When AI Agentic Beats a Single LLM for RO Business Processes
Ask a mid-market Romanian CFO why last month's invoice-matching automation failed, and the answer is specific: the single-LLM prompt could not handle both the SAF-T XML export and the ANAF e-Factura schema in one cont…
When AI Agentic Beats a Single LLM for RO Business Processes
Ask a mid-market Romanian CFO why last month’s invoice-matching automation failed, and the answer is specific: the single-LLM prompt could not handle both the SAF-T XML export and the ANAF e-Factura schema in one context window without silently dropping line items. That is the exact break-point where an orchestrator plus specialized agents starts paying its keep.
Single-LLM works until the process forks
A single-LLM setup — one model, one prompt, one call — is defensible when the task is bounded: summarize a contract, classify a ticket, draft a reply. The moment the process forks into parallel branches with different tools, data-access rules, and audit requirements, one prompt cannot hold state cleanly. The NIST AI Risk Management Framework treats this as a governance boundary: measurability drops as scope widens.
For Romanian workflows tied to Directive (EU) 2022/2555 (NIS2) or OPANAF 1.783/2022 on e-Factura, the audit trail must show which component touched which record. A monolithic prompt cannot produce that trail — a point our team has argued in the context of the AI Act timeline shift.
What an orchestrator gives you
An orchestrator routes a task to specialized agents, each with a narrow tool belt and its own memory scope. Concretely, for a RO purchase-to-pay flow:
orchestrator: p2p_router_v3
agents:
- name: efactura_ingest
tools: [anaf_spv_client, xml_validator]
scope: read-only
- name: gl_matcher
tools: [erp_query, sql_readonly]
max_tokens_per_call: 8000
- name: exception_handler
tools: [teams_notify, ticket_create]
requires_human_review: true
audit_log: /var/log/p2p/2026-08-26.jsonlEach agent’s calls land in a single append-only log. That log satisfies both the GDPR Article 30 records-of-processing obligation and internal SOX-style controls.
The pattern also lets you version and swap one agent without retraining the whole flow — critical when the ENISA multilayer framework for AI cybersecurity evolves and you need to harden a specific interface.
When it does not make sense
Do not deploy an orchestrator for a 200-token classification task. The routing overhead, the extra tool round-trips, and the observability stack cost more than the single LLM call they replace. Our internal rule at CAI: fewer than three distinct tools, fewer than two data-domain boundaries, single audience — stay with one model. Cross either threshold and the orchestrator wins on cost within one quarter. The failure clusters we track in agentic systems show that most outages stem from adding an orchestrator too early, not too late.
Our position
Agentic architecture is a governance choice before it is an engineering choice. The real question is not “which framework?” but “which regulator will read the log?” — because that reader dictates where the boundaries between agents fall. Have a look at our IRIS pillar work on how agents pick tools before you commit to a topology.