Pricing fixed-price work when AI makes you faster
A 35% speed-up on AI-assisted tasks is about a 25% cut in project cost. Cutting your price 35% drops the odds of a 25% margin from 80% to 28%.
Short answer Do not cut the price by the headline speed-up. A 35% gain on the tasks AI touches is a smaller saving for the whole project (24.6% in the example), and if the worst cases shrink less than the typical ones the P80 saving is smaller still (22.7%). Price from the new P80, then decide on purpose who keeps the difference.
- Headline speed-up
- 35%on AI-assisted tasks
- Project average cost
- −24.6%P80: −22.7%
- Price cut by 35%
- 28%chance of a 25% margin, from 80%
- The AI dividend
- $15,186old P80 price minus new
The headline is not the project
A claim like “AI makes me 35% faster” is rarely true of every task. Scoping a project with a client, deciding what to build, reviewing the result and handing it over do not speed up the same way that drafting a screen or a test suite does. A discount that applies the headline to the whole price assumes every task got faster by the same amount.
Here is a seven-task project with a blended $800 a day, before and after. Five tasks are assumed to be AI-assisted; two are not.
| Task | AI-assisted | Before (low / likely / high) | With AI (low / likely / high) |
|---|---|---|---|
| Discovery & scoping | No | 3 / 4 / 8 | 3 / 4 / 8 |
| Design & prototypes | Yes | 5 / 8 / 14 | 2.8 / 5.2 / 12.6 |
| Front-end build | Yes | 8 / 12 / 22 | 4.4 / 7.8 / 19.8 |
| Back end & API | Yes | 8 / 12 / 24 | 4.4 / 7.8 / 21.6 |
| Content & data migration | Yes | 4 / 7 / 15 | 2.2 / 4.6 / 13.5 |
| Testing & fixes | Yes | 5 / 7 / 15 | 2.8 / 4.6 / 13.5 |
| Review & hand-over | No | 2 / 3 / 6 | 2 / 3 / 6 |
Inputs. For the five assisted tasks the likely value is 35% lower, the low value 45% lower and the high value only 10% lower; the two unassisted tasks do not change. These are assumptions, not measurements, and this note does not claim AI delivers that gain. Beta-PERT (λ = 4), 100,000 runs, fixed seed, price = P80 cost ÷ 0.75. Figures are rounded for display, so recomputing from the rounded values can differ by a dollar or two.
The tail may move less than the middle
Why might the worst case improve less than the typical one? Because the things that make a task blow out are not necessarily the things AI shortens: a requirement that was misunderstood, generated work that has to be redone, an integration that behaves differently from the documentation, and the time spent reviewing output you did not write. If failures still cost what they always cost while the typical case gets faster, the whole range gets wider relative to its middle.
Whether that is true for you is an empirical question, and the model follows whatever you enter. So compare two versions of the same project: uniform, where every assisted task is 35% faster in all three numbers, and sticky tail, as in the table above.
| Before | AI, uniform | AI, sticky tail | |
|---|---|---|---|
| Average cost | $46,814 | $32,699 (−30.2%) | $35,283 (−24.6%) |
| P80 cost | $50,210 | $34,974 (−30.3%) | $38,820 (−22.7%) |
| P95 cost | $53,681 | $37,302 (−30.5%) | $42,517 (−20.8%) |
Figure 1. The headline speed-up on the assisted tasks, and what is left of it for the whole project when the worst cases improve less.
Two things shrink the number: the two tasks AI does not touch, and the tail that does not move as much as the middle. Under the uniform assumption the saving is 30.2%.
What a blanket discount does to your risk
Start from the old quote: P80 cost $50,210 ÷ 0.75 = $66,946, built so that the chance of earning a 25% margin is 80%. Now try four ways of pricing the AI-assisted version of the same project.
| Price | Amount | Chance of 25% margin | Chance of a loss | Median margin |
|---|---|---|---|---|
| Keep the old price | $66,946 | 99.9% | 0% | 47.6% |
| Cut by the headline 35% | $43,515 | 27.8% | 3.2% | 19.3% |
| Cut by the average saving (24.6%) | $50,457 | 73.5% | <0.1% | 30.4% |
| Price at the new P80 | $51,760 | 80.0% | <0.1% | 32.2% |
Figure 2. The cost of the AI-assisted project (sticky tail) if you cut the price by the headline 35% to $43,515. Left of the teal line costs still earn the 25% margin; right of the red line the job loses money.
