AI agents that complete real work, with approvals and an audit trail
An agent is not a chatbot with ambition. It is a system that takes actions in your real accounts, which is exactly why the approval gates and the audit log matter more than the model behind it.
What is AI agents?
AI Agents are software workers that plan a task, use tools such as your CRM, inbox, files and databases, and carry out multi step work like triage, research, drafting and data entry. They suit Australian organisations with high volume back office processes that are too variable for fixed rules but still need approvals, logging and human oversight.
Get a fixed written quote- Typical timeline
- 8 to 18 weeks
- What drives cost
- How complex the process is, how many systems the agent has to work across, and the level of assurance and human review required.
- Best for
- Variable, multi step back office work with a clear definition of done
- You own
- The code, the prompts, the logs and every credential
- Built with
- Tool use, approval gates, audit logs, evaluation harnesses
Your handover
What separates an agent from a chatbot or a script
A chatbot answers. A script follows a path you wrote in advance. An agent is given a goal, decides which tools to call in which order, reads the results and adjusts, then either finishes or asks for help. That loop is what makes agents useful for work that arrives in slightly different shapes each time: a supplier email that might be an invoice, a query or a complaint, or a tender document that has to be read, summarised and filed against the right opportunity.
- 01Process map with the exceptions documented
- 02Agent with a scoped, least privilege tool set
- 03Approval gates and configurable spend and volume caps
- 04Scoring harness built from real historical cases
- 05Full audit log of inputs, reasoning and actions taken
- 06Cost per completed task instrumented from day one
- Runbook covering failures, stop switch and rollback
- Staged rollout plan with review thresholds
- Handover training for the team supervising it
The loop is also what makes them risky
The loop is also what makes them risky. A script that fails does the same wrong thing every time and is easy to find. An agent that fails does something novel, plausible and occasionally expensive, and it may do it inside a system that sends emails to your customers. So we constrain the tool set to the narrowest possible surface, make every action reversible where we can, and put a human approval in front of anything that touches money, a customer or a permanent record.
Most agent projects are automation projects wearing a costume
Enquiries arrive asking for an agent and describe a process with fixed inputs, fixed steps and one correct outcome. Copy the order details from an email into the ERP. Move a file when a form is submitted. Alert the account manager if a job has not been invoiced within a week. None of that needs a model. It needs a reliable integration, and it will run for years at a predictable cost without hallucinating anything.
More on most agent projects are automation projects wearing a costume
We say so during scoping, and it usually saves the budget rather than losing us work. A large portion of what businesses want is better handled by n8n or Power Automate with clean logic and error handling, with a model called only for the one genuinely fuzzy step, such as classifying an email or extracting fields from an unstructured attachment. That hybrid pattern is the one we recommend most often, because it puts intelligence exactly where it earns its cost and determinism everywhere else.
How the engagement runs
How we build an agent you can trust with production access
We build against a copy of your systems first, with a task set drawn from real historical work where the correct outcome is already known. That gives a score to improve rather than an impression to argue about, and it makes it obvious when a change to the prompt or the tool descriptions has quietly broken something that used to work.
- 01Process mappingThe current steps, the exceptions, the decisions and who is accountable for each
- 02Split the workWhich steps are deterministic, which genuinely need judgement, which should stay with people
- 03Task setReal historical cases with known correct outcomes, used as the scoring harness
- 04Tool designNarrow, well described tools with least privilege credentials and hard limits on what they can touch
- 05GuardrailsApproval gates, spend and volume caps, blocked actions and a stop switch anyone can hit
- 06Shadow runThe agent proposes, a person approves every action, disagreements are recorded
- 07Supervised liveAutonomous on low risk cases, with sampled review and full audit logging
- 08WidenThresholds relaxed only where the measured error rate justifies it
Two decisions on your side that keep the project moving
Rollout is staged deliberately. The agent proposes and a human approves everything. Then it acts alone on low value cases while a person reviews after the fact. Then approval thresholds move as the evidence supports it. We have never regretted moving slowly through those stages, and we have been called in to clean up after teams that skipped them.
Reliability compounds, and that is the whole problem
Suppose each step of an agent is right ninety five percent of the time, which is respectable. Chain ten steps together with no checks and the whole task completes correctly around six times in ten. That arithmetic is why demos are dazzling and production is disappointing, and why the answer is not a better model but a better decomposition: fewer model decisions, more deterministic steps, validation after each stage, and a definition of done the agent can be checked against.
At pilot volumes nobody notices
Cost behaves the same way. An agent that reasons, retries and reflects can make many model calls to finish one task, and each one is billed by the token. At pilot volumes nobody notices. At full volume the sensible measure is cost per completed task compared against the staff minutes it replaces, and we instrument that from the first week so the comparison is real. Some tasks fail that test outright, and it is far cheaper to discover it during a pilot than after a rollout.
Approvals, audit logs and who is accountable
Every action an agent takes is logged with the inputs it saw, the reasoning trace, the tool it called and the result, stored somewhere your team can search without asking us. When a customer rings in three months asking why they received a particular email, you need to answer in minutes. Accountability does not transfer to the software either. If an agent sends a misleading quote, that is your business making a misleading representation, and the Australian Consumer Law does not have an exemption for automation.
