ChatGPT
Claude
Perplexity
Gemini
GrokA custom GPT is a quick pilot for an Irish admissions office, not a ready-made queries desk
A custom GPT can answer common questions within a day. It will not on its own record consent, keep a log you control, write to your CRM or know that a CAO date has moved. For an Irish college or university, the gaps in data protection and accuracy matter more than the model's fluency.
This guide is for admissions, marketing and IT leads choosing between a custom GPT, a build on the API and a dedicated admissions agent. It builds on our guide to AI chatbots for student recruitment.
What a custom GPT contains
A custom GPT is ChatGPT configured with your instructions, documents and optional connections. 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.
The chat is quick to set up and the integration is not. Writing a query into your CRM needs authentication, error handling and monitoring, and a named person has to own it after the pilot ends.
Where it helps and where it stops
| Admissions job | Custom GPT alone | Gap to close |
|---|---|---|
| Answering questions from public course pages | Reasonable for a pilot | Keep sources current |
| CAO points, Leaving Cert subject requirements | Risky | Dated source; decline when unsure |
| Change of Mind and Available Places timing | Risky | Verified calendar |
| Grants, the Student Contribution and SUSI questions | Risky | Link to official sources; hand off |
| QQI levels, PLC and mature-applicant routes | Risky | Verified pathway content |
| Capturing a query with consent | Not built in | Consent text and stored record |
| Writing to the CRM | Needs custom API work | Endpoint, authentication, alerts |
| Booking an open day | Needs custom API work | Calendar integration |
Most Irish undergraduate applicants apply through the Central Applications Office, so an assistant must be exact about what the CAO handles and what your institution handles directly. Applicants who get that wrong can miss a step.
Accuracy under pressure
Points requirements, offer rounds and deadlines change by year and course. A model states a figure fluently whether or not its source is current. An applicant who plans a Change of Mind around a wrong answer will not forgive it easily.
Use retrieval from sources you maintain, show the date of the source, and make the assistant say it is unsure when evidence is thin. Our guide to chatbot hallucinations and guardrails lists the measures. For quality claims about programmes, QQI is the reference point for qualification levels and quality assurance, and your compliance lead should confirm how automated answers fit.
GDPR and the Data Protection Commission
Ireland applies the GDPR directly, supplemented by the Data Protection Act 2018. Applicants may reveal health, disability or family circumstances in a chat, so you need a lawful basis, a privacy notice that mentions the assistant, a data protection impact assessment and a clear view of who the processor is.
The Data Protection Commission publishes guidance for organisations, and it is also the supervisory authority you would deal with in the event of a complaint or breach. Bring your data protection officer in before launch. If applicants under 18 may use the assistant, say how that is handled.
International transfers deserve a specific line in the assessment. If a vendor processes data outside the EEA, you need a documented transfer mechanism and a view on the vendor's sub-processors.
Vendor terms
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 for a custom GPT inside ChatGPT are separate and 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 retention schedule and DPIA should reflect that.
Logs and escalation
Keep a searchable record of what the assistant said, when and on which source, and decide how long you hold transcripts and how a data subject request is handled. Define how a conversation passes to a staff member with its context. A chat window set up by one person rarely offers either in an exportable form.
Three routes beyond the pilot
Build on the API with retrieval. Your developers own the model calls, grounding, front end and log. You gain control and accept 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.
Buy a dedicated admissions agent. Specialist products ship with consent capture, escalation, CRM connectors and reporting. Skolbot is one example: a website chatbot that answers prospect questions, qualifies queries, books open days or appointments and pushes leads to the CRM, with a WhatsApp add-on.
Combine them. Use a custom GPT internally to draft replies for advisers, and 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 | Limited | High | Set by contract |
| CRM and booking | You build it | You build it | Usually included |
| Upkeep | Staff time | Your developers | Vendor |
| Best fit | Internal drafting | Engineering capacity | Public queries service |
This is a planning aid, not measured data. For criteria, read how to choose an AI agent for student recruitment, and for build versus buy, SaaS, custom or open source.
A path for the next quarter
- Pilot internally on public pages and record every wrong answer.
- Compare the failures with the table above.
- Decide whether you can own an API build.
- If not, ask vendors to show consent capture, escalation, log export and processing location.
- Complete a DPIA before applicants use the tool.
An illustrative workflow, not a measured result: a mature applicant asks about a PLC route one evening, the assistant answers from a dated page, offers a callback and records the query 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 when a data subject asks? Does the vendor give you the documents needed for a DPIA, including the transfer mechanism for any processing outside the EEA?
Ask how the assistant treats questions that depend on CAO rules, such as points, restricted-entry assessments or the Change of Mind period, and whether it hands them to staff. Ask how it behaves with a mature applicant or an international enquirer whose route differs from the standard Leaving Cert path. Then run a trial: send twenty of your hardest real questions and review the answers with an admissions officer 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.
