Executive summary (read this first)

How much does AI software cost in 2025? The short answer: less than last year, and more predictable than you think—provided you scope tightly and design for governance, observability, and cost control from day one.

Planning‑grade numbers (pilot to production):

• AI assistant (Q&A with citations): A$20k–A$60k (US$15k–US$40k / £12k–£32k / NZ$24k–NZ$65k) build; hundreds per month to run at modest volumes.
• AI automation / one agent workflow (with approvals): A$40k–A$120k (US$25k–US$80k / £20k–£63k / NZ$48k–NZ$130k) build; low thousands per month at scale.
• Custom AI app (portal + multiple workflows): A$90k–A$250k+ (US$60k–US$170k+ / £47k–£135k+ / NZ$108k–NZ$300k+) across phases; run costs vary with usage but can be budgeted and capped.

In this guide you’ll get:

• What drives cost (build + run) in simple terms.
• Budget bands by project type and complexity (AU/US/UK/NZ examples).
• A napkin ROI calculator and four worked examples.
• A run‑cost formula (so you can forecast monthly spend), with switches that lower costs fast.
• Pricing models (fixed price, T&M with cap, phased) and which one suits your risk.
• Region notes (GST/VAT, data residency expectations) for Australia, the United States, the United Kingdom and New Zealand.
• Procurement checklists you can paste straight into your RFP.

Promise: No jargon, no hype—just numbers and plain English.

Who this guide is for

• Business owners, founders, CEOs, CMOs, COOs evaluating AI development services.
• IT leaders/CTOs who must justify cost and control risk.
• Operations, customer service, sales, finance teams who want time back and fewer errors.
• Organisations operating in Australia, the USA, the UK, and New Zealand.

Quick answers (the 60‑second cheat sheet)

• Fastest win: a grounded AI assistant for your website or internal knowledge (citations + refusal). Time to pilot: 2–5 weeks. Build: A$20k–A$60k.
• Next step: one approval‑gated automation (e.g., refunds, invoice posting). Pilot + hardening: A$40k–A$120k over 4–8 weeks.
• When it gets expensive: big scopes, messy content, uncertain ownership, and “one giant prompt” projects without governance.
• How to keep costs predictable: tight scope, citations + refusal (less rework), caps & alerts, routing to smaller models, caching, and a weekly quality & cost review.
• What great partners show you: traces (where facts came from), dashboards (quality/latency/cost), approval flows, and rollback stories.

What actually drives cost (in plain English)

There are two buckets: build (one‑off project work) and run (monthly usage). Both can be managed.

1) Build costs (one‑off)

1. Discovery & design (1–2 weeks): clarity on one persona, one flagship workflow, content inventory, acceptance criteria.
2. Content work: ingest → clean → chunk → tag your policies, FAQs, manuals and price books. (Good content = fewer hallucinations and cheaper prompts.)
3. Assistant with citations & refusal: the Q&A layer that answers based on your content and shows its sources—or politely declines if evidence is thin.
4. Automation/agent workflow: “does things” with approvals and an audit trail (e.g., raise a refund, draft a ledger entry).
5. Integrations: CRM/helpdesk/ERP/payment systems (time varies by vendor quality).
6. Governance & security: tenant isolation, RBAC/SSO, region pinning (AU/US/UK/NZ), PII redaction, logs you can show to procurement.
7. LLMOps & evaluation: traces, golden test set, canary + rollback, dashboards for grounding %, latency and cost.
8. Change management & training: short videos, runbooks, and a “how we approve/refuse” cheat sheet.

2) Run costs (monthly)

1. Model usage (tokens): the text in and out of the AI; biggest driver early on.
2. Retrieval stack: vector searches, optional keyword/rerank passes; usually cents per thousand queries.
3. Storage: object store for your content + vector DB; modest unless you store everything forever.
4. Logging/observability: traces for audit and debugging; sized by traffic.
5. Third‑party APIs: e.g., payment or verification costs inside workflows.
6. Support & approvals: the time humans spend approving exceptions (still cheaper than doing the whole task).

Good news: Run costs are controllable. Prompt hygiene, caching, routing, and quality retrieval cut spend dramatically.

Budget bands by project type (with scope notes)

Use these to plan. Your exact quote should tie to deliverables and acceptance criteria—not vibes.

