One workflow on AI, counted before and after.
We take one routine task your staff does by hand today, such as reading invoices, searching your own files or drafting from your templates, and set up AI to do it, with a person still checking the result. It runs in your own cloud account, and we count the task before and after so you can see what changed. Most clients start with a fixed-price sprint and decide on the rest from the count.
Start small. Assess when there is more than one candidate. Build. Run.
Four engagements that feed each other. Most clients start with the first one.
1. AI quick-win sprint
$3,500 fixed · 2 to 4 weeks
One workflow, on your data, running in your own AWS account: invoice or intake extraction, search over your own files, drafting from your templates, a records-request queue. We count the task before and after, a person stays in the loop, and you keep what we built whether or not there is a next step.
Start a sprint2. AI Readiness Assessment
From $7,500, fixed price · 3 to 4 weeks
For when there is more than one candidate, or a board to convince: an inventory of your data and systems, three to five opportunities scored by payback, the governance gaps that would stop each one, and a roadmap you can hand to any vendor.
About the assessment3. Implementation
Fixed price after a sprint or discovery
Extraction from invoices, orders and forms. Search and question-answering over your files. Connecting a model to your ERP, CRM or practice-management system. Deployed in your account with identity, logging and data controls you can show an auditor. Time and materials only where the scope genuinely cannot be bounded, with an hours report.
Discuss a build4. Managed operations
From $750 per month
Monitoring, model and prompt updates as providers retire models, a monthly cost review, and support hours with next-business-day response: $750 a month covers a small deployment with 3 hours; $2,500 covers several workloads with 10. Or take the handover and run it yourself.
Ask about a retainerWe take one discovery, sprint or assessment and one build or retainer at a time. The people you talk to are the people doing the work.
Payback, not promises.
This is the shape of every estimate we produce. Baseline first, fully loaded labor cost, a human still in the loop, and the low case shown next to the expected one and the upside.
Illustrative example. The numbers show the method, not a guarantee. Your estimate uses your counts.
How we score an opportunity| Low | Expected | Upside | |
|---|---|---|---|
| Keying time saved, of 25 hours a week, with a person still reviewing | 30% | 40% | 50% |
| Hours reclaimed a year, at 50 working weeks | 375 | 500 | 625 |
| Labor value at $40 an hour, fully loaded | $15,000 | $20,000 | $25,000 |
| Duplicate payments prevented, of about $18,000 a year | $9,000 | $13,500 | $18,000 |
| Benefit a year | $24,000 | $33,500 | $43,000 |
| One-time cost: the build, $14,800, and 40 hours of your team's time for training | $16,400 | $16,400 | $16,400 |
| Running cost a year: cloud and model usage, in your account | $3,000 | $3,000 | $3,000 |
| Net in year one | $4,600 | $14,100 | $23,600 |
| Payback on the one-time cost | 9 months | 6 months | 5 months |
| Net each year after | $21,000 | $30,500 | $40,000 |
Labor value is hours reclaimed times $40 an hour, and it only counts if the hours go somewhere useful; here the team moves to collections, which is where the CFO wanted it all along. Payback is the one-time cost divided by the monthly benefit after running costs. The low case assumes 30 percent of keying time saved and half of the duplicates prevented; expected, 40 percent and three-quarters; upside, 50 percent and all. We show the low case first and build the estimate from your counts. If you would rather we run it, the small-deployment retainer adds $9,000 a year and the low case still clears $12,000.
Most AI projects stall for three reasons. None of them is the model.
Nobody measured the baseline.
If the current process was never timed or counted, the project cannot prove it helped. We count for a week before we estimate anything.
Data went where it should not.
Pasting customer records, PHI and contracts into a chatbot is the failure that makes the news. We decide what leaves your environment before a line of code is written.
The people whose work changed were told last.
Adoption fails more often than technology does. Plan on a fifth to a third of the effort going to training and process redesign; we say so up front, and the people doing the work see the tool first.
We design against all three from the first conversation, because we build what we recommend and then live with how it runs.
Start with one workflow.
We reply within one business day with questions or a time to talk. You get a written scope before any number.