Generative AI tools and LLM integration for your organisation
Generative AI tools use large language models and other generative models to draft, summarise, answer questions and create images from a plain-language prompt. Arke Technologies, in Johannesburg, selects, integrates and governs generative AI for South African organisations, public bodies and schools, connecting models to your own documents and systems under POPIA controls.
What is generative AI?
Generative AI is a type of artificial intelligence that creates new content, such as text, images, audio, video or code, by learning patterns in large amounts of existing data. Defined simply: traditional AI classifies or predicts; generative AI produces.
Is generative AI a type of AI? Yes, a branch of machine learning, and most generative AI models are deep neural networks. Generative AI and large language models are often treated as one thing, but a large language model (LLM) is one kind of generative AI model, trained on text to predict what comes next. The main generative AI algorithms are:
- Transformers, the architecture behind LLMs such as Claude and Gemini.
- Diffusion models, which build an image or video by removing noise step by step; most generative AI images and video come from them.
- Generative adversarial networks (GANs), where one network generates and another judges.
- Variational autoencoders, which compress data and generate variations of it.
Is generative AI supervised or unsupervised? Both, at different stages: models are pre-trained with self-supervised learning on unlabelled data, then refined with supervised fine-tuning and human feedback. Read what is generative AI and AI agents vs generative AI. This service is part of our AI consulting and implementation practice.
How we integrate large language models
The value comes from connecting a model to your own information. Our generative AI services cover LLM integration end to end:
- Grounding. Retrieval-augmented generation (RAG) searches approved sources, such as policies, contracts, SharePoint libraries or an internal wiki, and gives the model only the passages it needs, so answers cite their source.
- Systems. Through APIs, the model reads a case or ticket and writes back where staff work; see API development and integration.
- Prompts. Tested instructions, versioned like code. Two generative AI prompt examples of the kind we write: "Summarise this tender response against the evaluation criteria in the attached table, citing the page for each point" and "Draft a reply to this ratepayer query using only the tariff policy provided; if the policy does not answer it, say so."
- Guardrails. Input and output filtering, refusal rules and human review where stakes warrant it.
- Cost controls. Logging, budgets, caching and a smaller model wherever it performs as well.
When a task needs several steps and tools, it becomes an agent; see AI agents for business.
Generative AI limitations and legal issues
Generative AI is strong at language and weak on facts it has not been given. A generative AI summit keynote shows the best case; plan for the ordinary one. The main generative AI limitations and legal issues:
Hallucination. Fluent, confident and wrong. We ground models in approved sources, require citations and keep people reviewing where it matters.- Stale knowledge. A model knows only its training data unless given current information.
- Bias. Outputs can reflect bias in training data, which matters in HR, credit and public services.
- Personal information. POPIA section 72 restricts transfers abroad, and section 71 restricts decisions based solely on automated processing.
- Copyright. The law on training data and on ownership of generated work is unsettled in several countries, so generated images and copy get human review and a record of how they were made.
- Advertising claims. Generated marketing copy must meet the Consumer Protection Act like any other.
- Prompt injection and leakage. Hidden instructions in a document can redirect a model, and staff pasting data into free tools can leak it. We apply filtering, least-privilege access, an acceptable-use policy and approved tools.
Generative AI skills, training and certification
Adoption decides the return. Staff need generative AI skills that fit their role: writing a clear prompt, knowing what may go into one, checking an output and reporting a wrong answer. We train on your own documents and tools.
For technical staff, a professional certification from a cloud provider such as AWS, Google Cloud or Microsoft shows grounding in the fundamentals, and several providers publish free introductory courses. A generative AI certification is a useful start, not a substitute for experience with your data, controls and users.
Governance, compliance, procurement and continuity
- Governance and reporting. A named lead; a monthly report on sampled accuracy, usage, cost and incidents; escalation paths; change control for prompts, sources and model versions. AI recommends, people decide, and every AI action is logged.
- Compliance. POPIA controls: data minimisation, encryption in transit and at rest, access controls and audit logging, and a documented basis under section 72 wherever prompts or documents are processed outside South Africa; ISO 27001-aligned security; King IV Principle 12 reporting; PFMA or MFMA documentation for public bodies. We never train models on client data without explicit consent.
- Procurement. Level 1 B-BBEE contributor, giving 135% procurement recognition (sworn affidavit available on request); company documents, bank confirmation letter, CSD supplier number and SARS tax compliance status on request; standard bidding documents (SBD forms) completed when a bid requires them. Buy by proposal, SLA and purchase order, or by tender; see procuring from Arke.
- Continuity. One contract and one accountable team cover the build and its support. You own your data, prompts, configurations, code and documentation; because we build so the model can be swapped, a model change does not mean a rebuild, and on exit we hand over everything, credentials included.