What is an AI Agent?

Most enterprise software pitches use "AI agent" as a vague honorific. Vendors apply it to chatbots, to BPM workflows, and to anything where a model is involved. The label is now confused enough that buying conversations break down inside the first slide.

Three traits separate an agent from the other things that get the label:

1. Tool use. The agent calls APIs, queries databases, runs code. Strip the tools and you have a chatbot.
2. Planning. The agent breaks a goal into steps before acting, and revises the plan when results require it.
3. Iteration. The agent observes results, decides whether more steps are needed, and loops until done. A single LLM call is not iteration.

A concrete example. A Finnish mid-market company receives 2,000 supplier invoices a month. An AI agent reads each invoice, queries the ERP for the matching purchase order, checks the supplier database for current payment terms, computes any deviation, and either books the invoice or routes the exception to a human. That is an agent. It uses tools (ERP, supplier DB), it plans (decides which checks to run for which invoice type), and it iterates (loops over invoices, handles errors, retries).

Agents earn their cost in three situations:

Multi-system tasks. The work spans multiple internal systems, each with its own schema and quirks. Invoice processing across ERP plus procurement plus vendor portal. Customer onboarding across CRM plus KYC plus legal review plus billing. Sales research across LinkedIn plus CRM plus news plus internal notes. A workflow can do this if every path is predictable. When the path varies by case, the agent's reasoning is what makes the work feasible without a rule engine that grows forever.

Variable input shapes. The incoming data is messy, semi-structured, and changes form over time. Free-text emails. PDFs with non-standard layouts. Transcripts of meetings. Logs from heterogeneous tools. A traditional pipeline breaks when the input changes. An agent can adapt because the LLM at its core handles ambiguity natively.

Decisions that need context. The right next step depends on combining information from several places. "Should this contract amendment go to legal?" depends on the contract value, the clause that changed, the jurisdiction, and the supplier's track record. Hard-coding all combinations is brittle. An agent reasons through them.

Nordic examples we have seen in client work:

• Procurement teams using agents to triage incoming RFPs, classify by capability, and route to the right account manager.
• Finance teams using agents to reconcile cross-border invoices that include VAT, currency, and supplier-specific terms.
• Customer success teams using agents to summarise CRM history before each meeting and draft a follow-up.
• Engineering teams using agents to triage incoming bug reports, link them to existing issues, and propose a fix.

Five components matter. Skipping any one usually causes the system to fail in production.

The foundation model. Claude (Anthropic), GPT (OpenAI), Gemini (Google) for closed-weight via API. Llama (Meta), Mistral (Mistral AI), DeepSeek for open-weight self-hosted. The choice depends on cost, latency, language support, and data-residency requirements. Nordic clients with strict data-residency rules tend to land on self-hosted Llama or Mistral. Clients comfortable with no-retention API agreements tend to land on Claude or GPT.

The orchestration layer. Frameworks like LangChain, LlamaIndex, the Anthropic SDK, OpenAI Agents SDK, CrewAI, AutoGen exist. In production, many teams hand-roll a thin orchestration layer in Python or TypeScript. The frameworks add abstraction that obscures the LLM calls and complicates debugging. They are useful for prototyping; less useful at scale.

Tools. The functions the agent can call. APIs to internal systems, database queries, document retrieval, payment APIs. MCP (Model Context Protocol) is becoming the default standard for exposing tools to agents because it standardises authentication and audit logging.

Memory. Short-term memory for the current task (the model's context window). Long-term memory for facts the agent should remember across sessions (a vector database, often Postgres with pgvector, or Qdrant, Weaviate, Pinecone). Memory is the most common source of agent failure: too much of it overwhelms the model; too little leaves it amnesiac.

Evaluation. A test suite that runs continuously. Deterministic tests where possible. LLM-as-judge for subjective output. Human-in-the-loop sampling for edge cases. Evals are non-negotiable. An agent without evals is an agent without quality control, and quality regresses silently as the underlying model updates.