ReallySolved · Deck appendix
← Back to the deck
Deck appendix · sources

Where every number came from

Every figure in the deck is below, grouped by what that group of sources proves together. Where a number is ours rather than somebody's published research, it is marked OUR ESTIMATE and the arithmetic is shown, so you can change our inputs and watch the answer move.

The last group is the one worth reading first: the numbers we decided not to use, and why. Several of them are bigger and would have flattered us more.

Every link on this page goes to the report itself, not to a publisher's front page. Open any of them & check the figure we quoted against what they actually say. Where 2 firms disagree, we say so on the row instead of quoting the one that suits us.

1. How many people at a company actually use AI

Together these 3 set the denominator for everything else on this page: in a digital products or services company, roughly 60 to 75 people in every 100 use AI at all, and 25 to 35 of them use it every day. We use 30 daily users per 100 staff, the midpoint.

Not in the deck, on purpose: the physical-goods figure (10 to 15 users per 100). Factories & logistics use far less AI, and they are not who we sell to. Including them would have made the market look bigger & the claim look worse.

2. Whether people check what AI tells them, & what it costs when they don't

Together these prove the thing our whole company is built on: most people do not check, and on hard questions AI made them less accurate, not more. That second finding is the important one, because it is the opposite of what everybody assumes.

OUR ESTIMATE, flagged: the widely repeated figure that 2 out of 3 people never fact-check AI output before passing it on appears in a lot of write-ups, but we have not found the original study behind it. So the deck does not lean on it. The Harvard/BCG trial carries that argument on its own.

3. How big the market for corrected AI data is, & how much the forecasts disagree

Together these size the pot we actually sell into: $2.1B to $3.0B today, growing about 30% a year. They also disagree with each other by 2 to 3 times about the future, which is why we quote a range & name the firms instead of picking the biggest one.

What we did with the disagreement: we used the low end. Our forecast is built from people times hours, bottom up, & then checked against the market in the year it would happen. We never took a share of a market as a forecast. At 1,000,000 sign-ups our number is 2.3% of the 2030 forecast. If we had used share of market, 1% of the 2034 forecasts would have been $210M to $260M, and we are not counting a cent of it.

4. What checking AI already costs a company, per employee

This group is not a market we sell into. It is money a company is already spending, which is why it is a bill we take off their desk rather than a new line in their budget. The arithmetic is ours & every input is shown so you can change it.

There are 2 separate costs here, & a company should be thinking about both. One is the time people spend checking. The other is the money lost on the answers nobody checked. They are different kinds of loss, so we count them separately & add them up.

1. Time. Staff hours spent checking & redoing AI output$2,070
2. Mistakes. The wrong answers that went out unchecked$3,105
Per employee, per year~$5,000
OUR ESTIMATE. Line 1, the inputs: 30 daily AI users per 100 employees (group 1, above), 2 hours a day each on AI, a 230-day working year, staff loaded at $50 an hour, and 30% of AI time spent checking or redoing what it produced.

That last one is the soft input, so here is the range rather than a single number: at 20% it is about $1,380 per employee per year, at 30% about $2,070, at 40% about $2,760. We use 30%.

Why it is a share & not a fixed number of minutes: somebody who uses AI for 2 hours a day cannot be doing a flat 15 minutes of checking. Checking scales with use. An earlier version of this page used a flat figure & it was wrong.
OUR ESTIMATE. Line 2, & why it is 1.5 times line 1. There is a long-standing rule in quality management, usually written 1-10-100: it costs about $1 to stop a mistake happening, about $10 to fix it once it is out, & about $100 once it has reached a customer. We count a mistake that ships at 1.5 times what checking costs, not 10.

We are not claiming the 2 numbers are equally solid. Line 1 is arithmetic on hours & wages. Line 2 is a multiplier we chose, deliberately at about a sixth of what the standard rule implies, because the true figure depends entirely on what the wrong answer touched. A wrong answer in a draft costs almost nothing. A wrong answer in a contract, a filing or a customer email can cost hundreds of thousands. That range is real & we are not going to pretend we can average it for you.
What it looks like for a whole company, if you want it that way. A 200-person digital company: about $1.0M a year, or $552,000 to $1.4M across the 20% to 40% range on line 1. The deck deliberately does not show this, because a per-head number works for a 12-person office & a 40,000-person bank alike, & anybody can do the multiplication.
OUR ESTIMATE. The same cost, added up across every company. The deck says companies will spend somewhere around $10B to $15B a year of their own staff time checking & fixing what AI gave them. It is the everybody-at-once version of line 1 above: the hours people burn checking, rather than anything anyone is being sold.

The word "will" is doing real work here. The forecasts underneath this run to 2030 to 2035, not to today. It is not a claim about what is being spent this year.

What we are not claiming. No published report carries this figure, so none is cited, & it stays marked as ours. Nothing else in the deck is built on top of it. It sits beside the market forecasts as a separate point, because this money is already on somebody's payroll: it is a bill we take off a desk, not a new line in a budget. Group 5 below has more on why the staff-hours side of this kind of estimate stays labelled an estimate.

How much of it we actually take back

Not all of it, & anybody who tells you otherwise is selling something. Here is the whole formula:

What you get back = how many of your people use it × how often we catch it × the number above.

The first number is yours. The second is ours, & it comes from our own accuracy testing rather than from anybody's research. Worked example: if 60% of your staff use it & we catch 70% of what would have gone wrong, that is 42% of the bill, or about $2,100 per employee per year. Run it on 20 people first & measure it before you scale it.

5. The numbers we did not use, & why

Every figure below is bigger & more dramatic than what we put in the deck. We left all of them out for the same reason: they trace back to marketing pages & vendor blogs rather than to research anybody can check. A partner who opens 2 of them finds that out in a minute, & then doubts the numbers that were sound.

The one we would happily use if it gets a real source: the estimate that companies collectively burn $25B to $35B a year checking AI's work. The tooling & services half of it has named reports behind it. The internal-staff-hours half does not, so for now it stays labelled as an estimate & nothing in the deck is built on it.

The 2 rules we hold ourselves to on this page

If any figure here does not hold up, we would rather hear it than defend it. The whole company exists because confident numbers are not the same as correct ones.