Eval
Against the client's dataset, not benchmarks.
IA-02 · Applied AI
Back-office agents
AI agents that execute multi-step tasks with reasoning. Critical actions are always signed off by a human until the confidence threshold is exceeded.
There is a class of tasks — classify, decide, execute steps in sequence, call external systems, synthesise — that current language models handle well with the right agent architecture. E-commerce returns, invoice reading and logging, quotes, first-level support, reconciliation.
We build agents with LangGraph, Anthropic and OpenAI models, and always with a validation layer. The critical action — the one that costs money or exposes the client — is signed off by a human until the model demonstrates statistical stability above the threshold.
In closed pilots, we achieved 78% autonomous resolution on fashion e-commerce returns, with an error rate below 1.5% and a cost per case of €0.28 versus €8.40 for a human agent.
AgentReceives return emailTrigger · webhookConfidence: high
ToolQueries order in ShopifyAPI · GET /ordersLatency: 240ms
ToolGenerates return labelAPI · POST /returnsLatency: 410ms
AgentDrafts response to the customerLLM · Claude 4Cost: €0.02
HumanApproves if refund > €200Threshold · policyLatency: 8 min
The Back-office agents process.
The process · four stagesBuild
Agent with tools, validation and escalation.
Pilot
4–6 weeks in production with monitoring.
Production
Deployment if the eval justifies it.
What is included
What Back-office agents includes.
The operational detail: what we deliver as part of the engagement and what we keep active afterwards.
-
Reasoning agents
Returns, support, reconciliation, invoices, quotes.
-
Proprietary eval framework
On the client's dataset. Business metrics.
-
Human in the loop
Critical action with a path to a human.
-
Continuous monitoring
Accuracy, latency, cost, scaling.
-
ERP and CRM integration
Via Summum Sistemas.
-
AI Act documentation
Annex IV where applicable.
Regulatory framework
The regulatory framework.
Agents may fall under the AI Act high-risk system category.
EU
AI Act
Applicable to this service
ISO
42001
Applicable to this service
RGPD
art.22
Applicable to this service
ENS
ENIA
Applicable to this service
Summum cluster
How it intersects with related services.
Production agents are not just AI: ERP, regulatory and SOC.
Frequently asked questions about Back-office agents.
What if the pilot fails?
We do not charge for implementation. It is contractual.
Does it replace people?
It usually reassigns. The team moves to complex queues.
Which models?
Multi-model. Claude for reasoning, GPT-4 for vision.