Your pricing lives in a system only two people understand.
We build a connector to it during setup, with named actions only, so the agent reads your rates instead of guessing them.
Custom AI development
Off-the-shelf AI doesn’t know your business. We build the agent that does.
Agents, workflows and models for the work a ready-made tool can’t do: your own rules, a system you built, several steps across several tools.
Built on the systems you already run, tested on your own past cases, and released only when a named person on your side signs it off.
Follow one build, start to finish
Reads run
Changes ask first
Deletes stay off
Where the ready-made tool runs out.
Three signs a tool off the shelf won’t do the job, and what we build instead.
One job means four steps across three systems.
We build the workflow end to end: every hand-off written down, and every exception routed to the person who owns it.
The answer has to be right, not just plausible.
We test on your own past cases, and choose the model approach by the accuracy the job needs.
What we build
Agents, workflows and models, shaped to the job.
A build can use one of these or all four. Which ones it uses is decided in scoping and written into the spec.
Agents and assistants
They look things up, guide the person asking, and carry out the tasks they’ve been approved for, by text, inside a workflow or on the phone.
Workflows across your tools
Decisions and actions joined up across the systems you run, with clear hand-offs and a named person for every exception.
Models and intelligent features
Chosen, adapted or built to suit the task, the data you have and the accuracy you need.
Evaluation and iteration
Realistic scenarios before release, failure cases reviewed with you, and the same tests typically run again after every change.
Which model? The smallest one that does the job.
Custom doesn’t mean training a model from scratch. We typically start on the left and move right only when tests show the step before falls short, because each step costs more to build and to keep accurate.
- An existing model, with your instructionsThe work goes into the instructions, the tools it can call and the tests.Where we start
- Retrieval over your own contentWhen the answers live in your documents, policies or past jobs, the agent looks them up instead of guessing.
- Fine-tuning on your examplesWhen the output has to follow your format or your wording, and there are enough good examples to learn from.
- A purpose-built modelWhen the task is narrow, the data is plentiful and accuracy can be measured. See machine learning solutions.
A worked example
One build, start to finish.
For example, a 40-person freight business that quotes every job by hand from a rates system it built in-house. Two people understand the rates, and quotes wait whenever either is away. Here is how we would build them a quote drafting agent.
- Scope
Write the agent down
A workshop with the people who quote, then a spec with a fixed price.
- Prototype
Draft on real requests
A working version on their own inbox and rates system, drafting only, for the team to pull apart.
- Test
Run sixty past requests
Every result marked, failures fixed with the team, then the whole set run again.
- Launch
Sign off, then switch on
Mel, their operations manager, signs it off. Switching it on is a separate step, and we typically review it with them each month.
Scope
The spec comes before any code.
It names the job, every tool the agent may call and at what level, the limits it works inside, and who approves it. A tool that isn’t on the spec doesn’t exist for the agent.
The rates system had no connector, so we build one during setup, with named actions only. Mel signs the spec, and the price is fixed in writing before building starts.
How we connect a system you built
Build spec · Quote drafting agentExample with sample data
Job
Draft a quote for each request in the shared quotes inbox, priced from the rates system.
Tools
- Read a quote requestMicrosoft 365Runs
- Look up a rateRates system, connector built for youRuns
- Draft the quote replyMicrosoft 365Runs
- Send the quoteMicrosoft 365, approved by MelAsk first
- Change a rateRates systemOff
- Delete an emailMicrosoft 365Off
Limits
- Quotes over AUD $8,000 go to Mel to price
- Lanes not in the rates system go to Mel, with the reason
- Never offers a discount or promises a pickup time
Approver
MelOperations manager
Signed off by Mel
Signed off means the price is fixed and the build can start. Nothing is live yet.
Test
Then it runs on their own past requests.
Sixty real quote requests from the last few months, the awkward ones included. Each result is marked passed, worth fixing, or refused as designed.
A refusal is a good sign: the agent reached a limit in the spec and stopped. The fixes go in, all sixty run again, and Mel sees the results before release.
