Automation builds
From $3,500 AUDOne process, rebuilt to run without you. Monitoring and documentation included.
For Australian teams preparing research, internal answers and drafts
AI agent development that turns research into work you can use.
We build agents that gather permitted information, check what they found and prepare something your team can use. You can inspect the sources, see what remains uncertain and keep approval over consequential actions.
Research. Reason. Review.
A closer look +
Who it is for
When finding the evidence takes longer than making the decision.
For Australian teams that repeatedly turn scattered documents into a research brief, a proposal draft or an internal answer. Your knowledge may sit in Google Drive, Microsoft 365 or a CRM. We check access and exact tool operations before choosing a model or agent framework.
What changes for you
01
A research packet with source references and unresolved questions
02
A draft prepared from the evidence your team permits
03
A visible limit on tools, spending and attempts
04
Human approval before an external commitment
Example system · Simulated business data
Inspect what sits behind the answer.
A fictional team asks whether its internal onboarding pack covers a remote starter. Open the evidence, tool boundary and proposed output.
QUESTION · Q-017
“What is missing before this starter can work remotely?”
Allowed scope is the internal checklist and equipment register. The agent cannot order equipment or email the starter.
01 · Read the permitted evidence
Checklist C-17 requires an account owner, an issued device and access approval. Register E-04 lists a laptop but has no handover receipt. These are fictional records, authored for this example.
- Permitted tool
- Read selected checklist and asset fields
- Excluded action
- Purchase, email or account provisioning
02 · Separate facts from missing evidence
Ordinary checks find the absent receipt and access approval. AI turns those findings into a concise draft. No tool call may mark an approval as complete.
- Supported finding
- Device assigned, handover unconfirmed
- Unresolved
- Access approval and receipt
03 · Review the useful draft
“The laptop is assigned in E-04. Confirm its handover receipt and record access approval against C-17 before marking this starter ready.”
- Output state
- Draft for the operations owner
- Next action
- A person checks the records and approves the next step
This is an authored inspection example. Opening a panel does not run an agent, search documents or execute a tool.
Give the agent a job and a smaller set of permissions
We define a narrow objective, approved sources and the tools it may call. A research agent may search an authorised document collection and read selected CRM fields. That does not give it permission to send email, change bank details or buy something.
The surrounding workflow enforces these boundaries. Retrieved documents and websites are treated as evidence, not instructions that can expand access. Secrets stay outside prompts where possible. Tool arguments are checked against the expected record, allowed fields and current user authority.
- Allowed sources and account permissions
- Read tools separated from write tools
- A maximum number of attempts and a usage budget
- Approval bound to the exact proposed action
Test the uncomfortable questions before release
We assemble representative questions with expected evidence and acceptable outputs. The evaluation includes missing documents, contradictory versions, irrelevant search results and text that tries to instruct the agent. We inspect whether it stays within its tool permissions and admits uncertainty.
Ordinary rules do the exact work. Required fields, identifiers, arithmetic and approval status do not need to be guessed by a language model. AI earns its place in tasks such as extracting meaning, organising evidence and drafting a human-readable answer.
You receive the prompt and tool configuration, evaluation examples, known limits and recovery notes. A future model change should be tested against those cases before it replaces the working version.
The example shows our design approach using simulated records. It is not a client case study or a measured performance result.
What it costs
Price the scope and the operation.
Cost depends on the knowledge sources, permitted tools, evaluation cases, approval interface and consequences of a wrong answer. Model usage varies with context and tool calls. We set a budget and measure representative cases rather than promise a fixed cost for every question.
We put the build scope, fee and payback estimate in writing using your baseline. You own the accounts, workflows and documentation. Software, hosting and AI usage are separate operating costs.
Typical projects
$3,500 to $12,000 AUDSeveral connected processes across two or more systems, where the data needs cleaning first.
Complex AI agent systems
$12,000 to $25,000 AUDAgents that read unstructured information, decide and write back, with human review and a full audit trail.
Ongoing care plans
$750 to $2,500 AUD per monthPriced on how much you run and what your business loses the day one of them stops.
When something breaks
A clear owner. A safe next step.
If retrieval returns nothing or sources disagree, the agent should say what is missing and stop the affected draft. Tool timeouts, invalid output and exhausted budgets go to a named reviewer. Repeating the same prompt indefinitely is not a recovery plan.
We account for a failed run, an empty result and a run that never starts. Your written care scope names alert ownership, recovery responsibilities and change boundaries. Support commitments belong in that scope.
Care is month to month and can be cancelled without penalty. You keep the system and its documentation, with no lock-in.
See ongoing supportBefore you decide
Your questions, answered
Is this a chatbot or an AI agent?
A chatbot is an interface for a conversation. An agent can choose from permitted tools to work towards a defined objective. Either can be useful. We start with the output you need and use the simpler approach when a fixed workflow or search interface is enough.
Can an agent send emails or change business records?
Only if that exact action is in the agreed scope and the account permissions support it. Consequential writes can be held for a person to approve the recipient, content and record revision. A broad instruction to be helpful is not permission to make commitments.
Can you guarantee it never makes a mistake?
No. We limit the consequences through source checks, structured validation, evaluations and human review. An agent should expose uncertainty and return an incomplete result when the evidence does not support a confident answer.
Will our data stay inside our business?
That depends on the complete data path. Documents may pass to a search service, model provider, platform logs and support tools. We document those destinations and their settings before choosing the architecture. Self-hosting one part does not remove the others.
Who owns the accounts and the finished system?
You do. We build in accounts you control and hand over the workflows, access map, documentation and available exports. Another capable operator can take over. Platform licences and subscriptions remain subject to their own terms.
How do you price the work and ongoing support?
Builds start from $3,500 AUD. Connected builds are $3,500 to $12,000 AUD and larger systems are $12,000 to $25,000 AUD. Care is separate at $750 to $2,500 AUD per month. We confirm the scope and fee in writing. Software, hosting and AI usage sit outside the build fee. Care is month to month and can be cancelled without penalty.
A useful first conversation
Bring the handover you need to make reliable
Show us a representative record, the tools involved and what the next person needs. We will discuss a useful first scope and the decisions it depends on.
Book a call