Time back. The first AI install most teams should run.
by Luis Gomes, Founder & Growth Lead
If the four-jobs piece introduced the four jobs as a menu, this is the deep-dive on the one most teams should pick first. Time back is the install with the cleanest math. It is also the one where the team feels the change in the first week, which matters more than the math when an owner is deciding whether AI is real for their company or theater.
Time back is exactly what it sounds like. The work that takes the senior person on the team hours every week, automated. Not eliminated — automated, with the senior person still owning the judgment calls and approving the output. The hours go back. The senior person uses them on the work you actually hired them to do.
What time back is, in practice
The work that gets automated is the repetitive output that has a stable shape week to week. The Monday report. The standup write-up. The weekly client update. The recurring Slack post that summarizes what the team did. The status update the founder writes to the team. The end-of-month financials commentary. The sales pipeline cleanup. The CRM data hygiene pass.
These are the tasks that have three things in common. They follow a stable template. Their inputs come from systems we already have (analytics, CRM, calendar, project tracker, repo, support ticketing). The senior person does them because nobody else can produce something that sounds right, even though nothing in the work actually requires the senior person's judgment.
That third part is the giveaway. If the senior person is doing it because nobody else can match the voice or the structure, the job is automatable. AI can match a voice and a structure. AI cannot make the judgment call that goes inside the structure — the senior person still does that part. But the assembly is done before the senior person starts.
What the math looks like
Take the marketing operator who spends six hours a week assembling the Monday report. Three hours pulling numbers from analytics. Two hours stitching them into the standard template. One hour writing the commentary at the top.
After install: the agent pulls numbers from the same systems on Sunday night. The agent assembles the standard template. The agent writes a draft commentary based on the numbers and what was in last week's report. Monday morning the operator reads the draft, fixes the commentary, signs off, and the report is sent.
Six hours becomes one. Five hours back. Per week. That is roughly two days a month, returned to a single senior operator who is the most expensive person on a small team. Across a small team of five, doing similar weekly pulls, that is around two weeks of senior capacity returned every month.
The math does not depend on a heroic AI step. The hard part of the install is the part nobody outside the team can do — capturing what "the way we write our Monday report" actually means. Which numbers go where. Which trends get called out. Which voice the commentary uses. Which things never get said. The agent runs on top of that capture.
What 30 days looks like
Week 1 is scoping. Sit with the senior person. Watch them do the task once. Capture the inputs (which systems, which queries, which fields). Capture the outputs (what the finished thing looks like, including the things they would never write). Capture the rules — the things they reach for when something looks off, the rules of thumb they apply, the sentences they would and would not say.
Week 2 is install. Connect the agent to the source systems through MCP-style connectors. Write the methodology document — the CLAUDE.md for this specific task. Generate the first draft. Sit with the operator while they edit it.
Week 3 is iteration. Run the task three times in a row. Capture every edit the operator makes. Fold the edits back into the methodology document. By the end of week 3 the operator should be editing less than they would have written from scratch.
Week 4 is hand-off. The operator runs the task through the system without anyone else in the room. We measure: how many minutes did this take, what fraction of the output needed an edit, did anything ship that should not have. Number is the number. If the install paid back, we pick the next job.
Where it goes wrong
Two failure modes. Both are predictable.
The first is when nobody captures the rules — when the team installs an agent and gives it the source systems but skips the methodology document. The agent generates fluent output that does not sound like the company. The operator spends the same hours editing it as they would have spent writing it. The install fails on a measurable signal: edit time per output. We watch this number; if it does not drop in week 3, we go back to the methodology document, not the model.
The second is when the input systems are too messy for an agent to read. CRM with five years of inconsistent fields, analytics with broken event names, project tracker with conventions that change quarterly. The agent can work around messiness, but only up to a point, and the work-around is fragile. In practice we find one of two things: either the team needs to fix one or two pipes before the install works, or we scope a smaller version of the task that uses only the clean systems.
What you get back when it works
You do not get back a marketer. You get back the marketer's most repetitive five hours a week. The marketer's other 35 hours are the same as before — except they are not also doing the report assembly anymore.
The compounding is two-step. Step one: the senior person uses those five hours on the work you actually hired them for. Step two: that work makes the next install easier, because the team has now seen what an AI install looks like and the resistance to the next one is lower.
Most teams that pick this job and run it well do a second install within the next 60 days. By then it has stopped feeling like an experiment.
What we install
We pick the task with you. Usually the one your senior operator named when you asked them what work they hate doing. We capture the inputs, the outputs, and the rules. We connect the systems. We sit with the team while they use the install for the first three runs. We hand it off. We measure what changed.
The install is bounded. 30 days. One job. One number. If it paid back, you pick the next box. If it did not, we said we would tell you, and we tell you.