Two other risks deserve naming
Two other risks deserve naming. Agents that read untrusted content can be manipulated by instructions hidden inside that content, so anything the agent reads from outside your organisation is treated as data, never as commands. And agents inherit whatever permissions you give them, so credentials are scoped to the minimum and rotated. Where the audit log or the documents contain personal information, we keep that store in Australian infrastructure and check what crosses a border for model inference, because APP 8 makes that disclosure your responsibility regardless of who operates the model.
When an AI agent is the wrong fit
If the process runs fewer than a handful of times a week, the build and supervision cost will not come back. If the steps never vary, use automation and keep the reliability. If nobody can articulate what a correct outcome looks like, an agent will produce activity that feels like progress and cannot be evaluated, which is the most expensive failure mode of all because it takes a year to notice.
The rest of the answer
Agents also fail badly where the consequence of a mistake is severe and hard to reverse: clinical decisions, credit assessment, anything with a statutory deadline. In those settings the right build is usually a drafting and triage assistant that prepares work for a qualified person, which captures most of the time saving and none of the tail risk. If your real bottleneck is reading paperwork rather than deciding what to do with it, Document AI is the narrower and more reliable tool. Professional firms in particular often find that the sequencing question matters more than the technology choice.
How we scope it
Four ways to scope your AI Agents project
We do not publish package prices, because the same brief can be a short build or a long one. These are the shapes the work usually takes. Tell us which one sounds like you and you will get a fixed written quote that spells out exactly what it covers.
Proof of value
One use case, evaluated honestly before it goes near a customer
Fixed written quote, agreed before work starts
- Process map with the exceptions documented
- Agent with a scoped, least privilege tool set
- Approval gates and configurable spend and volume caps
Production build
In production, with a human approval step and an evaluation set
Fixed written quote, agreed before work starts
- Everything in Proof of value
- Scoring harness built from real historical cases
- Full audit log of inputs, reasoning and actions taken
- Cost per completed task instrumented from day one
Embedded platform
Built into the product rather than bolted onto it
Fixed written quote, agreed before work starts
- Everything in Production build
- Runbook covering failures, stop switch and rollback
- Staged rollout plan with review thresholds
- Handover training for the team supervising it
AI Agents Model care
Monitoring, evaluation and retraining as the inputs drift
Rolling monthly, quoted in writing
- Evaluation set rerun as the model and the inputs change
- Cost and quality reported monthly, not assumed
- Prompt, tool and guardrail changes as the work shifts
- Rolling, cancel with 30 days notice
These are shapes, not menus. Most quotes end up somewhere between two of them, and we will say so when the honest answer is the smallest one. Describe the problem and we will tell you which it is.
Questions buyers usually ask
Frequently asked questions
How long does an AI agent project take?
Typically 8 to 18 weeks. Mapping the process and building the scoring harness usually take longer than building the agent, because they require time from the people who currently do the work. The shadow running period alone is generally two to four weeks, and we do not recommend compressing it. Narrow single task agents can land at the shorter end.
What does an AI agent cost to run, not just to build?
Two components. The build is quoted in writing after a scoping phase, driven by process complexity, the number of systems involved and the level of assurance required. Running cost is model usage billed per token plus the staff time spent supervising it. We measure cost per completed task during the pilot so you can compare it against the manual cost before committing to a wider rollout.
Who owns the agent and the data it touches?
You own the repository, the prompts, the tool definitions, the evaluation cases and the logs. Model provider and infrastructure accounts are opened in your name. Access sits with you, and we are a collaborator you can cut off at any time. Nothing about the design depends on us remaining involved, which is deliberate and also the only honest way to sell this work.
What happens when the agent gets something wrong?
The log tells you exactly what it saw and did, the case goes into the scoring harness so the same failure is caught in future, and if the error class is serious the approval threshold for that action tightens again. Every agent we build has a stop switch that halts it without a deployment. Expect errors and design for them rather than assuming they will not occur.
Can an agent work inside our existing systems?
Usually yes, through APIs. Where a system has no API we look at supported integration layers before considering anything that drives a user interface, since those break with every vendor update and are miserable to maintain. If your systems cannot be integrated cleanly, that is a separate piece of API work and we scope it as one rather than hiding it in the agent build.
Will this replace roles in our team?
Honestly, it changes them more often than it removes them. The usual outcome is that a person stops doing the repetitive portion and starts supervising exceptions, which needs judgement and more product knowledge, not less. If your business case depends on removing headcount, say so during scoping so we can be blunt about whether the measured error rate will ever support running without a reviewer.
What does AI agent development in Australia actually involve?
Agentic workflow implementation is mostly plumbing and judgement rather than model choice. We map the process, define the narrow tool set the agent may touch, write the approval gates that stop it acting alone on anything expensive, then run it in shadow mode against real work until the error rate is known. LLM tool use and approvals get designed together, because an agent that can call a tool without a human check is only safe where the action is reversible.
Related services
Describe the process before the technology
Tell us the steps, the exceptions and how often it runs. We reply within one business day, and if automation is the better answer than an agent we will say that before quoting anything.