Most posts about personal AI systems show the output. This note is about the bill: the hours, the services, and what broke. Everything here comes from my own files, logs and posts. Where I don't have a number, I say so.
| Biggest cost | My time, mostly spent on the knowledge base, not on prompts or code |
| Logged agent work | About 1,400 finished tasks, about 610 hours of run time, June to September 2026 |
| Money | A frontier model plus several API services. No file records the total |
| Payback | Negative for a long time. Positive now on volume work, still not on voice |
Time
What I said in July 2025
My first public post on NicAI was honest about the maths:
- Hours invested > hours saved, for now.
- Time saving per task ranged from minus 100% (the task took far longer than doing it myself) to 30-80%. Nothing came close to 100%.
- 80% of the effort went into building and grooming the knowledge base: everything I know, do and decide, in Markdown and JSON.
- Another large share went into testing workflows and models, such as RAG setups.
That was after months of coding at night. By then the system had gone through many versions, from a first working setup in early 2024 on OpenAI APIs, my own vector database and a lot of Python scripts, to a single script in mid-2025 that sent each task to an OpenAI model with a hosted vector store of my files. My legacy Python folder still holds over 40 of those projects.
What the logs show in 2026
Since early June 2026 every agent task ends with a line in a TASKS.md file: what it did, when it finished, how long it ran. Across the 4 task files:
| Measure | Value |
|---|---|
| Finished tasks | 1,427 |
| Tasks with a logged duration | 1,390, from 12th June to 23rd September 2026 |
| Total run time | About 610 hours |
| Median task | 12 minutes |
| Tasks of 15 minutes or less | 824 |
| Tasks of 1 hour or more | 124 |
| Busiest month | July 2026, about 250 hours |
2 caveats. This is agent run time, not my time: agents run in parallel, and some long tasks include overnight runs or pauses for rate limits. And it doesn't log my own hours building skills, reviewing drafts or fixing what an agent got wrong. I have never tracked those.
By May 2026, a Saturday with 6 agents in parallel delivered what would have taken me weeks (see Working with agents). So the balance has turned on volume work. It hasn't turned on writing in my voice, where every draft still gets my edit (see The NicAI voice stack).
Money
No file in NicAI records what it costs per month, so this section names the costs and leaves out amounts.
| Cost | What it pays for |
|---|---|
| Frontier model | Claude Code, which runs almost everything today. Earlier, OpenAI APIs |
| Research APIs | Perplexity, xAI Grok, Firecrawl, Cloudflare page rendering, Apollo company data |
| Media | ElevenLabs for speech, ImageKit for image hosting |
| Hardware | A Mac Studio I already owned, for local models and nightly jobs |
The credit-based APIs are where money shows up as a failure. In early July 2026 the web-scraping credits ran out mid-batch, and agents fell back to other search tools.
What failed
Limits kill parallel work
In early July 2026, a session usage limit killed 22 parallel agents in the middle of a council run (see The Council ). Only 1 finished. The batch survived because each agent had already written a self-contained build script to disk, so the rest finished as plain Python with no model. On another batch the same limit forced a lean run, with self-review instead of a full panel. A long build in September survived 4 rate-limit stops the same way, on fragments saved to disk.
Rules go stale
Skills are instructions, and instructions age. One false claim, that a logo service was dead, spread into 5 skills, 2 agent briefs and a script before I removed it everywhere on 19th September 2026. Duplicated data aged the same way: a second copy of my company database, kept in sync by a nightly job, ran until I cut it back to one file in July 2026.
Big files don't work as knowledge
A single 12 MB knowledge-base file was the source for product answers until July 2026. I retired it for a wiki of small linked pages that agents rebuild every night (see LLM Wiki). A small page is easier for an agent to find, cite and update than a paragraph in a giant file.
Too many near-duplicate skills
A separate PowerPoint deck skill lived next to the HTML deck skill until 4th August 2026, when I merged them. 2 skills for one job meant 2 sets of rules to keep in step.
Unclear briefs cost whole batches
One batch of outreach briefs went through 5 archived versions. Version 4 was superseded the same morning it was made, by 4 new instructions.
Long sessions drift
The longer a session runs, the worse the output gets, well before the context window is full. I wrote up the problem in Context Rot and the Re-fresh Skill. Each task now gets its own folder with a short _CONTEXT.md, so a fresh session can pick up the work.
Some automation isn't worth it
In May 2026 I stopped automating outreach. NicAI prepares the brief and the draft. I send it by hand.
Lessons
- Most of the work is data, not AI. In 2025 I put it at 80% of the effort. Documented knowledge is the part that carries over to whatever model comes next.
- Log every task. A one-line entry with a duration costs nothing, and it gives the only real numbers in this note. Log your own hours too, which I didn't.
- Design for interruption. Limits, credits and crashes will stop long runs. Agents that write their progress to disk survive them.
- Plan for fallbacks. Credits run out mid-batch. Give the agent a second route.
- Small skills, one job each, one source of truth. Merge duplicates early. Delete stale rules as soon as they're wrong, everywhere they appear.
- Brief once, fully. A batch redone 5 times costs more than a complete brief before the first run.
- Measure the edit, not the draft. The 169 draft-vs-sent pairs show exactly what I change. A score only says that something is off.
- Keep the human steps you care about. Outreach goes out by hand, and nothing leaves in my name without my review.
Gaps in these numbers
- My own hours on NicAI aren't logged anywhere, so the real time cost is higher than any figure here.
- The service costs aren't recorded in one place, so there is no monthly total.
- Local models on the Mac Studio are still experiments, so almost all run time is paid cloud time. Moving private work to local models is on the list in What is NicAI.