A) AI Assistant (Q&A with citations + refusal)

Scope: Clean ingestion of your content; metadata + indexing; Q&A that always cites sources and refuses when evidence is weak; basic dashboard and run‑cost caps.

• Build:
  • AU: A$20k–A$60k
  • US: US$15k–US$40k
  • UK: £12k–£32k
  • NZ: NZ$24k–NZ$65k
• Timeline: 2–5 weeks to pilot.
• Run: Hundreds per month at modest volumes; scales linearly but stays predictable with caps.

Cost accelerators: chaotic document sets, heavy PDF/OCR issues, or a mandate to “support eight languages on day one”.
Cost dampeners: a tidy folder of policies/FAQs/product pages and a single language to start.

B) One Agent Workflow (automation with approvals + audit)

Scope: One “does something” flow (e.g., returns triage, refund under A$250 auto‑approve, over A$250 with manager approval), deterministic tool schema, idempotency, audit export, and dashboards.

• Build:
  • AU: A$40k–A$120k
  • US: US$25k–US$80k
  • UK: £20k–£63k
  • NZ: NZ$48k–NZ$130k
• Timeline: 4–8 weeks including hardening.
• Run: Low thousands per month once active users ramp; exception rates and approval times matter more than tokens.

Cost accelerators: brittle legacy APIs, missing approval rules, or “just let the agent decide everything” (don’t).
Cost dampeners: clear thresholds, read‑only “dry run” first, and a single, trusted system of record.

C) Custom AI App (portal + multiple workflows)

Scope: Assistant + 2–4 agent flows + integrations + multi‑tenant admin + SSO + metering/billing + residency controls; production‑grade evaluation and rollback.

• Build:
  • AU: A$90k–A$250k+ (phased)
  • US: US$60k–US$170k+
  • UK: £47k–£135k+
  • NZ: NZ$108k–NZ$300k+
• Timeline: 8–16 weeks across phases.
• Run: From low thousands into tens of thousands monthly at scale; predictable with entitlements and caps.

Cost accelerators: “all‑in‑one from day one”, multi‑region, bespoke SSO/SCIM, or deep legacy data clean‑up.
Cost dampeners: phase the release, start with one persona and two workflows, and offer residency as a Pro/Enterprise toggle.

The run‑cost formula you can trust

For assistants (Q&A): Cost per successful answer=Tokens+Retrieval stack+Logging/ObservabilitySuccessful grounded answers\textbf{Cost per successful answer} = \frac{\text{Tokens} + \text{Retrieval stack} + \text{Logging/Observability}}{\text{Successful grounded answers}}Cost per successful answer=Successful grounded answersTokens+Retrieval stack+Logging/Observability​

For agents (automations): Cost per successful task=Planning tokens+Retrieval+Tool/API fees+Logging/Obs+Support for approvals/exceptionsTasks completed without rework\textbf{Cost per successful task} = \frac{\text{Planning tokens} + \text{Retrieval} + \text{Tool/API fees} + \text{Logging/Obs} + \text{Support for approvals/exceptions}}{\text{Tasks completed without rework}}Cost per successful task=Tasks completed without reworkPlanning tokens+Retrieval+Tool/API fees+Logging/Obs+Support for approvals/exceptions​

What reduces both costs fast:

• Prompt hygiene (short, modular system prompts; no copy‑pasted lore).
• Caching (answers and embeddings).
• Routing (small models for formatting/classification; reserve big ones for reasoning).
• Better retrieval (clean chunks + metadata filters → fewer tokens and retries).
• Selective reranking (only on risky queries).
• Entitlements & caps (protect your margin and your customer’s bill).

Worked examples (so the numbers feel real)

These are illustrative. Replace with your usage and rates for precise forecasts.

Example 1 — Website Product Assistant (30k visits/month)

• Traffic: 3,000 questions/month reach the assistant; 40% deflected from emailing support.
• Tokens/answer: 1,000 average (including context).
• Retrieval: 3 vector queries + 1 keyword on 50% of questions.
• Logging/Obs: light.
• Grounded answers: 90% (10% refused with helpful links—still a win).

Indicative COGS per grounded answer: A$0.012–A$0.02
Monthly COGS: ~ A$32–A$55 (3,000 × 0.9 × per‑answer cost).
Business impact: 1,200 fewer support emails (40% deflection) × 3 minutes saved each × A$40/hour ≈ A$2,400 saved/month.
Payback on a A$30k build: ~ 13 months without counting conversion uplift; with modest conversion gains, payback typically < 9 months.

