08 to 2026-10-0201 to 456457 to 3,8373,838 to 7,0767,077 to 10,19010,191 to 21,347records
| Month | records | Active days | Days in window |
|---|---|---|---|
| 2026-03 | 1,846 | 15 | 24 |
| 2026-04 | 14,057 | 24 | 30 |
| 2026-05 | 25,226 | 24 | 31 |
| 2026-06 | 161,610 | 30 | 30 |
| 2026-07 | 245,101 | 31 | 31 |
| 2026-08 | 350,793 | 31 | 31 |
| 2026-09 | 288,261 | 30 | 30 |
| 2026-10 | 13,065 | 2 | 2 |
A day is shaded when any session wrote a record to it, which says sessions were running, not that anyone was at a desk.
ai-pilot-data.json, activityHeatmap, 8 March 2026 to 2 October 2026, UTC days, generated 2026-10-02
What this cannot show is the counterfactual. Nobody ran the same project in two-week sprints to compare. The weekly commit and issue counts for the case-study window are on cost analysis; they count every author in the organisation’s repositories, so they describe the project, not one person’s cadence.
03From meetings to messages
1,099,959 transcript records across 336 sessions and 102 projects. The coordination happened in the conversation, not on a calendar, and the record can only show the conversation half of that.
Totals cover the period since 8 March 2026, through 2 October 2026, read from the durable activity record. That record has been kept since 7 July 2026. From before then it holds only the sessions whose logs were still on disk, so the earlier months are incomplete.
An earlier version of this page set that count beside the words “0 meetings.” The count is real. The zero was a flourish: nothing in the record knows what was on a calendar. The same goes for the split of a working week. The cost model assumes a developer on a conventional team spends 35% of it writing code, 20% in meetings and 12% in review. There is no such split for the recorded side, because nobody logged one. What the record does have is when the work happened.
Three shares are written into the model. The remaining 33% is not broken down there, so it is not broken down here: its row is an outline, not a measurement.
cost-model.json, industryBenchmarks: assumptions written into the model, not measurements
| Bucket | records |
|---|---|
| 00:00 | 33,100 |
| 01:00 | 21,775 |
| 02:00 | 10,295 |
| 03:00 | 4,431 |
| 04:00 | 1,663 |
| 05:00 | 1,856 |
| 06:00 | 3,021 |
| 07:00 | 5,696 |
| 08:00 | 30,188 |
| 09:00 | 44,491 |
| 10:00 | 58,484 |
| 11:00 | 66,418 |
| 12:00 | 72,471 |
| 13:00 | 78,745 |
| 14:00 | 87,408 |
| 15:00 | 91,167 |
| 16:00 | 77,478 |
| 17:00 | 75,835 |
| 18:00 | 60,308 |
| 19:00 | 50,127 |
| 20:00 | 53,322 |
| 21:00 | 57,416 |
| 22:00 | 63,359 |
| 23:00 | 50,905 |
47.6% of the records fall outside 09:00 to 17:00. A record is a prompt, a tool result or one block of a model reply, so this is when sessions were running, not when anyone was typing.
ai-pilot-data.json, hourlyDistribution, 8 March 2026 to 2 October 2026, hours in America/Detroit, generated 2026-10-02
04The cache effect
98.4% of input came from cache. That is the whole reason a long-running context is affordable: re-reading the project each turn is what the cache is paying for.
An earlier version of this page called that instant onboarding, and said the model retains the context of prior work. It doesn’t. The model keeps nothing between turns: every turn it is handed the conversation and the files again, and the cache is what makes handing them over affordable. It is closer to say that onboarding happens on every turn. The conversation is the knowledge base.
120.3B tokens read from cache, 1.9B written to it and 13.1M fresh. The rows are drawn to scale: the write to cache is a sliver, and fresh input, about 146 times smaller again, is the two-pixel mark. Output over the same period was 321.9M tokens. Tokens used by subagents are not in the record.
ai-pilot-data.json, tokenEconomy, 8 March 2026 to 2 October 2026, generated 2026-10-02
05Compound velocity
102 projects aren’t 102 isolated efforts. They’re an interconnected ecosystem, and the connection has a shape.
Sessions generate code
1,099,959 transcript records of building, in 336 sessions.
Code feeds pipelines
Some of that code is scripts that count the work itself: commits, sessions, tokens.
Pipelines power dashboards
The pilot record, the work timelines and this page read what those scripts write.
Dashboards surface insights
A number on a page gets looked at. A number in a log file does not.
Insights accelerate building
The next session starts from what the last one found, tools included.
05 feeds 01Each pass starts from what the last one built.
Tools built for one project get used by the next. That is the part that compounds, and it is also the part this page cannot measure: there is no counter for a script that saved an afternoon six months later.
The one turn of the loop I can document is a correction. The cost model behind this page used to regenerate every four hours, and as the session logs aged out from under it the numbers decayed until the file held one active day and a 1,710x velocity multiplier, and the page was printing “9 months to 1 days.” That was caught the plain way, by reading the page. The model is now frozen as a dated case study, 6 of its figures carry a written correction, and cost analysis says which of its own ratios it will not headline.
What it adds up to
The old model isn’t broken. It’s just no longer the only option. AI doesn’t replace developers. It changes the ratio.
How far the ratio moves is the part one case study cannot settle. The cost comparison leans on a team that was never hired, the session record it was built from only begins 2026-02-09, and two of the three ratios the pipeline computed are ones this site declines to headline. What is left is still worth reading: the curve and the weekly counts are on cost analysis, and the whole session record is on the pilot record.
cost-model.json, frozen case study of 2025-12-01 to 2026-03-26, recorded 2026-03-26; ai-pilot-data.json, generated 2026-10-02