AI automation for the work your team repeats every week
ComplxAI builds AI automation for Australian businesses: agentic workflows that take on the intake, document handling, follow-ups, reconciliations and reporting your team does by hand. Fixed rules where the rules are known, an agent where the input is messy, human sign-off where an action has consequences.
- Human approval built in, thresholds set by you
- IP yours from day one of any paid stage
- Every stage priced in writing, exit at any boundary
Which repeatable processes are worth automating
Most operations teams know where the hours go: a form re-keyed, an invoice matched to a job by hand. Neither needs a new product; the process needs to run itself, with a reviewer where a decision carries weight.
We map the process, split fixed logic from agent judgement, and ship it inside the systems you already run. Where the job calls for a custom agent with tool use, memory or several agents coordinating, AI agent development covers the architecture. Still choosing a process? Discovery scores candidates by impact, effort and risk; AI strategy turns the shortlist into a plan.
A good fit if…
- One team repeats the same multi-step task every week
- Messy inputs (email, PDFs, forms), well-defined outcome
- Errors surface late; headcount grows to keep up with admin
- The data lives in a CRM, ledger or database with an API
Six processes an agentic workflow takes on
Repeatable, high-volume, done by hand.
Intake and triage
Email, form and portal enquiries classified, checked against your records, given a draft reply, routed to a named owner when unusual.
Document processing and data entry
Invoices, referrals and statements read, validated against the source of truth and written to the ledger or CRM.
Follow-ups and scheduling
Quotes, renewals, missing documents and unpaid invoices chased on your cadence; replies read, bookings placed in the calendar.
Reconciliations
Payments to invoices, claims to plans, statements to the ledger; clean matches clear alone, mismatches reach a reviewer with evidence.
Reporting
Weekly numbers pulled from your tools, checked for gaps, assembled and sent, each figure traceable to its query.
Quoting and estimates
Requests parsed, the right rate card applied, a draft quote built in your template and held for approval.
Four tests before a workflow is automated
Applied in Discovery, before anything is built.
High volume, low variance
The same shape of work, many times a week, so a small saving per item becomes hours.
A clear source of truth
A system that says what is correct: the ledger, the CRM, the price list. Validation needs one.
Actions that can be held
Reading, matching and drafting can run alone; sending, paying and committing wait for sign-off. That line sets the autonomy level.
Quick for your team to check
An output confirmed in seconds keeps the supervised rollout short and the error rate measurable.
Deterministic rails, agent steps and a human gate
Not one large model given access to everything: a pipeline with three kinds of step, every action logged.
- Triggerswhere work arrivesInbox and formsUploaded documentsCRM and ledger events
- Deterministic railsfixed logic in codeValidationRouting rulesSource-of-truth lookupsIdempotent writes
- Agent stepsjudgement on messy inputClassifyExtract fieldsMatch recordsDraft the reply or quote
- Human gatewhere consequences sitApproval queueConfidence thresholdsNamed exception owners
- Systems of recordwhere results landXeroMYOBHubSpotSalesforceEmail and calendarInternal databases
- Recordwhat happened and whyAction log per runEscalation reasonsError metrics
Rails are ordinary code, agent steps do the reading and judgement, the gate sits where the consequences are. Tool design, memory, multi-agent coordination and the runtime belong to AI agent development; connections to Xero, HubSpot, Salesforce and your databases to our AI integration services.
From shadow run to earned autonomy
Six steps, run inside the engagement described in how we work.
- Map the processWalk the task as done today: inputs, systems, decisions, exceptions.
- Split rules from judgementFixed logic, agent steps, and what needs a person's approval.
- Shadow runLive inputs, no writes; outputs compared with what your team did.
- SupervisedIt acts, your reviewer approves. Every override feeds the evaluation set.
- Earned autonomySteps that clear the bar you set run alone; the rest keep their approval.
- Operate and measureReporting on hours returned and error rate, documentation, option to take it in-house.
Controls you set, results you can measure
An automated process is only useful if its owners can see what it did and stop it in a second. Guardrails are configuration, not promises.
Measurement is set up before go-live, so the result is a before-and-after comparison; the evaluation harness and tracing come from our AI engineering practice.
Guardrails
- Permissions scoped per step, never a blanket key
- Spend and rate limits per run, plus a pause switch
- A log of every action with the evidence it used
What we measure
- Hours returned to the team each week
- Cycle time from arrival to done
- Error rate against the source of truth
- Escalation and override rate per step
AI automation we have shipped
90% of operations automated
For DisabilityAssessments we automated matching, alerting, invoice processing and reconciliations; the client reports 90% of operations automated on 10% of the human resource.
Manual workload down 75%
For Luxpip we built agentic solutions for internal operations alongside its web and mobile apps and B2B API; the client reports a 75% drop in manual workload and a 40% scale-up.
2× revenue in six months
For Modedu we built agentic workflows across operations on its custom SaaS platform; the client reports 2× revenue within six months, and ARR growth from $400k to $1M.
Agentic through core operations
For Billy we built a custom platform with agentic workflows through its core operations, delivered by the same small senior team. More on what we have built.
On either side of an automated process
AI agent development
Tool use, memory, multi-agent coordination.
Custom AI agentsAI integration
Connections into your CRM, ledger, inbox, databases.
Integrating AI with existing systemsAI development
When the process needs an application around it.
Custom AI developmentCustom software
The portals and platforms workflows run inside.
Custom software developmentAI strategy
A board-ready plan with build, buy or defer calls.
AI strategy consultingAI engineering
Evaluation, tracing and the runtime underneath.
Production AI engineeringQuestions operations leaders ask
Bring the process and whoever runs it.
Book a free intro callHow is this different from rules-based or no-code automation?
Rules hold while the input is tidy and break on a scanned PDF or a near-match record. An agentic workflow keeps the rules where they hold and adds agent steps for reading, matching and drafting, in code you own. We compare the two approaches in AI agents vs traditional automation.
What happens when it gets something wrong?
The log shows what it read, decided and why. A reviewer corrects the item, the case joins the evaluation set and the next release is tested against it. Low-confidence items are held for review first.
Do we have to replace our current systems?
Usually not. It sits on top of the CRM, ledger, inbox and databases you already use and writes through their APIs. If a system has no usable API, we say so in Discovery.
How is AI automation priced?
Two free conversations, then Discovery from a fixed fee, Strategy priced upfront and Implementation in 12-week blocks, with an exit at each boundary. The AI project cost calculator gives a rough range.
Tell us which process your team repeats most
Bring the task, the systems it touches and who checks it today. Two opening conversations cost nothing; nothing later starts without a written price.