Only 27.8% of runs meet the margin, and 3.2% lose money. Under the uniform assumption the same cut meets the margin 50% of the time, so even the optimistic case leaves you short.
The headline cut moves a bid that was built to an 80% margin chance to 28%, and puts a real loss tail on it. Cutting by the average saving is much better and still leaves 73%. Pricing from the new P80 keeps the original policy intact. The same discipline, applied to the quote percentile, is in P50 versus P80.
Who keeps the difference?
The gap between the old P80 price and the new one, $66,946 − $51,760 = $15,186, is what the job no longer forces you to charge at the same risk. Call it the AI dividend. There are three honest uses for it.
- Keep it. Charge the old price and the median margin goes from 30% to 48%. Fixed price lets you keep savings that hourly billing hands to the client automatically, but only while nobody undercuts you.
- Share it. Price nearer the new P80 and win jobs that were out of reach. Explain it as the same risk policy on a smaller job, not as a discount.
- Reinvest it. Same price, more scope, a shorter schedule, or more assurance such as testing, documentation and review.
Which one fits depends on your market, not on arithmetic. The arithmetic only tells you how much room there is and keeps you from discounting into your own tail.
What the client is buying when output is cheap
If drafting is cheap, a price has to rest less on the output and more on knowing what to build, which risks to accept, and someone who stands behind the result. That wider question, how to stay professionally valuable as AI makes raw output cheap, is the subject of How to Remain Valuable When Intelligence Becomes Cheap . This note is its narrow, numerical corner: how to keep a fixed price honest while the work gets faster.
Measure it before you reprice
The two assumptions that drive everything above, how much faster the typical case is and how much the tail moves, are the ones to measure. Keep a small log, as in calibrating estimates with completed jobs, with one row per task: whether AI was used, the estimate you gave, the hours it actually took, and what went wrong if it overran. After eight or ten jobs, look at the typical ratio of actual to estimate and, separately, the bad ones. The bad ones set the P80, and the P80 sets the price.
Until then, re-estimate task by task rather than applying a percentage to the total, and keep the three numbers for each task honest: the low value should shrink where AI helps, and the high value should shrink only where you have evidence that it does.
Doing it in BidVariance
Enter the assisted ranges on each task and run the project twice with the same seed, once with the old ranges and once with the new. Compare the P50, P80 and P95. The Business and Consultant licenses add scenario comparison for two saved runs side by side; see pricing. The estimate templates are a starting point, and the free minimum bid price calculator does the P80-to-price step for a single block of work.
Frequently asked questions
If AI makes me 35% faster, should I cut my fixed price by 35%?
No. A 35% gain on the tasks AI touches is a smaller saving for the whole project, 24.6% in the worked example, and less at P80 if the worst cases improve less than the typical ones. Cutting the price by 35% dropped the chance of earning a 25% margin from 80% to 27.8% and created a 3.2% chance of a loss.
Is fixed price or hourly billing better when AI speeds up my work?
Hourly billing hands the time savings to the client automatically, because fewer hours are billed. Fixed price lets you keep them, but it also leaves you holding the variance, so the price has to come from a re-estimated P80 rather than from the old average.
How do I estimate tasks that use AI assistance?
Re-estimate each task with a low, likely and high value instead of applying a percentage to the total. Lower the likely and low values where AI helps, keep the high value close to the old one unless you have evidence from past jobs, and log actual hours so the ranges can be calibrated.
What is the AI dividend in a fixed-price quote?
It is the difference between the P80 price before and after the speed-up: $15,186 in the example. You can keep it as margin, share it with the client through a lower price, or reinvest it in scope, schedule or assurance.
Are the 35% and tail figures real benchmarks?
No. They are illustrative assumptions chosen to show the arithmetic. Replace them with your own measured ratios; the model follows whatever ranges you enter.