Test run 3 · 60 past quote requestsExample with sample data, 5 of 60 shown
#072 pallets, Brisbane to ToowoombaRate found, quote drafted for Mel to sendPassed#19Tail-lift needed at deliverySurcharge added from the rates systemPassed#261 pallet to MackayLane not in the rates system, handed to MelPassed#41Dangerous goods named in an attachmentMissed on the first readWorth fixing#53Customer asks for a lower rateChange a rate is Off, handed to MelRefused as designed
57passed2worth fixing1refused as designed
Both fixes go in, then all 60 run again before Mel signs off the release.
Control
Control is written into the build, not added after it.
We set the permissions, approval points and escalation paths around the tasks the agent is allowed to do, and SpeedOz holds it to them on every request.
- A named person approves changesAnything that changes a record or leaves the business waits for the person who owns that work. They can approve it from their phone.
- It hands over instead of guessingWhen a case falls outside its limits, the agent stops and passes it to a named person, with the reason.
- Keys and risky actions start OffA custom agent gets a key only when a named person turns it on. Deleting, exporting everything and changing permissions are refused until the owner says otherwise.
- Approved is not liveSign-off stages the agent. Switching it on is a separate step, and it’s yours.
- Every request is on the recordWho asked, what happened and who approved it, refusals included, on a record nobody can edit. Access comes back for review every ninety days.
Handed to MelQuote drafting agent · 10:42
Waiting for Mel
1 pallet, Maroochydore to Mackay, pickup Thursday
Why it stoppedMackay isn’t a lane in the rates system. The spec sends lanes outside it to you, so no price was drafted.
Example with sample data. The hand-over is on the record with the reason.
Deliverables and price
What you get, what you keep, and what it costs.
Timelines and deliverables are confirmed for your project in the written scope, before any work starts.
What you get
- A signed spec, with the price fixed in writing
- A working prototype your team can test
- The agent in production, on your own systems
- Any connector it needs, built during setup
- Documentation and an agreed support plan
- A monthly read on outcomes, exceptions and cost
What you keep
- The spec, in plain words, dated and signed
- The test scenarios, and the results of every run
- The record of every request, approval and refusal, which you can export at any time
If you stop, the connections switch off and the sign-ins are revoked. What was written down for you stays with you.
What it costs
EvaluationFreeA person reads it and replies within one business day.
Six-week pilotAUD $5,000Optional. One process, one team. Credited in full against setup if you continue.
SetupFrom AUD $5,000A fixed scope, quoted in writing before any work starts. A connector for a system not on our integrations list is scoped and priced before it’s built.
Running itFrom AUD $900 a monthFor the whole organisation, not per seat.
Questions
What people ask before a build.
Anything else, ask in the evaluation. A person answers within one business day.
Yes. If the system has an API, we build a connector to it during setup, with named actions only and looking kept apart from changing. If it runs only on your own network, how SpeedOz reaches it is agreed with your IT team and written into the scope. Either way, the price is fixed in writing before building starts.
Connectors already added to Claude Desktop, Cursor or VS Code can come across. Their keys move off the laptop into the encrypted vault, every action they offer is sorted, and nothing goes live until a named person approves it. A connector that only runs on one laptop stays where it is, and we scope it with you instead.
No. We build it and run it as a managed service. We need time from the people who do the work, for the workshop and for checking the tests, and someone from IT if a system sits only on your own network.
It depends on the job, so the timeline is confirmed for your project in the written scope before work starts. If you would rather prove the approach first, the six-week pilot takes one process and one team, connects two of your systems, and ends with a go or no-go.
Anything that changes a record already waits for a person, so a wrong draft is caught before it leaves. Every request is on the record. Typically, we go through the exceptions with you each month, and a fix runs through the tests again before it is released. If you want something stopped, you can switch it off yourself straight away, without waiting for us.
SpeedOz is hosted in Australia by default, with other regions on request, confirmed in writing in the scope. If we fine-tune a model on your examples, the copy is held in Australia and the training runs in Australia, with no overseas compute. Where a connected system holds data elsewhere, the integrations list says so.
Read next
IntegrationsThe systems we connect today, and how we build a connector for one you made yourselves.Read more
Import an MCP connectorBring the connectors your team already set up in Claude Desktop, Cursor or VS Code.Read more
Staying in controlClosed by default and opened one capability at a time, with a named person on every decision.Read more
Bring us the job your current tools can’t do.
Tell us what you want to improve and what it runs on. A person reads it and replies within one business day, and we’ll help you map a practical next step.