Machine learning built on your data, for your decisions.
Some decisions repeat thousands of times: which account to work first, which claim will be denied, which invoice looks wrong. If you have the history, a model can rank them better than a rule or a gut call. We build that model on your data, around how your team works, and put it inside the tools they already use.
- Trained on de-identified data
- Back-tested on your own history
- You own the model
25%
improvement in collections from targeted effort
We helped a third-party collections firm build a propensity-to-pay model on their own, de-identified data. Not a vendor score bought off the shelf: a model trained on their accounts, their outcomes and their way of working, so collectors start each day on the accounts most likely to pay. Targeting effort that way improved collections by 25%.
Why custom beats a bought score
- Your data and your outcomes. A bought score was trained on someone else's accounts; yours is trained on what actually happened in your operation.
- Your definitions. What counts as “paid”? As “denied”? As “worked”? The model uses your answers, not a vendor's.
- Your constraints. Which accounts must never be touched, which queues have a deadline, which inputs must stay out.
- Explanations your team can read. Why this account scored high, in the words they use, not a number with no story.
- You own it. The model, the code that trains it and the documentation are yours.
Other places it earns its place
- Collections and receivables
- Propensity to pay, best time and channel to contact, settlement likelihood, payment-date prediction for cash forecasting.
- Claims and billing
- Denial prediction before submission, underpayment detection, which claims to follow up first, expected reimbursement.
- Work queues of any kind
- Rank by likely value or urgency so staff start at the top.
- Documents and messages
- Classify and route incoming mail, email and faxes; flag the ones that need a person.
- Finance and audit
- Duplicate or unusual invoices and payments, expense and vendor anomalies.
- Customers
- Churn risk, lead and quote scoring, likelihood to renew.
- Planning
- Call, claim or order volume forecasts for staffing; inventory demand.
- Data quality
- Matching the same person or company across systems that spell it differently.
The same shape every time: a decision that repeats, history with known outcomes, and a queue where the score changes what someone does next. See collections and healthcare billing for the work around it.
When you don't need it
- A rule does the job. If you can write the decision in a sentence, a script beats a model.
- There isn't enough history. A model learns from outcomes; without a few thousand of them, it is guessing.
- Nobody will act on the score. A ranking nobody works is a report nobody reads.
- The decision is too rare. A model for something that happens twice a month costs more than it returns.
We'll tell you. A sorted report is sometimes the right answer. Automation vs. AI: which one your task needs.
How it works
- 1
Feasibility check on a sample of your data
Is there signal, and is it worth it? A few weeks with a de-identified sample, ending in a written answer either way.
- 2
Build and back-test against your own history
So you see how it would have done on last year's accounts before you rely on it on this year's.
- 3
Put the score where people work
The queue, the report, the system of record. Not a dashboard nobody opens.
- 4
Monitor and retrain as your data changes
Outcomes drift. The model is checked against them and retrained on a schedule, under the managed plan.
Your data
Models are trained on de-identified data. Your data stays yours, is used only for your model, and is never pooled with another client's.
A score that ranks work is different from a decision about a person. We design for the first, document what goes into the model, and leave out inputs that shouldn't be there.
Pricing and guarantee
Model Feasibility Check
$2,500–$5,000
Credited toward the build.
A written answer on a sample of your data: is there signal, and is it worth building? If the answer is no, you have the answer and the credit is the cost.
Build
$25,000–$75,000
Typical range. Fixed quote before work begins.
First deliverable within 30 days: a baseline model back-tested on your own history, so you can see how it would have done. Monitoring and retraining sit under the managed plan.
All prices on the pricing page. For the rest of the work around a model, see AI and automation consulting.
Machine learning questions.
How much data do we need?
Enough history with known outcomes to learn from: usually a few thousand cases where you know what happened. The feasibility check answers this on your actual data before you spend on a build.
Do we need a data scientist on staff?
No. We build it, document it and put it inside the tools your team already uses. Your team needs to act on the score, not maintain the model.
Is this the same as ChatGPT?
No: it's a model trained on your records to predict one thing. It does not chat, and it does not know anything outside your data. That narrowness is what makes it reliable.
Who owns the model?
You do. The model, the training code and the documentation are yours, and your data is never pooled with another client's.
What happens when our data changes?
Models drift as the world does. Under the managed plan we monitor the score against real outcomes and retrain when it slips. Without the plan, retraining is quoted as it comes up.
Can we see why an account scored the way it did?
Yes. Every score comes with the inputs that drove it, in plain language, so a collector or a biller can see the reason and disagree with it.
Tell us the decision that repeats.
Which account, which claim, which invoice. A few sentences and a sense of how much history you have. We come back with whether a model is worth a feasibility check, or whether a rule will do.
Free evaluation. Fixed quote before work begins.