By Greg Nowak. Last updated 2026-10-04.
Agentic AI becomes interesting when a piece of work needs more than a good answer. An agent can inspect information, choose an approved tool, act, check the result, and decide whether another step is needed. For a business, that might mean preparing a customer record from an enquiry or gathering evidence for a service request.
The question is whether that freedom improves a real process. If the steps are predictable, ordinary automation is usually easier to run. If the work needs interpretation but a person will take the next action, an AI assistant may be enough. An agent earns its place when the route to a defined outcome varies from case to case.
What makes an AI system an agent?
An agent works in a loop: understand the task, select an action, use a tool, inspect what happened, then continue, finish, or ask for help. Its instructions define the goal and boundaries; its tools provide access to documents, search, APIs, or business systems. A record of the current task helps it avoid repeating work and supports review when something goes wrong.
That does not make it an independent employee. The organisation still decides which systems it may reach, which actions it may take, and when a person must approve the result. It also needs a stopping rule, such as a maximum number of steps, a time limit, or an escalation after repeated failures.
| Choose | When the work looks like this | Example |
|---|---|---|
| Rules-based automation | Inputs and steps are stable; exceptions are rare. | Copy a validated form submission into a CRM. |
| AI assistant | Interpretation or drafting helps, but a person remains responsible for acting. | Summarise a brief and suggest follow-up questions. |
| Agentic workflow | The next step depends on what the system finds, and it must use several approved tools. | Check an enquiry, find missing details, prepare a CRM update, and route a draft reply for approval. |
This progression reflects the advice in OpenAI’s agent guide and Anthropic’s engineering guidance: add model-driven steps where simpler paths struggle, and start with one agent before coordinating several.
Where is the extra complexity justified?
Look for work with a clear finish line but variable inputs: enquiries written in different formats, documents that need interpretation, requests that require checking several systems, or exceptions that have made a rules engine difficult to maintain. Good early candidates are frequent enough to measure and low risk enough that mistakes can be caught and corrected.
Follow a few real cases from arrival to completion before selecting software. Ask the people doing the work what information they check, which decisions require judgement, where they re-enter data, and what happens when information is missing. If the team cannot agree on what a correct outcome looks like, clarify the process first. An agent cannot make an unclear operating policy dependable.
For example, an agency could use an agent to read an incoming brief, identify missing requirements, consult approved public information, and prepare a CRM entry plus a draft response. The useful boundary is clear: staff review the record and message before either is committed or sent. Pricing, accepting work, and making binding promises remain explicit human decisions.
Give the agent authority one action at a time
Start with the access needed to observe the process. Add write permissions only for named actions with a defined owner. A read-only company lookup carries different consequences from changing a customer record, sending an email, deleting a file, or initiating a payment. Build approval around those consequences, not around how confident the model sounds.
Keep credentials in a secrets system, outside prompts and uploaded files. Treat webpages, emails, retrieved documents, and tool responses as data rather than instructions: they may contain text that tries to redirect the agent. OWASP’s agentic application guidance is a useful reference when reviewing tool misuse and other security risks. Technical controls should include narrow permissions, checks on tool inputs, and a safe path to stop or hand the task to a person.
Design for failure as well as success. A tool can time out after a change was made, so a blind retry may create a duplicate. Record whether an action completed, make repeat operations safe where possible, and send ambiguous outcomes for review. Keep a useful audit trail of the request, tool calls, approvals, result, and workflow version while limiting the personal data and secrets it retains.
How to run a pilot that answers a business question
- Define one outcome. Choose a bounded workflow and state what a correct completion, an escalation, and an unacceptable error look like.
- Measure the current process. Capture handling time, corrections, and hand-offs so the pilot has a fair comparison.
- Build a small test set. Include ordinary cases, missing information, conflicting instructions, tool failures, and content that tries to change the agent’s goal.
- Run with review. Let the agent prepare work while a person approves consequential actions. Record why reviewers change or reject its output.
- Expand only with evidence. Compare completed outcomes, correction effort, escalation quality, time, and cost. Remove an approval step only when the results support it.
Watch the work created around the agent, too. A quick draft that requires substantial cleanup has not improved the process. Someone should own the instructions, permissions, test cases, and incident response after the pilot; NIST’s generative AI risk profile provides a broader framework for that ongoing responsibility.
Agentic AI is most useful when it joins sound process design with carefully limited action. If you have a workflow in mind, Greg can help map it, choose an appropriate architecture, and shape a pilot your team can actually operate. See how Greg works as a digital project manager.
Related on GrN.dk
- Microsoft Access Database Resources: When to Fix, Split, or Replace Your System
- Your AI Agent Has Shell Access. What Can It Reach?
- A Voice Agent Is Only Ready When the Human Handoff Works
Need help with this kind of work?
Discuss a practical AI pilot with Greg Get in touch with Greg.
