Generative AI consulting in Australia
Large language models are very good at a short list of tasks and confidently wrong about the rest. ComplxAI provides generative AI consulting in Australia that sorts your candidate use cases into those two piles, settles the model and data position for each, and designs the grounding, guardrails and approvals around the output. Then we build it.
- Two free conversations before any paid stage
- Discovery in 2–4 weeks, ending in a scored use-case map
- The IP is yours from day one of any paid stage
LLM consulting, not general AI advice
This page covers one family of technology: large language models and the systems built on them. Our AI consulting in Australia is the wider practice for deciding which kind of AI belongs on the roadmap; custom AI development is the next step once a use case is scoped.
The honest picture is narrower than the marketing. Given the right context, a model saves real hours on the tasks listed here; asked for the ones it should not be trusted with, it answers just as fluently. That is the failure to design against.
Where a language model earns its place
- Drafting letters, reports and proposals from notes
- Summarising documents, meetings and long threads
- Extracting fields from invoices, forms and contracts
- Classifying and routing incoming requests
- Answering questions from a curated knowledgebase
- Writing and explaining code
Where it should not be trusted alone
- Exact arithmetic, or anything a database already knows
- Deterministic rules such as eligibility or pricing; those belong in code
- Any output nobody can verify before acting on it
What our generative AI consultants do
Scoping through to a governed system running in your own cloud account.
Use-case identification and prioritisation
We walk the processes where staff read, write, search or transcribe, then score each candidate on impact, effort and risk, after checking whether a rule, a lookup or a predictive model would do the job.
Model landscape and selection
Proprietary APIs from OpenAI, Anthropic and Google, open-weight models you can host, and AWS Bedrock, deployed in the region you need. For Australian data residency that is the AWS Sydney region, which serves some Bedrock models but not all, so we confirm availability per model and pin the region. Each task gets the model that fits, with the option to swap.
Grounding with retrieval
Answers come from your content, not the model's memory: ingestion and chunking, embeddings in Postgres with pgvector or a managed vector database, hybrid search, permission filters and citations to the source.
Guardrails and governance
Input policies that catch injection and out-of-scope requests, redaction of personal information before text leaves your account where the task calls for it, output checks against a schema, approval steps where a person must sign off, and an audit log.
Evaluation, prompts and versions
A golden set written with your experts, a regression run on every change, and judge models checked against human graders. Prompts are versioned like code: one component of the system, not the deliverable.
Cost management
Small tasks routed to small models, repeated context cached, background work batched, and telemetry that reports spend per feature and outcome, so the invoice is a forecast rather than a surprise.
What sits around the model
Six layers. The model is the easiest one to change.
- Intakewhere requests arriveProduct screensStaff toolsShared inboxDocument dropAPI endpoint
- Pre-model controlspolicy gateIdentity and permissionsInput policyInjection screeningPII redactionPer-feature budgets
- KnowledgegroundingDocument ingestionChunk boundaries and metadataEmbedding index, pgvector or managedKeyword plus vector searchRow-level access filtersCitation to the passageRe-index on change
- Model layergenerationTask-to-model routingBedrock, region confirmed per modelProprietary APIsSelf-hosted open-weightPrompts versioned in gitJSON-schema outputsFallback model
- Post-model controlsreview and actionOutput policySchema validationConfidence thresholdsApproval queueAudit log
- Evidencequality, cost, operationsGolden set from your expertsEval run on every changeJudge model graded against peopleSpend per featureAlarms and dashboards
Built serverless and event-driven on AWS and deployed as code into your account, so a slow model call never blocks the rest and swapping a model is a configuration change. The engineering detail is on our production AI engineering page.
Generative AI vs predictive machine learning
A brief for a language model can turn out to be a forecasting or scoring problem, which is the territory of our machine learning consulting.
| Question | Generative AI (LLMs) | Predictive machine learning |
|---|---|---|
| What it produces | Text, code, structured fields or a conversation | A number, a probability or a category |
| What it needs from you | Documents, examples and a clear rubric | Structured records with known outcomes |
| How it fails | Plausible but false output, drift after prompt or model changes, injection | Drift as conditions change, bias inherited from history |
| Where the cost sits | Per-token inference that grows with use unless routed and cached | Data preparation and training up front; inference is usually cheap and predictable after |
From candidate list to live rollout
LLM implementation follows the five stages in how we work; where the board wants a GenAI strategy in writing first, our AI strategy stage produces it, priced upfront. Inside a 12-week block, the LLM work runs in this order.
- ShortlistCandidates scored on impact, effort and risk; LLM fit confirmed or ruled out.