Example 2 — Returns & Refund Agent (eCommerce)

• Volume: 4,000 returns/month; agent proposes outcome; auto‑approves ≤ A$250, escalates above.
• Approval time: 20 seconds average; 90% pass at first attempt.
• COGS per successful task: ~ A$0.08–A$0.25 (tokens + retrieval + logging + tiny share of support time).
• Monthly COGS: A$320–A$1,000.
• Value: 4 minutes saved per case × A$35/hour × 4,000 ≈ A$9,333/month saved + more consistent decisions and fewer chargebacks.

Payback on A$90k build: ~ 10 months (faster if chargebacks shrink or NPS lifts).

Example 3 — Finance Inbox (AP Automation)

• Volume: 6,000 invoices/month; agent drafts ledger lines; human approves any posting.
• COGS per task: A$0.10–A$0.30 depending on complexity and API fees.
• Monthly COGS: A$600–A$1,800.
• Value: 3 minutes saved per invoice × A$60/hour × 6,000 ≈ A$18,000/month.
• Payback on A$120k build: ~ 7 months.

Example 4 — Sales Copilot (Proposals & Follow‑Ups)

• Volume: 1,000 proposals/month drafted from call notes; humans edit and send.
• COGS per draft: A$0.05–A$0.15.
• Monthly COGS: A$50–A$150.
• Value: 10 minutes saved × A$80/hour × 1,000 ≈ A$13,333/month + win‑rate lift (even +2–3% is material).
• Payback on A$60k build: ~ 5 months.

Pricing models (pros & cons for buyers)

1) Fixed price (by phase)

• Best for: well‑defined scope (assistant + one workflow).
• Pros: budget certainty; acceptance criteria aligned to outcomes.
• Cons: change requests slow you down if scope drifts.

2) Time & Materials with a cap

• Best for: iterative work where discovery will surface surprises.
• Pros: flexibility; you keep momentum.
• Cons: you need a strong weekly steerco and transparent dashboards.

3) Phased (discovery → pilot → hardening)

• Best for: sensible default.
• Pros: go/no‑go at each gate; lessons feed the next phase.
• Cons: requires discipline to stop scope creep.

4) Outcome‑linked bonus (optional)

• Best for: when you can measure uplift clearly (deflection rate, time saved).
• Pros: aligned incentives.
• Cons: needs agreed baselines and a fair measurement window.

Tip: Whatever you choose, insist on acceptance criteria tied to grounded answers %, refusal correctness %, task success %, latency budgets, and “cost per successful answer/task” targets.

Hidden costs & traps (and how to avoid them)

• Messy content → extra cleaning; fix the pipeline once; it pays back quickly.
• One giant prompt → unpredictable results and cost; split roles (retrieval, planner, verifier) and version each.
• No refusal behaviour → hallucinations and support pain; refuse politely with helpful links.
• Co‑mingled indices/logs → privacy risk and lost deals; use per‑tenant namespaces.
• Unlimited usage promises → surprise bills; set entitlements, caps and alerts at 70/85/95%.
• Ignoring run costs in quotes → ask for a cost model and levers to reduce it.
• Skipping LLMOps → regressions after launch; add golden sets, canaries, and rollback.
• Approvals afterthought → risky automations; define thresholds and audit up front.

ROI: the three buckets and a quick calculator

1) Time saved (hours back): Time value=hours saved per user per week×loaded hourly rate×users×48\text{Time value} = \text{hours saved per user per week} \times \text{loaded hourly rate} \times \text{users} \times 48Time value=hours saved per user per week×loaded hourly rate×users×48

2) Revenue gained (conversion uplift, faster proposals): Revenue uplift=(baseline revenue×uplift %)\text{Revenue uplift} = (\text{baseline revenue} \times \text{uplift \%})Revenue uplift=(baseline revenue×uplift %)

3) Risk reduced (fewer chargebacks, fewer errors):
Establish a baseline (e.g., average monthly chargebacks) × expected reduction %.

ROI snapshot: Annual benefit−(Build+12 months run)\text{Annual benefit} – (\text{Build} + \text{12 months run})Annual benefit−(Build+12 months run)

If positive with a payback < 12 months, you’re in strong shape.

