Honest numbers, starting with our own.
Trying AI and getting time back are two different things. Below is our first measured number from a live build, then published studies of what AI does to specific jobs, then the reasons most AI projects never make it into daily use and what we do about each one. Every figure is labeled and linked.
Measured where it is ours. Sourced and linked where it is not.
One client, one office, a quarter to a third of a full-time job.
Our first own number, from agents running on a Florida mechanical contractor's own Mac mini. The counts come straight from the agents' records. The minutes per task are our estimate until the client's staff time them, so we show the full range.
A Florida mechanical contractor
11 to 13 hrs
of office work a week, done by the agents on one client's own Mac mini
Full range 6 to 24 hours a week, depending on how long each task would take by hand.
Measurement continuing
Counted, Two full weeks, September 21 to October 4, 2026
- 48supplier invoices and statements read and sorted, 38 of them filed straight to the right person
- 23bid invitations logged and summarized
- 13subcontractor quotes filed
- quote follow-ups to subcontractors went out on time, without anyone in the office having to chase them
- 40morning briefs and weekly safety posts delivered
- 66staff questions answered
How we got the number
- The counts are measured, taken from the agents' own records, week by week.
- The minutes each task would take a person are our estimate. The client's staff have not timed them yet, so we show the range, not just the middle.
- Some of it is work nobody did before, like the morning briefs. We count it as the time a person would need to produce the same thing.
- Measurement is continuing: a timed sample week with the client's staff comes next, and this number will be updated with it.
The client is not named because we have not asked to use their name. No prices, no client details, no figures from their books.
What AI does to one specific job, measured.
Controlled studies that timed people doing a real task with and without AI. Each one is peer-reviewed and linked. None of them are our clients and none of them are prices.
50 to 130%
More legal work done in the same time, on five of six legal tasks
Randomized trial: upper-level law students did realistic legal tasks with a legal research AI tool, a reasoning model, or no AI. Work quality also went up. The biggest gains were on drafting a persuasive letter and analyzing a complaint.
40% less time
On professional writing tasks from people's own line of work
Randomized experiment with 453 college-educated professionals doing writing tasks from their own line of work. Time fell by 40 percent and graded quality rose by 18 percent.
25% faster
On realistic professional-services work, with 12% more of it finished
Preregistered experiment with 758 consultants on 18 realistic tasks. Quality went up too. On one task picked to be outside what AI does well, people with AI did worse, which is why every build has a person checking what matters.
15% more
Customer questions resolved per hour
5,172 customer-support agents, measured on the job as an AI assistant was rolled out. The least experienced people gained the most in both speed and quality.
In each study, people did the job with AI help and were measured against people who did not. Our builds go a step further and hand the routine part of the job to an agent, with a person checking what matters.
The floor: general AI tools, averaged across everyone.
These national figures count anyone who opened a chat tool, for any reason, across every kind of job. They are the baseline, not the target.
About 2 hrs
A week, per person, from general AI tools alone
Workers who used generative AI saved 5.4 percent of their hours, about 2.2 hours of a 40-hour week. The May 2026 tracker puts savings at 2.2 percent of all U.S. work hours, users and non-users together.
1 in 5
Weekly users who saved 4 hours or more
20.5 percent of people who used generative AI in a given week said it saved them four hours or more that week. Daily users reported more: 33.5 percent saved four hours or more.
$46.89
What an hour of a private-sector employee costs
Average total employer cost per hour worked, wages plus benefits, private industry, June 2026. The number to multiply your hours by.
At 2 to 4 hours a week per person and $46.89 an hour, that is $406 to $813 a month per person in time you are already paying for.
That is the floor: what people get from general chat tools on their own, averaged across every kind of worker. The task-level studies and our own measured number above show what happens when AI is pointed at one specific job.
Whether it pays for itself is one division.
You do not need a model or a consultant's spreadsheet to answer it. You need three numbers about your own week, and you should see them before you sign anything.
The formula
Months to payback = what you spend in year one, divided by what it saves you each month.
Hours the job takes each month
Count it before anything is built: the intake, the follow-ups, the retyping. That count is your baseline.
Times what an hour costs you
Use your own payroll number. If you do not have one handy, the national private-sector average is $46.89 an hour.
Times the share that gets handed off
Rarely all of it. A person still reviews what matters, and that time stays on the books.
That gives you the monthly saving. The year-one cost comes from your written scope, after the intro call, so it is not on this page: every build is priced against your own week.
