Fixed-access network investigation remains labour-intensive because faults are expressed across heterogeneous evidence: customer reports, topology, alarms, optical measurements, device telemetry, configuration histories, tickets, software releases and operational knowledge. This review examines how retrieval-augmented generation (RAG) and agentic artificial intelligence can be combined to automate investigation without allowing a language model to substitute plausible narrative for verified engineering evidence. The study synthesises research published from 2020 to 2025 on retrieval, tool-using language agents, root-cause analysis, optical-access diagnostics, intent-based management and zero-touch operations, together with relevant fixed-access standards. It proposes an evidence-governed architecture in which a planning agent decomposes an incident, permission-aware retrievers assemble time-bounded context, specialised tools execute deterministic queries, a causal reasoning layer ranks hypotheses, and an independent verifier checks numerical results, provenance and policy compliance. The review argues that the most defensible role for generative models is orchestration and explanation, whereas calculations, state inspection and change execution should remain within typed, auditable tools. A staged workflow is developed for incident framing, evidence alignment, topology reconstruction, hypothesis testing, counterfactual checking, confidence calibration, human approval and organisational learning. Evaluation should therefore measure retrieval coverage, factual and numerical correctness, root-cause ranking, time-to-isolation, unsafe-action rate and operator trust rather than linguistic fluency alone. The resulting framework supports faster and more consistent diagnosis across fibre, copper, residential gateway and disaggregated access environments while preserving human authority over consequential network actions.