AI project cost calculator
Describe the system you have in mind and this AI project cost calculator returns an indicative range in AUD for the build, the monthly model and infrastructure bill, and what it takes to keep it running. The ranges are wide on purpose. This is a starting point for a conversation, not a quote.
- Runs in your browser, nothing to sign up for
- Rounded ranges in AUD, never a single figure
- Every ComplxAI stage is priced in writing before it begins
Describe the system, read the range
Eleven controls in, four ranges out. Rough answers are fine, and every figure in the card updates as you change a control. Nothing here is a quote.
Project shape
Data and integrations
Load and scale
Risk, quality and operations
Adjust the controls to see an indicative range for a system of this shape.
Indicative, not a quote, and not ComplxAI rates. The ranges were reviewed and confirmed by ComplxAI in September 2026. They are a model, not measured engagement data, and no ComplxAI price is read from them. All figures in AUD. Implementation is rounded to the nearest $5k and monthly lines to a sensible step. Upkeep is modelled as a share of the build cost a year, shown per month. The written price for your project comes from Discovery.
Turn this into a scoped proposalHow the estimate works
Eleven inputs, one model, no hidden weighting. Each control changes the result in the way described here, and every constant sits in one block at the top of the page script: ranges reviewed and confirmed by ComplxAI in September 2026, a model rather than measured engagement data, and never a ComplxAI rate.
- Project type
- Sets the starting range for the build and the starting complexity score. An internal tool starts low. An enterprise platform starts high because it carries orchestration, permissions and several interfaces before any model is called.
- Expected users
- Nudges the complexity score and the infrastructure tier. More people means more access control, more edge cases and more support load, but it does not change the build range on its own.
- Daily requests
- Drives the monthly model and API line almost directly: requests per month multiplied by the per-thousand cost of the chosen model tier. High volumes also lift the infrastructure tier and the complexity score.
- System integrations
- Adds a fixed allowance per connected system, because each one needs authentication, error handling, idempotent writes and a reconciliation job. The complexity score steps up at one, three and six systems.
- Data complexity
- Multiplies the whole build range. Clean, structured data leaves the multiplier at one. Messy, unstructured data raises it, since ingestion, cleaning and evaluation absorb the difference.
- RAG required
- Adds a retrieval allowance when the project type does not already include a knowledge layer, and raises the model line because retrieved context makes every call larger. A RAG knowledge system already has it built in.
- Agent behaviour
- Adds an allowance for a single agent with approval steps, or a larger one for several coordinating agents, less whatever the project type already includes. A project type that carries an agent in its base range, such as an AI agent for a workflow or an enterprise multi-agent platform, is scored at that level even when this control is left at None. Agents also fan out into several model calls per request, which lifts the API line.
- Security tier
- Regulated or sensitive data multiplies the build range to cover data boundaries, redaction, audit trails and review, and moves the infrastructure tier up. It is a multiplier rather than an add-on because it touches every layer.
- Model quality
- Chooses the per-thousand-request cost band for the model line. Economy models suit classification and extraction, premium models suit reasoning-heavy work, and most production systems route between them by task.
- Expected scale in 12 months
- Customer-facing systems multiply the build range, since latency, abuse handling and uptime carry real engineering. Company-wide and customer-facing scale also raise the infrastructure tier and the complexity score.
- Support requirement
- Does not change the build range. Every build comes with a 14-day support window after launch, then paid support if you want it. This control changes who carries the upkeep line after that window, which the model sets as a share of the build cost a year and shows per month.
Implementation figures are rounded to the nearest $5k and monthly figures to a sensible step, so the calculator never shows a precision it does not have. The complexity tier is a score across nine of the eleven inputs, since model quality and support requirement move the monthly lines rather than the tier. It is not a function of the dollar figure, which is why a cheap build can still read as high complexity when the data or the security context demands it.
What drives AI project cost
The eleven controls above feed eight underlying drivers, and those drivers set the range. Here is what sits behind each one, and what you can do before a project starts to land at the lower end of it.
