How to Build an AI Agent for Enterprise, Step by Step

An AI agent is useful in an enterprise when it does one job the same way every time. That sounds modest, and it is the whole point: a triage agent that classifies tickets consistently is worth more than a brilliant assistant that answers differently on Tuesday. This lesson builds that kind of agent, an IT service desk triage agent, and shows each decision along the way.

You will finish with an agent that reads a support ticket, sets a category and a priority using your incident priority matrix, and drafts a first reply to the requester, returned as JSON that other systems can read.

Step 1: Write the agent brief

Before you open the builder, write four lines on paper. They become the agent’s description and instructions, and they are what reviewers will ask about later.

• The job. Triage incoming IT support tickets: classify, prioritise, draft a first reply.
• The inputs. The ticket text as the requester wrote it. Nothing else.
• The output. Category, priority from P1 to P4, a one-sentence reason and the reply.
• Done means. An engineer can act on the priority without re-reading the ticket, and the reply is safe to send.

The brief also tells you what the agent must not do. This one never resets passwords, never asks for credentials and never promises a fix time. Scope written down now is scope you can test later.

Step 2: Name the agent and fix its output format

In Agents Hub, open Agent Builder and choose Create New Agent, then Create Manually. The builder has five steps: Basics, Model, Tools & Skills, Prompt and Review. Agents created here are assigned to the workspaces selected at the top of the page.

Fill in the Basics:

• Agent Name: it_ticket_triage. Names take letters, numbers and underscores.
• Version: 1.0. You will create 1.1 when testing changes the instructions.
• Domain: General.
• Description: the job line from your brief.
• Output Format: JSON. This is the most important field on the page for an agent whose answer feeds another system.
• Competencies: ticket triage and service desk. These tags describe what the agent is good at and help people find it.

Step 3: Choose a model

The Model step lists the models your deployment makes available. Search by name and pick one suited to short classification and drafting. This lesson uses Qwen3 Next 80B A3B Instruct, a capable open-weight model that answers quickly.

Advanced parameters such as temperature and maximum tokens can stay at their defaults for now. If you later see the agent phrase the same ticket in very different ways, a lower temperature is the first setting to try.

Step 4: Leave the tools empty for now

The Tools & Skills step lets you grant tools the agent may call and skills it may load. Triage needs only the ticket text, so grant nothing. An agent can only reach what it has been given, and an agent with no tools cannot act on any system.

Step 5: Write the instructions

The Prompt step holds the system prompt: the standing instructions the agent follows on every run. Write it from your brief, in this order: role, method, rules, output.

You are the IT service desk triage assistant for our organisation.

For every ticket you receive:
1. Classify it into exactly one category: Access, Hardware, Network,
   Email & Calendar, Collaboration, Software, Security or Service Management.
2. Set a priority using the incident priority matrix:
   - P1: a whole site, a business-critical system or a security incident,
     including lost devices and compromised credentials.
   - P2: a team or a key person cannot work and there is no workaround.
   - P3: one person is affected, or a workaround exists.
   - P4: requests, questions and cosmetic issues.
3. Draft a short, polite first reply in British English.

Rules:
- Never ask for, or repeat, passwords, MFA codes or recovery keys.
- If the ticket mentions a lost or stolen device, a clicked phishing link
  or an entered password, set P1 and tell the requester to phone the
  service desk now.

Return JSON only:
{"category": "...", "priority": "P1|P2|P3|P4",
 "reasoning": "one sentence", "reply": "..."}

Two habits matter more than clever wording. Put the priority matrix itself in the instructions rather than asking the model to guess what P1 means, and end with the exact JSON shape so every answer can be checked the same way.

Step 6: Create the agent and run it in the Playground

The Review step shows the whole configuration on one screen: name, version, domain, model, output format, workspaces and description. Check it, then create the agent.

Open Playground, select it_ticket_triage and send a real-sounding ticket:

My laptop was stolen from my car last night. It has the Q3 finance forecast on it. What should I do?

The agent answers in a few seconds with the category Security, priority P1, a one-line reason and a reply that tells the requester to contact the service desk immediately. The classification is exactly what the rules asked for.

Now read the reply as a reviewer would. In the run shown here, the reply gives the service desk phone number as 0800 123 4567. That number appears nowhere in the instructions: the model invented it. In another run of the same test, the reply left a placeholder, “[insert number]”, and promised a remote wipe that no one had authorised. These are exactly the defects a first run exists to catch, and they are where the next lessons pick up: testing an agent against realistic cases, fixing its instructions and comparing versions side by side before anyone relies on it.