Vela The Next Chapter Join waitlist

AI & Technology

When AI helps. When it doesn't.

Last updated: May 2026

1. Overview

Vela uses a combination of deterministic matching logic and, in limited circumstances, AI language model assistance. This policy explains what each component does, where AI is used, where it is not, and how we ensure that guidance remains accurate and accountable.

2. How the Matching Engine Works

The core of Vela's guidance service is a deterministic matching engine — a rules-based system applied to human-verified data. It does not use AI. Given the same learner profile, verified data set, and season, it will always produce the same result.

The engine:

  • Calculates your APS score using the NSC rules stored in our database
  • Filters programmes against your marks, subjects, province, and pathway preference
  • Ranks matches using a scoring formula applied to your interest discovery responses and employment demand data
  • Evaluates funding eligibility against programme-specific and learner-specific rules
  • Builds a 5-step action plan ordered by deadline urgency

No AI model makes eligibility decisions, calculates your APS, or determines which bursaries you qualify for. These are arithmetic and rule-based operations applied to vetted source data.

3. Where AI Is Used

Vela uses Claude Haiku, developed by Anthropic (anthropic.com), as a fallback for off-flow inputs — messages that fall outside the structured conversation flow, such as free-text questions or unclear responses.

Claude Haiku is used to:

  • Provide a helpful response when a learner sends a message that the structured bot cannot interpret
  • Clarify a question or re-direct the learner to the correct step in the flow

Claude Haiku is not used to:

  • Decide which programmes a learner qualifies for
  • Assess a learner's APS score
  • Determine bursary or NSFAS eligibility
  • Generate salary ranges or employment data
  • Write the structured sections of a learner's action plan

Where AI-generated text appears in a Vela response, it is labelled or contextualised appropriately. Structured plan content (route names, fit indicators, deadlines, bursary amounts) is always drawn from the verified database, not generated by AI.

4. Learner Data and AI Training

Learner personal information is never shared with or used to train any AI model, including Claude or any other third-party model. Data sent to the Claude Haiku API for off-flow handling is used only to generate a response to that specific message — it is not retained by Anthropic for training purposes under our current API agreement. For details of Anthropic's data practices, see anthropic.com/privacy.

Our matching engine, interest scoring weights, and eligibility rules were developed by Vela's team in collaboration with our advisory board. They were not trained by an AI model on learner data.

5. Human-Verified Data

All programme requirements, application deadlines, bursary criteria, salary ranges, and career path information in Vela's database is verified by a human against authoritative South African sources before it is activated. No AI-generated content is activated in the matching database without human review. Every record carries a verified_date, source URL, and the identity of the person who verified it.

This human verification layer is the primary safeguard against AI-generated inaccuracies appearing in learner guidance.

6. Accuracy and Known Limitations

AI language models can produce plausible-sounding but incorrect information — this is sometimes called "hallucination." Because Vela's structured guidance is drawn from a vetted database rather than AI generation, hallucination risk in core guidance is low. The residual risk exists in off-flow AI-assisted responses.

If you receive a response from Vela that appears factually incorrect — particularly any claims about specific programme requirements, deadlines, bursary amounts, or salary figures — please report it to hello@velaguide.co.za. We treat accuracy errors as production failures.

7. Decisions That Remain Human

The following decisions are made by humans at Vela or by the institutions and funders involved — never by AI:

  • Whether a data source is approved for use in the matching engine
  • Whether a programme or bursary record is activated or deactivated
  • Whether a salary range is accurate enough to surface to a learner
  • Whether a sponsored tier learner qualifies for free access
  • Any manual intervention during a bot outage
  • Admission decisions (made by institutions)
  • Bursary award decisions (made by funders)

8. Oversight and Accountability

Vela's use of AI is reviewed as part of our annual compliance cycle. We track AI-assisted responses in our analytics system (without personal information) to identify patterns of off-flow queries that should be addressed by improving the structured flow. Our goal is to reduce reliance on AI fallback over time, not expand it.

9. Updates

As AI technology and regulation in South Africa develop, we will update this policy accordingly. South Africa is developing a National AI Policy Framework — we will align with applicable obligations as they are enacted.

10. Contact

AI policy questions: hello@velaguide.co.za