Scope and project type
The largest lever is what the system has to do, not which model it uses. An assistant that answers questions from a document set is a different build from an agent that reads a request, checks three systems and writes back with an approval step in between. Fix the scope to one workflow and one owner before asking for a price. Each additional user group, channel or edge case adds interfaces, states and tests, and a system that serves customers directly needs more engineering around failure than one used by a single team.
Data readiness
Data is where estimates go wrong most often. A well-kept database with named fields and a known owner is the easy case. Scanned PDFs, mailbox archives, spreadsheets with hand-typed categories and documents with conflicting versions need ingestion, cleaning, deduplication and a decision about which source is right. That work is invisible in a demo and dominant in production. The cheapest thing you can do before a project starts is to name the systems of record, find out who owns them and look at what the exports actually contain.
Integrations
Each connected system carries its own authentication, rate limits, failure modes and data shapes. Reading from a CRM is one thing. Writing back to it safely, with retries that do not create duplicates and a reconciliation job that catches what was missed, is where the effort sits. Xero, MYOB, HubSpot and Salesforce are the systems we are asked to connect most often, and the allowance per system is real whichever it is. Count them honestly, including email, calendars and internal databases. The shared plumbing for authentication, retries and logging is built once, so the first integration tends to carry more of the effort than the ones that follow.
Evaluation depth
A system that must be right needs a way to prove it. That means a set of real inputs and expected outputs written with the people who know the domain, a harness that scores every change and a release gate that fails when quality drops. The depth scales with consequence: a drafting assistant a person reviews can start light, while an agent that sends invoices or answers customers needs regression tests before every release. Evaluation is not overhead. It is what separates a promising prototype from something the business can rely on, and it is a line in every serious estimate.
Security tier
Regulated or sensitive data changes the design, not just the paperwork. Personal, health, financial or legally privileged information needs data boundaries, redaction before anything leaves your account, permission-aware retrieval so people only see what they may, least-privilege access and an audit trail of every model and tool call. The system is deployed in the region you need; for Australian data residency that is the AWS Sydney region, and the model endpoint has to respect the same boundary. None of this is optional in those settings, which is why the calculator treats it as a multiplier on the whole build rather than a line at the bottom.
Model choice
Model providers are components, and the choice moves the monthly line far more than the build. Economy models handle classification, extraction and routing at a fraction of the cost of the largest ones. Premium models earn their price on genuine reasoning. Most production systems route by task, so the real bill sits between the bands. The engineering that makes routing possible, together with caching, structured outputs and per-feature cost telemetry, is part of the build, and it is what stops the API bill from growing faster than usage.
Scale
A pilot for one team can tolerate a slow response and a manual fix. A company-wide rollout needs single sign-on, role-based access, onboarding and someone to answer questions. A customer-facing system needs latency targets, abuse handling, rate limiting and uptime that somebody is accountable for. The same feature can therefore land in three different ranges depending on where it will be in a year. Build for the scale you will reach in twelve months, not the one you have today and not the one on the five-year plan.
Operations
Software that uses models drifts. Providers change versions, source data changes shape, and the questions users ask move away from the ones you tested. Somebody has to watch the traces, re-run the evaluation suite, patch integrations and answer the page when a job fails overnight. Budget for it as a share of the build each year, whether that is your own team working from the documentation and handover ComplxAI includes, or ongoing operation by ComplxAI, priced in writing like every other stage. Systems with no operations budget are the ones that quietly stop being used.
What each complexity tier usually looks like
The tier is a shape, not a price. Two systems in the same tier can differ in cost by a wide margin, but they tend to need the same kind of engineering and the same kind of engagement. For concrete examples of these shapes, see what we've built.
