AI consulting for the systems your business already runs on. With a human holding the keys.
Codium is an artificial intelligence consultant and software company in one. We work out where AI will pay off in your business, then build it into the software you already run: reading documents, answering questions from your own data, and agents that get real work done across email, tasks, files and chat. Every action is permitted, logged with its reasoning, and approved by a person where it matters.
- Built into your softwareAI features inside the systems you already use, not another tool on the side.
- Human in the loopAgents act within limits you set and ask a person before anything consequential.
- Private by designYour data isn't used to train models and doesn't leave an environment you control.
- Every action loggedWhat was done, why, by whom and on what data, so there's always an audit trail.
Three signs you need an AI consultant, not another demo
Most people who call us about AI arrive at one of these three moments.
You've been told to "do something with AI"
The board or the owner wants an AI plan by next quarter. Every vendor says yes. Nobody has worked out which tasks in your business would actually change, or what it would cost to find out.
ChatGPT impressed you, then stalled
Staff use it for emails and summaries, but it can't see your jobs, your pricing or your customers, so the real work is still done by hand. The gap is integration, and that's what we do.
Your data can't go to a public AI service
Client records, health information, contracts or anything under an NDA. You want the benefit without the data leaving an environment you control. Private hosting in Australia and on-premise AI are both options.
AI consulting that ends in working software
Most AI consulting companies hand over a strategy deck and leave the build to someone else. Most developers build whatever they're asked to. We do both halves, so the advice has to survive contact with your real data, your real systems and your real staff.
AI opportunity and readiness assessment
We sit with the people doing the work and find the tasks that are mostly reading, typing, sorting and chasing. For each one: the hours it takes, the data it needs, and what a wrong answer would cost. That gives you the list worth automating, ranked, and the data gaps to fix first.
Options and model advice
API models, private hosting in Australia, a server in your own office, or no AI at all. We compare them on accuracy, monthly cost, privacy and how hard each would be to swap later, then recommend one. If a rule-based script does the job, we say so.
Guardrails and data policy
Before anything is built, we write down what the AI may see, what it may do, who approves what and how it's logged. That document becomes the design, and it's what your board, your auditors and your clients will ask to see.
The consulting phase is Discovery: standalone, fixed price, two to four weeks. You keep the report and the roadmap whether or not you build with us, and you can take it to your own IT team or another supplier. How the process runs
A chatbot answers. An agent gets the job done.
Much of the talk about AI in business software has been about chatbots. The bigger shift is agents: given the right tools and the right limits, an agent can review history, pull information from several systems, prepare a quote, create a task or update a record, and hand the decision to a person when it matters.
"A customer wants a quote"
-
×Replies with a generic answer and a link to the contact page
-
×Can't see your pricing, stock or past jobs
-
×Someone still has to do the quote by hand
-
×Nobody knows what it told the customer
Document extraction
Bills, forms, contracts and emails read automatically: supplier, dates, line items and amounts pulled out, checked and filed where they belong.
Search and summaries
Ask questions of your own business knowledge: job history, documents, policies and past decisions, summarised with sources rather than guessed.
Classification and routing
Incoming requests, tickets, leads and messages sorted, prioritised and sent to the right person or workflow without someone triaging an inbox.
Agents with tools
Multi-step work across your systems: prepare the quote, raise the task, update the record, draft the reply. Within limits, and with a person approving what matters.
Where AI earns its place in a business system
The AI features that pay off are usually the dull ones, removing hours of reading, typing and chasing every week. This is what we build most often.
01Bills and invoices, read and matched
AI-powered OCR reads supplier bills, pre-fills the details, matches them to purchase orders and routes anything unusual to a person. We built exactly this for the JACK construction platform, with approvals flowing through to Xero.
See JACK's AI bill processing
02Customer assistants
Assistants that answer from your real product, order and account data, hand off to a human cleanly, and never promise something the system can't do.
03Internal knowledge search
Staff ask plain-language questions across documents, job notes and past tickets, and get answers with the source attached.
04Inbox and task agents
An agent that reads the shared inbox, creates the tasks, chases the missing information and drafts the replies for someone to approve.
05AI reporting
Ask questions of the data already in your system and get answers, tables, CSVs or PDFs back, on demand or on a schedule.
Ask your data
06Content and drafting
Quotes, proposals, job reports and customer updates drafted from the data already in your system, in your tone, ready for a person to check and send.
Got a job that's mostly reading, typing and chasing?
That's usually where an AI feature pays for itself first. Tell us what the task is and we'll tell you honestly whether AI is the right tool for it, and what the guardrails would look like.
Run it past usUseful enough to act. Controlled enough to trust.
Large language models are non-deterministic. That's fine for drafting a reply; it's not fine for paying a bill. So every AI feature we build is designed around what it may see, what it may do, who has to approve it and how you'd find out afterwards.
Human in the loop
Agents work up to a line you draw: amounts, record types, customers or actions. Past that line they stop and ask, and the decision stays with your people.
