AI and automation that earns its place
Of the nine systems we describe in our case studies, three use an AI model we deploy, and in each of those a person stays in the loop. That ratio is the point: we automate what can be automated simply, and use AI only where nothing simpler works.
Six kinds of work, all running in a real agency
Closed-book assistants
Staff ask questions in Microsoft Teams and get answers quoted from your own handbook, policies and the regulations you load, with the source cited and an honest “the documents do not cover that.”
Document intake and filing
Scans, orders, route sheets and invoices read, matched to the right record and filed. A person confirms anything uncertain.
Compliance monitoring
Credential boards, monthly exclusion screening, and a watcher that tells you when a regulation you rely on has changed.
Payroll and billing checks
Reconcile visits against sheets and invoices against notes, and see only the exceptions, so the meeting is about the five problems, not the five hundred rows that are fine.
Dashboards and audit trails
Recurring tasks as buttons, results by email and Teams, and an append-only record an auditor can be shown.
Communications
Phone menus, voicemail turned into tracked tasks, and reminders that reach the right person at the right time.
AI or not?
We use plain automation unless at least one of these is true:
- The input is unstructured language or images whose shape varies run to run
- The rules are too many or too fuzzy to write down
- A person currently reads it and makes a judgment call
If none apply, a schedule, a script or a lookup table wins: cheaper, easier to explain to a surveyor, and with fewer places for data to go. We say so in the proposal. See how we work.
AI proposes. A person decides. Anything an AI reads from a document is treated as a suggestion until someone confirms it. That is how we keep AI inside a HIPAA program.
An AI governance starter kit
For agencies that want to say yes to AI safely. A short engagement that produces:
- A one-page AI use policy staff will actually read
- An approved-tools list with what data each may see
- The data-class rule (public, internal, PHI, decisions)
- A short training session for all staff (see training)
- A review step for adding new tools later
Eight questions for any AI vendor
- Where does our data go, and in what country?
- Is it used to train your models?
- Will you sign a BAA, and does it cover this feature and plan?
- Can the tool be limited to our documents, with the web off?
- Does it cite its sources?
- What is logged, and who can read the logs?
- Who confirms the output before it matters?
- How do we leave, and what do we get back?
AI and automation in use
The three systems that use AI, and what happens around the AI so that it stays safe.
Staff get cited answers from the agency’s own policies, and learn when a regulation changes
Two closed-book assistants in Microsoft Teams answer only from the agency’s documents, with the source cited, and a weekly watcher flags when a saved regulation has changed.
Cited answers and a quiet watcher that speaks up only when it matters.
Running in productionRead the case study Payroll and operationsRoute sheets to payroll, with exceptions caught before checks are cut
Paper and photographed route sheets are read, matched to EMR visits and filed, and payroll gets an exceptions-only list instead of a pile.
94.4% of handwritten rows read correctly by a small local model.
Running in productionRead the case study ComplianceEvery caregiver credential in one trustworthy picture
A weekly audit and a color-coded board replace three sources that disagreed, and a monthly federal exclusion screen runs alongside.
Audit every Monday, board always current, exclusion screen monthly.
Running in productionRead the case study