Practical AI implementation guides, comparisons and the business implications of global developments.
12 articles listed
Nine proprietary models from four vendors: list price, cache reads, context window, clouds and EU data handling. When is each the reasonable choice?
OpenAI tightened the GPT-6 Astra system card on 9 September. What it means for access control, logging and your procurement file in Hungary and CEE.
Same list price, cache reads at a quarter on Fable 5.1, and a different leader depending on who ran the benchmark. When each model is the reasonable choice.
GPT-6 Astra launched on September 3: faster computer use, $10/$50 per million tokens, and OpenAI's first model at the Critical cybersecurity level.
A prioritized list of automatable processes with cost and payback estimates. From HUF 500,000 + VAT; what it contains, and when to pilot instead.
Per Anthropic, a typical workload costs about 25% less at unchanged token prices; the switch pays off most for long-running agentic systems.
Only the provider's watermark shows reliably that a text was written by AI; post-hoc detectors are inaccurate. Gemini watermarks since 2024, Claude since 2026.
Step by step: what the assistant should do (inquiry summaries, quote drafts, follow-ups), what your CRM must provide, and where a human stays in the loop.
Copilot is enough while the work stays in Teams and SharePoint. When a process posts into an ERP with no ready connector and runs unattended, build it custom.
If data cannot leave the company, on-premise; if speed and model quality decide, cloud; most systems we build end up hybrid. The decision criteria.
Data on eight AI consulting firms in Hungary, on identical criteria: founding, focus, references, pricing, EU AI Act. As of 20 August and 4 September 2026.
The four risk categories, provider and deployer roles, what the high-risk obligations mean in practice, and a realistic preparation sequence.