Private by design
Your data is not used to train anyone's models. It's processed through API or self-hosted models configured for that, and it doesn't leave an environment you control.
Private AI hosting in Australia On-premise AI and private LLM serversPermissions and data access
An agent gets its own identity with explicit scopes: which systems, which records, read or write. It can never see more than the person it's working for, and usually sees less.
Audit trail with reasoning
Every action is logged with the data it used, the reasoning it gave and who approved it. You can answer "why did it do that?" for any step, months later.
An agent is only as useful as what it can reach
An agent needs to read the information a task depends on and, where permitted, act in the systems your team already works in. We're API specialists, so connecting them is familiar ground.
Tasks and projects
Files and documents
Chat
Accounting and CRM
Your business systems
Each connection carries its own permissions, so an agent allowed to read the inbox isn't automatically allowed to send from it. Software integration
Prove it on a small task, then widen the scope
We start with a job that costs your team hours, prove AI can do it reliably with the right guardrails, and expand from there. The choice of model comes later.
Consult: find the task
First, we work out where AI will actually save time, and where it shouldn't be used at all.
In Discovery we look at the repetitive reading, typing, sorting and chasing in your workflows, the systems that hold the data, and what a wrong answer would cost. That tells us what to automate fully, what needs an approval step, and what stays with people. The Discovery report is the consulting deliverable: the ranked task list, the options we compared, the guardrails we recommend and an estimate for the build.
Prototype with real data
Next, a working prototype on your real documents and records, with the guardrails designed in from the start.
AI-assisted development lets us build and test this quickly. We measure accuracy on your data, tune prompts and models, decide what the agent may read and write, where it must ask a person, and what gets logged. You see it working before committing to the build.
Deploy, watch, widen
Then it goes live inside your system, with monitoring, review and a human approving what matters.
The feature is built into your application on our usual stack, hosted in an environment you control, with its audit log and approval queue in the interface your team already uses. As confidence grows, the limits are loosened and the scope widened, one task at a time.
Models change. Your system shouldn't have to.
We build against the provider's API through an abstraction, so when a better or cheaper model arrives, swapping it in is a configuration change, not a rebuild.
What does an AI consultant actually do?
A good one finds the tasks where AI will save real hours, tells you which model and hosting suit your data, writes down the guardrails, and gives you a costed plan. Ours do that in Discovery, then stay to build it, so the advice is tested against real documents and real systems rather than left in a slide deck.
Do we need an AI consultant or a developer?
Both, ideally the same team. Consulting without building produces a strategy nobody can price. Building without consulting produces a chatbot nobody uses. We scope and advise first and build second, and if the honest advice is that you don't need AI yet, you'll hear it in week one.
Can you advise without building?
Yes. Discovery is a standalone, fixed-price phase. You get the opportunity assessment, the options comparison, the guardrails document and a roadmap with an estimate, all yours to keep. Take it to your own IT team or another supplier if you prefer.
What does AI consulting cost?
Discovery is fixed price for two to four weeks, and the figure is agreed before we start. The build that follows is estimated in the report. Running costs after go-live are mostly model usage and hosting, and we set them out in the same document. Our guide to AI consulting costs in Australia has the detail.
Is our data used to train AI models?
No. We use model providers through API agreements that exclude training on your data, or self-hosted models where that's the better fit, and everything runs inside an environment you control. What the model sees is limited to what the task needs, and nothing is sent that shouldn't be.
Can an agent do something we didn't want it to?
Not outside the limits you set. Each agent has explicit permissions for what it can read and write, hard limits such as amounts or record types, and approval steps where a person must say yes. Anything outside that, it can't do, and everything it does do is logged with its reasoning.
What about hallucinations and mistakes?
Large language models are non-deterministic, so we plan for mistakes. Answers cite their sources, outputs are validated against your data and rules, low-confidence results are routed to a person, and we measure accuracy on your real records before anything goes live.
Which models and providers do you use?
Whichever suits the task, cost and privacy requirements, from the major API providers to open models you can host yourself. We build against an abstraction, so switching models later is a configuration change. Laravel, which we use extensively, now has an AI SDK for exactly this.
Can you add AI to the system we already have?
Usually, yes. AI features bolt onto existing applications through their APIs and databases, which is how we built AI bill processing into JACK with a two-way Xero integration. If we built your system, it's straightforward. If someone else did, we start with a review.
How do we know what it did and why?
Every action is written to an audit log: the input it read, the tools it called, the reasoning it gave, the output, and who approved it. The log lives in your system, so you can review it, report on it and answer questions about any decision months later.
Got a task you'd trust an agent with, if it asked first?
Tell us what it is. You'll be talking to an AI consulting company that will also tell you when not to use AI: one of our digital strategists will work through whether it fits, what the guardrails look like and where to start.
Not sure where to start? Discovery is a standalone, fixed-price phase: two to four weeks working through your workflows and settling the scope, timeline and cost. You keep everything we produce and there's no obligation to build with us. If off-the-shelf software is the better answer, we'll say so.