Customer operations
- Ticket routing
- Complaint intake
- Escalation triage
Drafted for enterprise
Teach AI how your best people work. Then own it.
Pick one job your team does by hand every day, like sorting tickets, claims, or documents. We train a small AI model on how your experts do it, prove it on cases it has never seen, and hand you the model to run in your own cloud.
Reviewers verified through
41.6%Right before training
72.0%Right after training
Our public test: 1,190 real consumer complaints, sorted by the same small model before and after training.
We already did it once
Same model. Half the mistakes.
- 52%fewer mistakes: 695 wrong before training, 333 after.
- +362more cases filed right, out of 1,190 it had never seen.
- 2xbetter on rare categories, the ones that cost the most when missed (0.360 to 0.718).
- 1,000random reshuffles of the test, and the gain held up. It was not luck.
41.6%Right before training
72.0%Right after training
The same test, on your work
We did it on public data. Here is your version.
Nothing about the method changes. Only the job and the definition of correct become yours.
Our public testYour pilot
The job
Sort consumer complaints into the right product and issue.
One queue your team sorts by hand today: tickets, claims, invoices, documents.
Correct means
The public archive’s own records.
Your experts’ rules, built into an answer key by our verified reviewers and signed off by your team.
The model
A small open model (Qwen3-8B) plus a small trained adapter.
The same approach, running in your cloud, owned by you.
The test
1,190 cases the model never saw.
Your own unseen cases, scored against a KPI you pick before we start.
The result
41.6% to 72.0% right. 52% fewer mistakes.
Your before and after, in writing. If it does not win, we say so.
Where it fits
If a person reads it and puts it in a bucket, we can train for it.
The method is strongest on high volume work where a wrong call costs time or money.
Finance
- Invoice exceptions
- Expense coding
- Transaction categories
Insurance
- Claims coding
- First notice of loss triage
- Document sorting
Legal and compliance
- Contract clause tagging
- Policy checks
- Review queues
What it could mean for you
Put in your volume. See what half the mistakes is worth.
5,840mistakes a month at 58.4% wrong
2,798mistakes a month at 28.0% wrong
3,042 fewer mistakesand about 254 hours of rework back every month, if your queue behaved like this test.
Uses the error rates from our public test. It shows scale, not a promise: your real number comes from a pilot on your own cases.
How a pilot works
Take no risks. The number is agreed before we train.
Four steps, one workflow, and a scorecard you can hold us to.
- 001 Pick
Pick one workflow
A queue with real volume and a real cost when it goes wrong. One job, not a platform rollout.
- 002 Define
Define correct
Your experts set the rules. Our verified reviewers turn them into an answer key your team signs off on.
- 003 Train
Train a small adapter
It snaps onto an open model. The base model stays untouched, and the adapter comes off anytime.
- 004 Prove
Prove it on your cases
Your process today, the untrained model, and the trained one, on the same unseen cases, scored on your KPI.
You walk away with
A scorecard on your own cases, the trained model running in your cloud, and the answer key your team approved. Yours to keep either way.
Why Drafted
People you can verify. A model you can keep.
Verified people, not anonymous crowds.
Every reviewer is enrolled at a partner university and identity checked. The same cohorts come back every semester, so quality does not reset.
You own the model.
Open weights, running in your cloud. No per call fees to someone else’s model, and no lock in.
Proof before promises.
We agree on the KPI before training starts, so the result cannot be moved afterward. If it does not win, you hear it from us first.
For your security team
The questions they will ask, answered.
Where does our data go?
It stays in your environment and sensitive fields are redacted on ingest. Our reviewers work only inside the access you grant, on the cases you approve.
Who owns the model?
You do. The open base model and the trained adapter run in your cloud. Keep it, retrain it, or remove it without us.
Which models do you use?
Open weight models you can run yourself. Our public test used Qwen3-8B. We start with the smallest model that clears your bar, because smaller models are cheaper to run.
Do we need a big rollout to start?
No. Start with one workflow. If the pilot does not beat your current process on your own answer key, you have spent little and learned something real.
Is the case study a customer result?
Not yet, and we say so plainly. It is our public rehearsal on open data. Your pilot is where it becomes a business number.
Start here
Bring one workflow and the number you care about.
Tell us the queue, the monthly volume, and what correct means to your team today. We will tell you if it is worth testing, including when it is not.
Goes straight to our CTO at rodrigo@joindrafted.com