AI Policy

Govern

Who is responsible, and how a decision about AI is made.

• One owner. TRIOPilot’s founder, our Security Owner, owns this policy. He approves every AI feature, every AI provider and every change in the data an AI feature sends, before any program can use it.
• One AI provider. Both features use Anthropic’s Claude through Anthropic’s commercial API. Anthropic is the only AI provider that receives program data, and we use no Google AI features. We train no AI model of our own.
• Training. Both TRIOPilot staff have completed responsible AI training. Our client engagement lead does not build or configure the AI features.
• Changes are published. When an AI feature changes, the same release updates this page, the Help Center, our Privacy Policy and our HECVAT. Before a change in how a program’s data is handled is published, we email each institution.
• Review. We review this policy every year, and after any new AI feature, any AI incident, or any change in the provider’s terms on training or retention.

Map

What each feature does, what it sends to the AI provider, and who it affects.

• AI Assistant

  What it does: Answers staff questions about their program’s records, builds student lists, answers questions about using TRIOPilot from the Help Center, and drafts contact logs and texts for staff to review.

  Who uses it: Advisors, assistants, observers and admins. Guests, students and applicants cannot.

  What the AI receives: The staff member’s question, the conversation so far, the staff member’s name and role, and the records the assistant looks up, limited to what that staff member can already open.

  Who turns it on: The program’s admins. It is off until they turn it on.

• Paper Applications

  What it does: Reads a scanned or photographed paper application into the Add a Student form, for staff to check.

  Who uses it: Advisors and admins, who can add students.

  What the AI receives: The uploaded pages as the applicant wrote them, with the program’s application questions. No other records.

  Who turns it on: Only TRIOPilot, when a program asks. It is off until then.

• Who is affected. Students and applicants, including minors in programs such as Talent Search and Upward Bound; applicants’ households, through answers such as household income; and program staff named in records.
• What the AI never does. Neither feature decides anything about a person. The AI does not accept, reject, rank or score applicants or students, set a status or change a record, and it never computes the numbers in a federal report. A person makes every change.
• Support work. AI tools from Anthropic used in support work, under Anthropic’s consumer terms, with model training turned off, can receive support reports and error details that can identify a student. They never change program data without our Security Owner’s approval and never send messages to programs.

Risks We Identified

• Program data leaves TRIOPilot for the AI provider.
• An assistant answer is wrong or incomplete.
• Handwriting is misread, so a wrong value could reach a student record.
• Text in a record or on a scanned page tries to give the AI instructions.
• An applicant writes a Social Security number on the page.
• Use or cost runs higher than planned.
• The AI provider is unavailable.

The steps under Measure and Manage answer each one.

Measure

How we test the features and watch them in use.

• Tests before every release. Automated tests run before every release and check the limits under Manage: for example, that the assistant has no way to save or change records, uses no privileged database access and stores no conversation text, and that a student is created from a scan only after a person checks it.
• Reading accuracy. We measure paper application reading on a test set of synthetic handwritten applications, filled in with invented answers: neat and cursive hands, light pencil, poor scans and phone photos, mostly in English and Spanish. The most recent full run read 308 of 318 answers exactly.
• New models. A model is offered for the AI Assistant only after a test of every one of its tools passes.
• Usage records. Every AI request is recorded with the program, the person, the time, the model, its size and its cost. For the assistant, the record lists which lookups it ran, never the text of the conversation.
• Error monitoring. Failed AI requests and lookups are recorded, and our Security Owner is alerted when the AI provider refuses our key or our credit runs out.
• Monthly allowance. We are alerted when a program reaches 80 percent of its monthly AI allowance and when it reaches the limit. Program admins see their program’s use in Settings.
• Reports. We check every report of a problem against these records.

Manage

The controls that limit what the AI can do, and what happens when something goes wrong.

• Off by default. Both features are off for every program until turned on. A program’s admins can turn the AI Assistant off at any time. TRIOPilot turns Paper Applications off for a program at any time, and scans still waiting are then not read. We can turn both features off for every program within minutes.
• The assistant reads and drafts; a person saves or sends. It reaches records only through a fixed set of 13 tools that apply the staff member’s own permissions. A contact log is saved only when the staff member chooses Log Contact, and a text is sent only when they choose Send, with the same checks as Bulk Texts, including the student’s consent. It cannot change or delete a record, send email or notify anyone. Its answers link the records they come from, so staff can check them, and it is told to label any count it gives as informal, never an official APR number.
• A person checks every scanned application. The AI that reads scans has no tools and cannot look up or change any record. TRIOPilot checks every answer it reads against the program’s questions and removes any number shaped like a Social Security number. Staff check each answer, with unclear ones marked Check This, and choose Add Student; no student is created from a scan without that step. A discarded scan is deleted at once, and every scan nobody adds is deleted 30 days after upload.
• Limits. Each program has a monthly limit on AI requests (500 unless TRIOPilot sets another) and on spending, set only by TRIOPilot. One person can ask the assistant at most 20 questions a minute.
• Data the provider keeps. Under Anthropic’s commercial terms, which include its data processing addendum, Anthropic may not train models on program data and deletes API inputs and outputs within 30 days; abuse monitoring or a legal hold can keep them longer. TRIOPilot’s database stores no AI Assistant conversation. Our error monitoring receives a conversation’s text only with a failed request, or with about 1 in 20 requests sampled to measure performance, and deletes it after 30 days.
• Your choices. Each institution decides for its own programs. To turn Paper Applications on or off, email support@triopilot.com. Staff can choose not to use the assistant, and can type an application in by hand instead of uploading it.
• Incidents. An AI problem that puts program data at risk is handled under our incident response plan: we turn the feature off for the program, or for every program, and tell each affected institution within 72 hours of confirming that its data was accessed, changed or lost without authority.

Known Limitations

• We have not run adversarial tests of attempts to instruct the AI through a record or a scanned page. Whatever the AI is told, the assistant has no way to save or send anything, and the scan reader has no tools.
• Assistant answers are not scored against a fixed test set; staff check them against the linked records.
• Reading accuracy is measured on synthetic applications only, mostly in English and Spanish.
• A student or applicant cannot opt out alone; the institution decides for its program.
• A scanned page keeps everything the applicant wrote, including any Social Security number, and is filed in the student’s documents once the student is added.
• AI requests do not use Anthropic’s United States processing option, so they may be processed outside the United States.
• The AI tools used in support work run under Anthropic’s consumer terms, which promise no retention period.

Report a Concern

To report a problem with an AI feature, such as a wrong answer, a misread application, or program data shown or sent where it should not be, email security@triopilot.com. Staff can also choose Share Feedback in TRIOPilot. Applicants and students can email support@triopilot.com or tell their program.

Our Security Owner reads every report. If program data could be at risk, we handle it as a security incident. Otherwise we check the records involved, and we can turn the feature off for the program while we fix the cause. We reply when the report is resolved, and we tell the institution when its records were affected.

See also our Security page and Privacy Policy.