Custom AI Agents for business processes that matter

Picking a process worth automating with an agent

The best candidates combine three traits: the work follows a recognisable pattern, it requires reading or reasoning that simple automation cannot do, and it consumes meaningful skilled time. Examples include triaging inbound requests, preparing first drafts of standard documents, reconciling information across systems and monitoring for exceptions.

Poor candidates are processes with unclear rules, rare occurrences or high consequences for any mistake. For those, an assistant that supports a person is usually safer than an agent that acts alone.

Designing guardrails before writing prompts

An agent's capabilities should be defined by what it is allowed to do, not by what the model can do. List every tool the agent may use and every action it may take, such as reading a record, drafting a reply or creating a task, and grant only those permissions.

Classify actions by risk. Low-risk actions can be automatic. Medium-risk actions can require a quick human approval. High-risk or irreversible actions, such as payments, deletions or legal commitments, should always require approval or remain manual. Every action should be logged with the reasoning that led to it.

Testing an agent properly

Build a test set from real historical cases, including messy, unusual and adversarial examples. Run the agent against the set before launch and after every significant change, measuring accuracy, completion rate and how often it correctly escalates.

Launch in stages: shadow mode, where the agent proposes actions that a person reviews, then partial automation for the clearest cases, then wider automation as confidence grows. This approach catches problems while they are cheap to fix.

Governance, privacy and accountability

Assign a business owner for every agent. They decide what it may do, approve changes and review performance. Technical ownership sits with the team that maintains it, which in our model is Apex.

Map the personal information the agent touches and apply the Australian Privacy Principles. Keep data within the systems and regions you approve, and make sure logs are retained long enough to investigate any issue. Clear governance is what allows an organisation to expand agent use with confidence.

Build, buy or partner

Off-the-shelf agents suit common tasks such as meeting notes or simple support. Building entirely in-house gives maximum control but requires engineering capacity and ongoing maintenance. A partner who designs, builds and manages agents sits in between, giving you a tailored system without hiring a team.

Whichever path you choose, insist on documentation, model flexibility and monitoring. Models will change. Your process knowledge, test sets and guardrails are the durable assets.

Measuring return on a custom agent

Define success before building: hours saved per week, turnaround time, error rate, throughput or revenue impact. Measure the baseline manually for a couple of weeks so the comparison is honest.

After launch, track accuracy, escalation rate, failures caught by monitoring and model usage costs alongside the business metrics. A good agent pays for itself through time saved and faster outcomes, and the data tells you when it is ready to take on more.

An illustrative agent: triaging inbound requests

Imagine a firm receiving hundreds of mixed requests by email each week: new enquiries, document submissions, invoice queries and appointment changes. An agent reads each email, classifies it, extracts key details, checks the relevant system and either completes a low-risk action, such as filing a document, or prepares a draft response and task for a person.

Staff start each day with a sorted queue rather than an overflowing inbox. Every action the agent took is visible with its reasoning, and anything uncertain is clearly marked for review.

When a custom agent is the wrong tool

If a process follows fixed rules with no reading or judgement involved, standard workflow automation is cheaper, faster and more predictable than an AI agent. If the process is rare or poorly defined, documenting it properly should come first. We will recommend the simplest solution that achieves the outcome.

Key terms explained

AI agent

Software that uses a language model to decide steps towards a goal and act through approved tools.

Tool use

An agent's ability to call systems such as a CRM, database or email platform to read or write information.

Guardrails

Limits on what an agent can do, including permissions, approval steps and rules about sensitive actions.

Human in the loop

A design where a person reviews or approves certain agent actions before they take effect.

Shadow mode

A testing stage where the agent proposes actions but a person performs them, allowing safe comparison.

Evaluation set

A collection of real examples used to measure an agent's accuracy before and after changes.