ChatGPT
Claude
Perplexity
Gemini
GrokA custom GPT is a fast pilot for an admissions office, not a finished inquiry desk
Many US admissions teams can have a custom GPT answering questions within a day. It will not, by itself, capture consent, keep a log you control, write to your CRM or know that a deadline changed this morning. Those gaps decide whether the pilot becomes a production service.
This guide is for enrollment, marketing and IT leaders at US colleges and universities deciding whether to build on ChatGPT, build on the API or buy a dedicated tool. It complements our overview of AI chatbots for student recruitment.
What you are really configuring
A custom GPT is ChatGPT with your instructions, your reference files and optional connections. OpenAI's developer documentation is explicit about the pieces. A custom GPT combines instructions, attached knowledge documents and actions that call third-party services through APIs, so a link to your CRM needs an API that you build and host.
For an enrollment office, that means the conversation part is quick and the system part is not. If your CRM is a major platform, someone still has to expose an endpoint, secure it and monitor it. That person becomes the owner of the assistant.
What it handles and what it does not
| Enrollment job | Custom GPT alone | Gap to close |
|---|---|---|
| Answering FAQs from public pages | Good for a pilot | Keep source pages current |
| Deadlines for Early Action, Early Decision and Regular Decision | Risky | Dated source, rule to decline when unsure |
| Financial aid and net price questions | Risky | Link to your official calculator; no individualized promises |
| Transfer credit and prerequisite questions | Risky | Hand off to an adviser |
| Capturing an inquiry with consent | Not built in | Consent wording and stored record |
| Writing to your CRM | Needs custom API work | Endpoint, authentication, alerts |
| Scheduling a campus visit | Needs custom API work | Calendar integration |
| Escalating to a counselor | Weak | Routing rules and context transfer |
The pattern is the same across rows. The model is the cheap part, and the surrounding controls are the work.
Accuracy and the Common App calendar
Applicants in the US move between the Common App, institutional portals and your own site, and they compare answers. If your assistant states a deadline or a requirement that differs from the application platform, the applicant trusts one of them and it is not always yours.
Use retrieval from a controlled set of pages, show the date a source was last checked, and make the assistant say it does not know when it does not. Our guide to chatbot hallucinations and guardrails lists the technical measures. For financial aid, send applicants to official tools instead of letting a model improvise.
Privacy: FERPA is only part of the picture
FERPA protects education records of students who attend your institution, and the US Department of Education's Student Privacy Policy Office publishes the rules and guidance. Records of prospects who never enroll are generally outside FERPA, which tempts teams to assume no privacy law applies. That assumption is wrong.
State consumer privacy laws, children's privacy rules for applicants under 13 or the state equivalents, and the Federal Trade Commission's authority over deceptive claims can all reach an admissions chatbot. The Federal Trade Commission has repeatedly said that claims about AI must be truthful, which includes what your assistant says about itself. Ask counsel which state laws cover your applicants, since residency of the applicant, not the campus, often decides.
Once a prospect enrolls, the same conversation history may become part of an education record. Plan for that transition so you do not hold chats you cannot classify.
Vendor terms: the API is not ChatGPT
According to OpenAI's API data documentation, data sent to the OpenAI API is not used to train or improve OpenAI models unless you explicitly opt in, and this describes the API only, not ChatGPT apps. The terms that apply to a custom GPT used in ChatGPT are separate, and they depend on the plan. Read the terms for your exact plan and sharing setup before applicant data is involved.
By default the same page says abuse-monitoring logs are retained for up to 30 days, unless longer retention is required by law. Your records schedule and your public-records obligations, if you are a public institution, should account for that.
Logs, audit and accessibility
You need a searchable record of what the assistant said, to whom, on which source and when. A chat window configured by a staff member rarely gives you that. Public institutions may also face records-retention and open-records requests that reach chatbot transcripts.
Accessibility matters too. Section 508 and the ADA expectations on your site extend to an embedded assistant, so test it with a screen reader and keyboard before launch.
Your three routes
Build on the API with retrieval. Your developers own the model calls, the grounding, the front end and the log. It gives control and creates maintenance. OpenAI announced the removal of the Assistants API on 26 August 2026 and names the Responses API as its replacement, so new builds should target the Responses API. A bespoke build inherits every platform change.
Buy a dedicated admissions agent. A specialist product ships with consent capture, escalation, CRM connection and reporting. Skolbot is one example: a website chatbot that answers prospect questions, qualifies inquiries, books visits or appointments and pushes the lead into the CRM, with a WhatsApp add-on.
Combine them. Use a custom GPT internally to draft responses for counselors to review, and keep a dedicated agent on the public site.
| Question | Custom GPT | API build | Dedicated agent |
|---|---|---|---|
| Time to first pilot | Short | Medium | Short to medium |
| Control of data flow and logs | Limited | High | Set by contract |
| CRM and scheduling | You build it | You build it | Usually included |
| Ongoing upkeep | Staff time | Your developers | Vendor |
| Best fit | Internal drafting | Teams with engineering | Public inquiry service |
This is a planning aid, not measured data. For criteria, see how to choose an AI agent for student recruitment, and for the build-versus-buy detail, SaaS, custom or open source.
A decision path you can follow this month
- Pilot internally on public pages only and log every wrong answer.
- Compare the gaps with the table above.
- Decide whether your IT team can own an API build.
- If not, ask shortlisted vendors to demonstrate consent capture, escalation and log export.
- Complete a privacy review with counsel before any applicant uses the tool.
An illustrative workflow, not a measured result: a prospect asks about a transfer pathway late in the evening, the assistant answers from a dated page, offers a counselor callback and records the request with consent in the CRM.
Questions to put to any vendor or internal builder
Whichever route you pick, the same questions expose the weak points. Ask them in writing and keep the answers with your procurement file.
Where is applicant data processed and stored, and which sub-processors touch it? Can transcripts be exported, searched and deleted on request? What happens to a conversation when a prospect later becomes an enrolled student, and who classifies it then? How does the assistant behave when a question concerns financial aid, immigration status or a minor?
Ask also how the tool handles a prospect who says they are under 13, since children's online privacy rules can apply. For public institutions, ask how transcripts fit your records-retention schedule and open-records process. Finally, request a test: send twenty of your hardest real questions and review the answers with an enrollment counselor before anyone external sees the tool.
If the builder cannot show the consent text, the escalation path and the log in a live demonstration, treat the pilot as a prototype and keep it away from applicants.
