How it works
You run a business.
Install it, work normally, get your report.
Nobody changes how they work. fda watches every working day of your team’s real work, then shows where it goes.
- 01Install the agentDeployed in an afternoon, light on CPU. Nobody changes how they work.Day 0
- 02Work normally, two weeksIt learns the real job: the tools, the loops, the switching nobody logs.Days 1–14
- 03Get your reportWhere the work actually goes, and what a teammate could take over. Yours to keep.Day 14
fda watches and quantifies
ranks each finding by how often, how many people and business impact
fda’s specialist agents queue the implementation plan
you press go; version one ships within 24 hours
Side by side with your team.
We build it with the people whose work it learned, sitting alongside them. It ships after the two-week scoping period, trained on your workflows.
fda learns the work. Your AI teammate does it.
Example · trained on every working day of your team’s real work
Remove monotony from your day-to-day.
Shortcut rows show key combinations only, never what was typed
| # | Task | Evidence | Repeats | Build |
|---|---|---|---|---|
| 01 | Deal desk → invoice handoff | 6 people | ×118 | low |
| 02 | Salesforce ↔ Notion field sync | 4 people | ×47 | medium |
| 03 | Weekly status report assembly | Customer success | ×63 | medium |
| 04 | Expense report reconciliation | Finance | ×12 | low |
We put fda on 4 of our own machines.
Nobody changed how they worked. Three weeks later, here is what it found. Every data point, click through all 5 pages.
25%
Of the team’s work was repeated
A fifth person’s worth, spread across four jobs
~25,500
App switches analysed
4 people · 3 weeks
~2,300
Copy-pastes into AI tools
Context carried by hand
We measured ourselves first. It was worse than we thought.
We put our forward deployed agent on 4 of our own machines (recruiting, client success and the founder’s desk) and recorded 3 weeks of real work. Every application, every browser tab, every copy and paste. No surveys.
Four people. Four different problems.
The inbox is the job. 19% of the work in Gmail, 25% in Slack. Runs 1,636 AI sessions and feeds every one by hand.
All-in, entirely by hand. 43% of the work in Claude and ChatGPT, with 748 switches running the same prompt twice.
The untapped seat. 24% of the work on LinkedIn, the lowest AI use on the team, and 528 failed saves retried by hand.
The human API. Reads a source, prompts Claude, pastes the answer back, about 550 times in three weeks.
Four different jobs. The shares are not a performance comparison.
Where the three weeks went
Work a human shouldn’t be doing. The tools don’t talk to each other, so people carry the data between them. None of this is on anyone’s job description, and all of it is measured.
6 moves. A 5th teammate.
- 01Connect AI to the stackMCP connectors putting Gmail, Slack and our ATS inside Claude: scoped, read-only, human-approved.All four4%
- 02Bridge Slack and email, then template bothA Gmail↔Slack bridge and canned replies for the 6–8 recurring message types.All four3%
- 03One prompt, both modelsOne prompt to Claude and GPT, one merged answer with the disagreements flagged.Founder2%
- 04AI drafting inside the inboxClassifies the thread and drafts the reply from thread plus ATS context. They edit and send.Recruiter, client success2%
- 05Fix the ATS write path528 failed saves retried and 428 post-submit waits in 14 days. Validation before submit, safe retries.Sourcer, and every ATS user1%
- 06Sync, schedule and digestList sync between the CRM, the mailer and the ATS; self-booking links; a campaign digest in Slack.All four2%
What the team gets back.
A fifth person’s worth of output from the same 4 people, with no new hires, no new tools, and nobody asked to work harder.
- 01Install fda
- 02Work normally
- 03Get your report
- 04Build your teammate
4 people · 3 weeks · ~25,500 app switches analysed · attain internal case study, July 2026
This is the same report you get after your two weeks. Take the sample report →
Want this report for your own team?
4 people, 3 weeks, every data point. Yours in two weeks.
More desks = better AI teammate.
Every machine you add sharpens the picture. What looks like one person’s quirk turns out to be how the whole company works.
One desk
A habit
How one person works. Easy to dismiss as personal preference.
One team
A pattern
The same workaround on four desks is a process nobody designed.
The company
A system
fda learns how the business actually runs.
Install it on every machine, so don’t be shy. Unlimited seats · Deploys in an afternoonRolling out to a team? Talk to us →
2 weeks of work. 1 new teammate.
Unlimited machines · Metadata only, unless a person chooses to record
Verify the build
macOS · v0.5.15 · 16.3 MB · SHA-256 F4C637BE … 2BAEE5E1
Windows · v0.5.15 · 5.4 MB · SHA-256 CA5C1906 … 7F81A16B
live in production
open one to see how we built it
Attain MCP Use Attain inside your AI app. Set up the connector →