- Data positionProvider, region, data-use terms, retention and redaction decided and written down.
- Grounded pilotA thin slice on real documents, with the golden set in place before tuning.
- ControlsInput and output policies, thresholds and the approval queue wired in and tested.
- Rollout and changeReleased team by team; reviewers trained on what it does and does not do; corrections fed into the eval set.
- Run and hand overDashboards, runbooks and documentation; take it in-house at any boundary.
Use cases, and who approves the output
Each names who signs off. The first two build on systems described in our case studies; the other two are patterns we can scope and build.
Personalised academic papers
For Modedu, an EdTech company, we built a custom SaaS platform with AI that produces personalised academic papers. For a pattern like this, the approval step is an educator who reviews and releases each paper; their corrections become evaluation examples.
A knowledgebase that answers questions
For DisabilityAssessments, an NDIS allied health provider, we built a central knowledgebase while automating its operations. Retrieval over a knowledgebase like that lets staff ask in plain language and get the source passage with the answer. Approval step: content owners sign off each change to the source; the model reads but never writes.
Extraction from contracts and forms
Fields pulled from incoming documents into a schema, validated by rules in code and posted to the system of record. Approval step: any field below the confidence threshold, and any document that moves money, is checked by a person first.
Drafted replies from a shared inbox
Incoming email classified, matched to knowledgebase content, and a reply drafted in your voice. Approval step: a staff member edits and sends; nothing goes out unread. When the workflow must also act in other systems, that becomes AI agent development.
Designed around Australian privacy obligations
Many language-model tasks touch personal information. Organisations covered by Australia's Privacy Act have obligations under the Australian Privacy Principles about how it is collected, used, disclosed and sent overseas, and passing a customer record to a model API can involve all four. We are engineers, not lawyers, and give no legal advice. What we produce is what your privacy officer or adviser needs to make the call.
Governance deliverables
- Data-flow map per use case: provider, region, terms and data-use setting
- Redaction of personal information before text leaves your account, where the task needs it; retention rules for prompts and outputs
- Named owner, written input and output policy, review cadence
- Audit trail of prompts, retrieved sources, versions and approvals
- Prompt and model changes reviewed like code changes; reviewer corrections fed back into the evaluation set
Consulting from the team that ships it
Advice you can hold us to
The consultants who score the use cases are the engineers who build the pipeline. What we recommend, we build and run.
Straight about the limits
We will say when a rule beats a model and when a use case should wait. Providers are components, not products we resell.
A small senior team that does the work
No layers between the people you meet and the people who build. Remote-first across Australia, onsite where it helps, in Australian business hours.
Evidence, not promises
Modedu reported 2× revenue within six months and ARR from $400k to $1M on a platform we built; DisabilityAssessments reported 90% of operations automated. See what we've built.
Where this work connects
AI consulting
Where AI fits, scored and sequenced.
AI consulting in AustraliaAI strategy
A board-ready plan with build, buy or defer calls.
AI strategy consultingAI development
The scoped use case, built as working software.
Custom AI developmentAI engineering
Retrieval, routing, evaluation and observability in depth.
Production AI engineeringAI agent development
When the model must take actions, under approval.
Custom AI agentsMachine learning
Forecasting and scoring problems.
Machine learning consultingQuestions we hear before an LLM project
Bring the rest to the free intro call.
Book a free intro callIs ChatGPT enough for us?
For one person drafting their own work with no customer data, often yes, and we will say so. It stops being enough when the task repeats across a team, the input includes personal or commercial information, the answer must come from your documents, or someone must show who approved what.
Will our data be used to train someone else's model?
It depends on the product and its terms. Consumer chat products and the business or API products from the same provider often carry different data-use terms, and many business and API offerings either exclude inputs from training by default or offer a setting that controls it. We cannot guarantee any provider's policy, so we read the current terms for the exact product and region, choose and record the setting that meets your requirement, and re-check when the provider or model changes.
How accurate will it be?
Nobody can give an honest number before the work starts; it depends on the task, the grounding material and the checking after the model. Extraction and classification against a clear schema tend to do well; open-ended answers need retrieval with citations and a person in the loop. The golden set turns the question into a measurement.
What does an LLM engagement cost?
The first two conversations are free: a 30-minute intro call and a one-hour pre-discovery. Discovery runs two to four weeks from a fixed fee. Strategy is priced upfront. Implementation is quoted per 12-week block, with an exit at every boundary. For a rough range, try the AI project cost calculator.
Bring the tasks you think a language model could take on
We will tell you which ones it can, what has to sit around it, and what each stage costs. The first two conversations are free.