Connected systems before AI agents

AI agents are often presented as the next layer of business automation: assistants that can plan, search, use tools, update systems and complete tasks across workflows.

The idea is attractive. Many companies have recurring work that could be faster, clearer or less manual. Finance teams chase receipts and reconcile exports. Sales and service teams search across CRM, ERP, documentation and inboxes. Managers wait for reports assembled from several systems. Customer portals show partial information because the back-end systems do not agree.

But an AI agent cannot solve disconnected systems by itself. If the underlying data is fragmented, permissions are unclear and integrations are fragile, the agent inherits the mess.

Before companies add agents, they often need to connect the systems that already run the business.

Integration defines what AI can do

An AI agent is only as useful as the tools, data and workflows it can safely access.

If it can only read a document, it may summarise. If it can query a database, it may retrieve current information. If it can call an API, it may prepare an action. If it can write back to a system, it may change business records. Each step increases potential value, but also increases the need for control.

That is why integration is not technical plumbing. It defines the boundary of automation.

A connected workflow needs answers to practical questions:

• Which systems contain the source of truth?
• Which data may be read by AI?
• Which actions may be prepared automatically?
• Which actions require human approval?
• Which permissions does the agent or service account need?
• What should be logged?
• Who reviews exceptions?
• What happens if the integration fails?

Without those answers, an agent becomes another layer on top of existing uncertainty.

Disconnected systems create the same problems in many places

Different business problems often have the same technical shape.

A finance team spends time reconciling spreadsheets because bank, expense, accounting and reporting tools do not exchange information cleanly.

A B2B customer portal disappoints users because product data, pricing, documentation, service history and order status live in separate systems.

A management report arrives late because data is exported manually from several platforms before someone cleans and combines it.

A support team answers the same questions repeatedly because policies, product documentation and customer context are not available where the work happens.

These are not primarily AI problems. They are system and workflow problems. AI may help once the underlying data flow is understood, but it cannot reliably automate a process that has no reliable process shape.

Why agents increase the need for control

Traditional automation usually follows defined rules. It may move a file, update a record, send a notification or run a scheduled job.

Agentic workflows are different. They may involve repeated reasoning steps, tool calls, retrieval, classification and conditional action. Gartner has warned that over 40% of agentic AI projects may be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.

OWASP also highlights “Excessive Agency” as a risk in LLM-based systems: when an AI system has too much functionality, permission or autonomy, it can perform damaging actions in response to unexpected or manipulated outputs.

That does not mean agents should be avoided. It means they should be designed like operational software.

What connected-system readiness looks like

Before adding an agent to a workflow, companies should usually work through six steps.

1. Map the workflow

Start with the recurring work, not the AI tool. Identify the trigger, inputs, decision points, systems, handoffs, exceptions and outputs.

The goal is to see what actually happens, including the informal workarounds.

2. Identify system and data boundaries

List where the relevant data lives. Separate system-of-record data from copied data, exported data and manually maintained spreadsheets.

This often reveals why the process is slow: the workflow crosses systems that were never designed to work together.

3. Connect the core systems

Build or improve APIs, data synchronisation, reporting pipelines or integration layers. The objective is not integration for its own sake. It is to make the workflow reliable enough for automation.

4. Automate simple steps first

Many useful improvements do not require AI. A status sync, approval notification, report generation job or validation rule may remove friction before any model is involved.

This also reduces the surface area for the eventual AI component.

5. Add AI where judgement or language handling is useful

AI is strongest where the workflow involves unstructured text, classification, summarisation, retrieval, drafting, triage or exception preparation.

Examples include extracting structured data from documents, summarising support history, classifying requests, preparing approval notes or searching internal documentation.

6. Monitor and maintain the workflow

A connected agentic workflow needs logs, alerts, cost tracking, permission reviews, model or prompt updates and a clear owner. It should not become a hidden automation that nobody can explain or stop.

A practical architecture pattern

A simple system-connected AI workflow may look like this:

Business systems → API / data layer → Workflow service → AI assistant or agent → Human review → System action

The important part is not the diagram. It is the separation of responsibility.

• Business systems remain the source of truth.
• The API or data layer controls access and structure.
• The workflow service defines the process.
• AI assists with language, retrieval or preparation.
• Humans approve high-impact actions.
• System actions are logged and monitored.

This is more robust than giving an AI tool broad access to everything and hoping it behaves.

The practical takeaway

AI agents should not be treated as shortcuts around system integration. They make integration more important.

If the business wants AI to support real work, it needs connected systems, clear data ownership, access boundaries, workflow design, logging and maintenance. The agent is not the foundation. It is a layer on top of the foundation.

Sources and inspiration

• Computerworld — Du klatter kostbare ressourcer væk på ineffektive arbejdsgange og for dårlig brug af digitale værktøjer: https://www.computerworld.dk/art/295612/du-klatter-kostbare-ressourcer-vaek-paa-ineffektive-arbejdsgange-og-for-daarlig-brug-af-digitale-vaerktoejer
• Computerworld — Regeringen jubler over AI-guld: Men ude i virkeligheden kæmper virksomhederne stadig med det samme gamle rod: https://www.computerworld.dk/art/295500/regeringen-jubler-over-ai-guld-men-ude-i-virkeligheden-kaemper-virksomhederne-stadig-med-det-samme-gamle-rod
• Computerworld — Nordisk industri har digitaliseret alt - undtagen kunden: https://www.computerworld.dk/art/295547/nordisk-industri-har-digitaliseret-alt-undtagen-kunden
• Computerworld — Din dyreste medarbejder i 2027 er en AI-agent, som du har glemt at slukke: https://www.computerworld.dk/art/295703/din-dyreste-medarbejder-i-2027-er-en-ai-agent-som-du-har-glemt-at-slukke
• Gartner — Over 40% of agentic AI projects may be cancelled by end-2027: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
• OWASP — Excessive Agency in LLM applications: https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
• BCG — Nordic AI: Value Creation or Value Bubble?: https://www.bcg.com/publications/2026/nordic-ai-value-creation-or-bubble
• McKinsey — B2B omnichannel experience and switching risk: https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/five-fundamental-truths-how-b2b-winners-keep-growing
• NIST AI RMF Core — Govern, Map, Measure, Manage: https://airc.nist.gov/airmf-resources/airmf/5-sec-core/