- Hours a month on one job60
- At $46.89 an hour$2,813 a month
- Two-thirds handed off$1,876 a month
Divide your year-one cost by $1,876 and you have the number of months it takes to pay for itself. We run this with your real hours on the intro call, before you commit to anything.
Big companies use AI at about twice the rate of small ones.
Size predicts who is using AI more than almost anything else. Large firms have people whose job is to figure it out. Most small businesses do not.
- 250 or more employees37%
- 100 to 249 employees32%
- All U.S. firms19.8%
- 4 or fewer employeesUnder 20%
Share of U.S. businesses that used AI in any business function, survey window ending May 3, 2026. Over the six months, use rose among firms with 20 or more employees and did not change meaningfully among firms with fewer than 20.
66%
say AI meets or beats expectations where an active AI plan exists
22%
say the same where there is no active plan
From a survey of 1,816 legal, tax, audit and compliance professionals. The gap is less about the technology and more about whether anyone has the job of putting it to work. In a small business, that is usually nobody. That is the job we take on.
You keep the machine. We keep score.
The agents run on a Mac mini in your office, and it is yours.
You buy the machine, it sits on your network, and your files stay on it. The machine is not rented back to you, and if we ever part ways it stays where it is.
- The hardwareYours
- Where your files liveYour office
- If we part waysIt stays
Why it matters: in the same Thomson Reuters survey, 96 percent of professionals said they want AI that safeguards confidential data. How we handle that.
AI projects fail for boring, predictable reasons.
The usual two: the scope kept growing, or the information going in was a mess. Neither is a mystery if somebody goes looking early. Our whole process is built to go looking early.
Scope creep
We write down exactly what we are building before we start, and you sign off on it. Good new ideas go on the list for next time instead of quietly stretching this build until it never finishes.
Data quality problems
We look at the actual information going in before we build anything. If it is a mess, we say so on day one and either clean it up or design around it, rather than discovering it the week you go live.
No human checkpoints
We decide up front where a person has to look before something goes out, and the deciding factor is what it would cost to get it wrong. Not how busy anyone feels that week.
Accuracy compounds badly
Two steps that are each right 85 percent of the time are only right 72 percent of the time together. We hold every step to 95 percent or better, because small gaps multiply into big ones.
Being right most of the time is not good enough.
When one step feeds the next, the mistakes multiply. Three steps that are each right 85 percent of the time are only right 61 percent of the time end to end. Push each step to 95 percent and the same job lands at 86 percent. A nine-point difference per step turns into a twenty-five point difference in what you get.
We hold every step to 95 percent or better, which is why we test against a copy of your information first, deliberately feed it bad inputs to see it stop, and make it answer only out of files you have checked, before it ever touches a real client record.
- 3-step workflow at 85% per stepPoor result: 61% overall
- 3-step workflow at 95% per stepGood result: 86% overall
- 5-step workflow at 85% per stepPoor result: 44% overall
- 5-step workflow at 95% per stepGood result: 77% overall
Illustrative - multiplicative accuracy math
These four steps are where those failures get caught.
The same steps run on every build, no exceptions and no shortcuts when a schedule gets tight. These four are the ones that catch the problems above while they are still cheap to fix.
We write down how the work really happens
Each job gets described in plain language before anything is built, and you read it back to us. That is where the surprises come out: the exception nobody mentioned, the spreadsheet that runs half the company.
We decide where a person stays in the loop
Some steps can run untouched. Some should never go out without a human look. We choose that line up front, based on what it would cost to get it wrong, and we write it into the build.
We build it somewhere safe first
The first version runs against a copy of your information, walled off from anything live. Mistakes happen there, where they are free, instead of on a real client file.
We try hard to break it
We throw the weird stuff at it on purpose: the bad input, the missing page, the question it was never meant to answer. We want to see it stop and ask for help rather than make something up.
We are early, and there is an upside to that for you.
We have one measured number so far, at the top of this page, and no named case studies yet. We are not going to invent any. What being early means for you: you work directly with the three partners, you get the most careful attention we will ever be able to give a build, and your numbers join ours on this page once your system is running.
If you want to be an early client and have your results on this page before others, that conversation starts with the intro call.
- Published client case studiesNot yet
- Numbers on this pageMeasured or sourced
- Who you work withThe partners
- Your resultsMeasured, then published
Honest by default
Run the math with your own numbers, not ours.
Half an hour, no cost. We put the published figures next to your actual week and tell you honestly what the payback looks like before you spend a dollar.