| Dimension | Low | Moderate | High | Very high |
|---|---|---|---|---|
| Typical shape | One workflow, one team, a person reviews every output. | A knowledge system or a single agent with approvals, several integrations. | Several workflows or agents, regulated data, company-wide rollout. | A platform: many agents, tenants or interfaces, customer-facing at volume. |
| Data | Structured, one system of record. | Mixed sources, some cleaning and deduplication. | Unstructured documents, conflicting sources, permission rules. | Many systems, continuous ingestion, freshness guarantees. |
| Integrations | None to two, mostly read-only. | Two to five, some write-back. | Five or more with write-back and reconciliation. | Many, event-driven, across every system of record. |
| Evaluation | Spot checks and a small test set. | Golden dataset and a scored harness. | Regression suite gating every release. | Continuous evaluation and per-tenant quality dashboards. |
| Security | Standard access controls. | Role-based access, secrets management. | Data boundaries, redaction, audit trail. | Tenant isolation, data residency, formal review. |
| Where the engagement starts | Discovery scopes it. Strategy is there if the plan needs a build, buy or defer call. | Discovery, then a plan for the first 12-week block with evaluation in scope from the start. | Discovery and Strategy, so security and evaluation are planned as deliverables before the first block. | A programme planned in Strategy and delivered block by block, each block priced in writing. |
From a range to a written price
The calculator is step zero. The number you can plan around comes from the five-stage engagement described in how we work, and the first two stages cost nothing.
- CalculatorAn indicative range in a few minutes. Nothing to sign up for, no figures stored.
- Intro callFree, 30 minutes. Whether the idea is worth pursuing at all.
- Pre-discoveryFree, one hour. Scope, data, systems and a first view of fit.
- Discovery2–4 weeks from a fixed fee. A use-case map scored by impact, effort and risk.
- Strategy2–4 weeks, priced upfront. A board-ready plan with build, buy and defer calls.
- Implementation12-week blocks, each priced in writing, working software every couple of weeks.
You can stop at any boundary, the IP is yours from the first paid stage, and there are no mandatory retainers. If the estimate suggests something the size of an AI product rather than an internal tool, Discovery is where that gets tested before anyone commits to a build.
Where to go next
AI product development
AI-native products and SaaS MVPs, from discovery to scale.
AI product developmentAI development
The build behind the internal-tool, RAG and feature ranges on this page, scoped and delivered in 12-week blocks.
Custom AI development servicesAI agent development
What the single-agent and multi-agent allowances actually buy: tool calls, approval steps and the evaluation around them.
Custom AI agentsCost report
Australian AI development cost report 2026, in preparation: the cost drivers and report structure.
Australian AI development cost reportROI calculator
Estimate what the same system could return in reclaimed hours and capacity.
ROI calculatorAI engineering
The production layer behind every range on this page: retrieval, evaluation, observability, AWS.
Production AI engineeringQuestions people ask before they ask for a price
If yours is not here, the intro call is free and takes half an hour.
Book a free intro callWhy is the range so wide?
Because a narrow one would be false. Two projects with the same inputs can differ by the state of the data, the number of edge cases in the workflow and how much evaluation the consequences demand. The calculator shows the band a project of that shape usually lands in. Discovery is where the band becomes a number.
Where do my inputs go?
The maths runs in your browser, and the figures you type are not stored or sent to us. What is recorded is an analytics event saying the calculator was used, along with the project type and the tier the result fell in. Those events go to the analytics tools our privacy policy names, currently Google Analytics 4 and the Meta Pixel, in the same way as any other page view on this site.
Does the implementation range include model and hosting costs?
No. The build range covers design, engineering, integration, evaluation, deployment, documentation, handover and a 14-day support window after launch. Model and API usage, infrastructure and upkeep are shown as separate monthly lines because you pay them for as long as the system runs, whoever operates it.
How do I turn this into a real price?
Book the free intro call, or request a proposal if you already know the scope. If the idea holds up we run a free one-hour pre-discovery, then Discovery from a fixed fee. Every stage after that is priced in writing before it begins, you can stop at any boundary, and the IP is yours from the first paid stage.
Should I use this or the ROI calculator?
Both, in either order. This page estimates what a system costs to build and run. The ROI calculator estimates what the time it frees up is worth. Together they give a first view of payback that Discovery can then test against your real numbers.
Want the range to become a number?
Bring the inputs you used here to a free 30-minute intro call. We will tell you whether the shape makes sense before anyone talks about price, and every stage after that is priced in writing.