If AI makes your analysis faster or cheaper, you are probably using it wrong (or someone is skipping a step you are still paying for). What it reliably buys, when the work is checked, is quality, and quality costs checking time.
Where the time goes
Any piece of analysis someone acts on goes through three steps. First you check the inputs: is the customer table complete, does “active” mean what the finance team means by it, did last month’s data load twice? Then you do the work: the queries, the model, the write-up. Then you check the outputs: do the numbers tie back to the source, does the method fit the question, does the conclusion follow from the numbers?
AI is very good at the middle step. It drafts the queries, the code and the prose in a fraction of the time. It cannot answer for the first or the last, because both rest on knowing the business, and someone has to put their name to them. Take a report: 2 days of drafting and 3 of checking, 5 in all. The checking is the larger part already, before AI touches anything.
AI speeds up the middle step; the checks on either side take as long as before
One report, four ways
Let’s put a price on it. Say a day of a senior analyst’s time costs €1,000 (put in your own figure; every ratio below holds at any price).
- Before AI, 2 days drafting and 3 checking cost €5,000.
- With AI drafting and the same checks, drafting falls to 1 day: 4 days and €4,000. That is the best case for saving money with AI, and it is a fifth off.
- Delivered in half the time, the report takes 2.5 days and costs €2,500. With drafting already down to 1 day, the other 1.5 days are all that is left for checking: half of what the report needed.
- Used to go twice as deep, AI still drafts in 1 day, but twice the analysis means twice the checking, 6 days. That is 7 days and €7,000, two fifths more than before AI.
Going twice as deep takes 7 days; half the time leaves half the checking
The same data as a table
| Case | Drafting, days | Checking, days | Total, days | Cost |
|---|---|---|---|---|
| Before AI | 2 | 3 | 5 | €5,000 |
| AI drafts, same checks | 1 | 3 | 4 | €4,000 |
| Half the time | 1 | 1.5 | 2.5 | €2,500 |
| AI used to go twice as deep | 1 | 6 | 7 | €7,000 |
So if AI halved your delivery time, someone stopped checking. There is no other step it could have come out of.
What the extra days are set against
A price only means something next to what the report decides. Take the invented subscription company that runs through this site’s demos, and its entry plan: 2,400 customers at €29 a month, 5% of them leaving each month. The report answers whether to fund a retention programme that promises to cut churn by 1%. On the arithmetic in what reducing churn by 1% is worth, a one-point cut brings in about €156,800 of revenue over two years, €125,400 of margin at the company’s 80% gross margin. That margin is the most the programme could cost over the same two years and still pay for itself.
“1%” can mean one percentage point, 5% to 4%, or 1% of the churn rate, 5% to 4.95%. Suppose the vendor meant the second and the report assumed the first. Every query runs, every chart renders, and the report overstates the programme by about €149,400 over two years: the relative cut is worth about €7,400. Nothing in the output looks wrong. Only someone who asks what the vendor actually promised finds it.
The checking a faster report skips costs little next to the mistake it would catch
The same data as a table
| Item | Euros |
|---|---|
| The checking cut to deliver in half the time | €1,500 |
| Three more days of checking, to go twice as deep | €3,000 |
| What the report overstates if it reads a relative 1% as a point | €149,400 |
Half the time saves €1,500 of checking. Going twice as deep costs €3,000 more of it. Either amount is small next to a decision worth up to €125,400 of margin, and the mistake above is found in the checking, not in the drafting.
What the research shows
The best evidence on what AI does to knowledge work comes from a preregistered experiment with Boston Consulting Group, published in Organization Science in 2026. 758 knowledge workers were randomly given GPT-4 or no AI. On 18 realistic tasks the researchers had picked to sit within what AI could do, from creative to analytical, the people with AI completed 12.2% more tasks, 25.1% more quickly, and their work was of significantly higher quality. On one complex task picked to sit outside it, they were 19% less likely to reach a correct answer than people without AI.
Even on tasks chosen to suit the AI, people finished 25.1% more quickly, well short of the half that halving a delivery time needs, and the gain in quality came with it. Where the task did not suit it, AI made people worse, and the authors describe the two kinds of task as having “a seemingly similar level of difficulty”. Telling them apart is part of the checking.
With AI, 25.1% quicker on tasks that suit it, and less often right on one that does not
18 tasks within what AI could do, against people without AI
One task outside it
The same data as a table
| Tasks | Measure | Difference |
|---|---|---|
| 18 within what AI could do | More tasks completed | 12.2% |
| 18 within what AI could do | Faster on average | 25.1% |
| One outside it | Less likely to reach a correct answer | 19% |
What skipping the last step costs has a public price, too. CFO Dive reported on 21 October 2025 that Deloitte had refunded over A$97,000 of a contract worth about A$440,000 to Australia’s Department of Employment and Workplace Relations for a report with AI-generated errors: a fabricated quote from a federal court judgment and references to research papers that do not exist. The errors came out after the report was published, and a revised version followed. A department spokesperson told CFO Dive that the substance of the review had been retained. A check of each quote against its source finds both kinds of error.
When AI does make analysis cheaper
There are real exceptions, and they share one feature: the checking got cheaper too, or it was never needed.
- A routine refresh: the same monthly report on new data, with its checks written as scripts: totals that must tie to the source, ranges a figure must stay inside, counts that must match last month’s within a tolerance. The checks are built once and then run on every refresh in seconds. Here the price per run can honestly fall, because the expensive checking was paid for once.
- A decision that is cheap to reverse: an internal first look, a hunch to test, a draft nobody will act on. Taking unchecked AI output is a fair choice when a wrong answer costs less than the checking would. It is a choice, though, and it should be made knowingly rather than discovered later.
- Checking that AI can do part of: reviewer agents can run part of the output check: reading the work against the brief, re-running numbers, hunting for gaps. That shortens the checking, though not the judgement at the end, and what AI review catches and misses is a question of its own.
None of these covers the open questions a business actually hires analysts for: why churn rose, what a price rise would do, which customers to target. There the method is part of the answer, nothing from last month says what this month should look like, and the checking cannot be scripted in advance. The evidence on where AI falls short sits almost entirely on that side.
How FinDaS prices it
This is why I price per deliverable, fixed, and written down before work starts. Billed by the day, the day AI saves in drafting shows up as a smaller invoice, and the next day to cut is a checking day, because there is nothing else left. At a fixed price the saved day goes into depth: one more cut of the data, one more scenario, the range around a figure instead of the figure alone. The deeper work is checked like the rest, and the method note on each deliverable lists which checks ran.
Billed by the day, the saved day comes out of the checking; at a fixed price it goes into depth
AI agents do the drafting, scripts and reviewer agents do the checks they can, and what you act on gets my signature at the end. The steps are on how it works.
Sources
- Dell’Acqua, McFowland, Mollick, Lifshitz, Kellogg, Rajendran, Krayer, Candelon and Lakhani, Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality, Organization Science 37(2), March–April 2026. First a Harvard Business School working paper, 2023.
- Alexei Alexis, Deloitte refunds over $60K for report with AI errors, Australian government says, CFO Dive, 21 October 2025.