01
ROUGHLYHALF
of AI pull requests that passed the SWE-bench automated grader still would not have been merged by actual maintainers.
The automated grader was about more optimistic than the maintainers.
Source · METR · March 2026Never Twice
Better AI creates more to review.
Everyone is racing to improve the output. We`re trying to shrink the review.
What is the AI Bottleneck?After the AI gets betterSame review team
Machine output
multiplies
Human review capacity
doesn't
You multiplied output. Your reviewers stayed human.
The Bottleneck
AI scales. Human attention doesn`t.
When your experts keep reviewing the same kinds of exceptions, the AI review queue grows faster than the team.
Never Twice learns from those interventions so repeated judgment stops coming back as repeated work.
Capacity over time
the widening gap
02
of AML transaction-monitoring alerts are estimated to be false positives.
Large volumes of machine-generated suspicion still demand human attention.
Source · Bank for International Settlements · Project Aurora03
of security alerts were false positives in research cited by Microsoft.
Another 42% went uninvestigated.
Source · Microsoft Security · February 2026What the record misses
The correction gets saved.
The reason usually doesn’t.
A reviewer catches something subtle, fixes it, and moves on.
A few days later, the same mistake comes back wearing different clothes.
Same reason · different item
Monday
same reason
Thursday
same reason
Next month
same reason
Never Twice
A resolved review shouldn`t be repeated.
Skagway captures why an expert intervened, validates when that judgment applies, and turns it into reusable behavior so substantially similar cases stop returning as routine human work.
Today
- AI output
- Human reviews
- Human corrects
- Item closes
- Similar item returns
- Human reviews again
Never Twice
- AI output
- Human intervenes
- Reason captured
- Judgment validated
- Similar future cases handled automatically
- Only new or required judgment reaches the human
Give us one AI review queue.
We make the repeated part shrink.
The economics
What's the cost when your AI output doubles?
What the queue costs now
You currently spend approximately:
$432K / year
416 expert hours / month
reviewing AI-generated work.
If machine output doubles and nothing else changes:
832 expert hours / month
If repeated review falls
Illustrative scenarios, not Skagway performance claims
25%
$108Kannual cost released
104 hoursreleased each month
32.1machine outputs per expert review hour
50%
$216Kannual cost released
208 hoursreleased each month
48.1machine outputs per expert review hour
75%
$324Kannual cost released
312 hoursreleased each month
96.2machine outputs per expert review hour
Machine outputs per expert review hour
24.0
Skagway metric: Autonomous Throughput per Expert Hour
Uses a 40-hour workweek and 52 weeks per year to allocate fully loaded annual cost. Figures describe the inputs you choose. They do not predict a Skagway result.
Before the trial
Most review queues are busy. A few are repeating themselves.
The work does not have to look identical. It only has to keep asking the same senior person to notice the same thing.
Is there a pattern here worth proving?
01
It comes back
A correction reappears in new wording, a new ticket, or another pull request.
02
Someone always knows
One person sees the thing everyone else is inclined to pass.
03
There is a record
The trail exists, even if no one has followed it closely yet.
The engagement
Never Twice Trial
One queue.
6–8 weeks.
Your current workflow stays in place.
The trial is built to prove, in shadow mode, how much repeated human review can safely disappear before anything changes in production.
Week 1
Read the queue
Establish what is arriving, who is touching it, and where time is going.
Week 2
Look for the pattern
Some corrections are one-offs. Others keep returning in new forms.
Weeks 3–4
Test the judgment
Follow the strongest thread with the people who see it first.
Weeks 5–6
Replay the evidence
See whether the pattern holds when it meets work it has not seen.
Weeks 7–8
Watch it live
Run beside the existing workflow and measure what could safely leave routine review.
Six to eight weeks is the expected range. The trial leaves a measured answer about what should remain with people—and what may not need to.
Trial pricing
$75,000 total
$40,000 to begin
$35,000
only if the trial demonstrates a measurable reduction in repeated human review under the agreed quality and safety threshold.
The Shadow Proof Guarantee
If the Never Twice Trial cannot demonstrate a measurable reduction in repeated human review during historical replay or shadow operation—without exceeding the quality and safety thresholds agreed before the trial—you do not owe the final $35,000.
The Never Twice Guarantee
Never pay us to rediscover the same lesson.
Once a review class has been formally validated and retired, if substantially the same root-cause pattern returns to routine human review during the engagement, correcting that recurrence is on us.
Start with one review workflow
Inquire
Tell us where your review workflow lives.
We'll look at it with you and determine whether there is a fit.
Frequently Asked
Before we begin
What counts as a review queue?
Any recurring workflow where AI-generated work or actions wait for a person to approve, edit, reject, or escalate them.
Do you replace our current tools?
No. We`ll work with whatever system you currently use.
What review can actually be removed?
Only repeatable judgment classes that can be validated safely. Novel, ambiguous, or legally required human decisions remain human.
What happens after the trial?
If the economics and safety case are strong, a Never Twice Deployment can extend the work to more review classes, workflows, or teams. If they are not, stop.