Region‑specific notes (AU, US, UK, NZ)

Australia (AU)

• Expect GST on invoices. Larger buyers (finance, public sector) often require AU data residency, private networking, and audit exports per tenant.
• Plain‑English privacy notices help adoption.

United States (US)

• Taxes vary by state; many buyers prefer usage‑based elements with annual prepay discounts.
• Expect questions on incident SLAs, sub‑processors, and data use (no training on your data without consent).

United Kingdom (UK)

• Quote ex‑VAT; VAT handled on invoicing. Public sector prefers UK/EU residency, accessibility, explainability, and simple overage rules.
• DPIAs are common for higher‑risk workflows.

New Zealand (NZ)

• Expect GST and interest in overseas disclosures. Government/health buyers may ask for NZ or near‑region hosting and robust audit.

If residency matters, make it a plan toggle in Pro/Enterprise: “Store and process data in AU/UK/US/NZ region”.

What a good proposal includes (copy this checklist)

• Scope: one assistant (citations + refusal) + one workflow (approvals + audit).
• Deliverables: content pipeline, retrieval policy, tool schema(s), approval thresholds, dashboards.
• Acceptance criteria: grounded answers %, refusal correctness %, task success %, P95 latency, cost per successful answer/task, rollback tested.
• Timeline: week‑by‑week milestones (2–5 weeks for assistant; 4–8 for workflow).
• Security: tenant isolation, PII redaction, region pinning, SSO/RBAC.
• LLMOps: golden set v1, canary + rollback, traces.
• Commercials: build by phase, run‑cost model and assumptions, overage rules, caps/alerts.
• Handover: runbooks, diagrams, admin training, support options.
• Assurance pack: data flows, DPA terms, sub‑processor list, deletion & certificate of destruction.

Year‑1 & Year‑2 budget planning (and the team you actually need)

Year‑1 (pilot to production)

• Build: A$40k–A$180k across an assistant + one agent workflow (+ extras if you add more flows).
• Run: A$300–A$3,000+/month depending on adoption.
• Internal time: product owner (0.2–0.5 FTE), content owner (0.1–0.3), approvers (minutes/day).

Year‑2 (scale & optimise)

• Additional flows or regions: A$40k–A$120k per added workflow, often less as patterns repeat.
• Run: scales with usage; caps and entitlements keep it predictable.
• Optimisation: prompt hygiene, routing, caching → COGS per answer/task drops 20–35% vs Year‑1.

Team you need (minimum viable on the vendor side): architect, AI/ML engineer, full‑stack dev, data/ETL engineer, designer (UX), QA/automation, delivery manager. You do not need a dozen people to succeed.

How to keep spend under control (seven levers)

1. Scope like a hawk: one persona, one flagship workflow + assistant.
2. Citations + refusal: less rework, more trust; don’t pay to generate confident nonsense.
3. Prompt hygiene: shorter, modular prompts save tokens every single call.
4. Routing: small models for simple steps; big ones only for hard reasoning.
5. Caching: response cache for frequent Q&As; embedding cache on content hash.
6. Hybrid retrieval: metadata filters → fewer, better chunks → cheaper contexts.
7. LLMOps discipline: golden sets, canaries, and automatic rollback; weekly review of cost per successful answer/task.

FAQ (plain English)

Q: What is the single biggest driver of AI cost?
A: Scope and content quality. Tight scope + clean content = smaller prompts, fewer retries, lower run costs.

Q: Can we cap monthly costs?
A: Yes. Use entitlements (answers/tasks/region/storage limits), soft warnings at 70/85/95%, and hard caps with admin override.

Q: Do we have to “train a model on our data”?
A: For most business apps, no. Use retrieval with citations so answers come from your sources and stay up to date.

Q: How fast can we see value?
A: 2–5 weeks for an assistant; 4–8 weeks for a workflow with approvals. Pick high‑volume, rule‑bound tasks first.

Q: Will AI replace our team?
A: It replaces toil, not people. Your team handles exceptions, approvals, and customer nuance.

Q: How do we avoid hallucinations?
A: Clean content → good retrieval → citations + refusal when evidence is thin. Measure grounding %.

Q: What if a model change increases our costs?
A: Keep changes behind canaries and rollback. Route simple tasks to smaller models; cache aggressively.

Q: Do we need a massive budget to start?
A: No. Many SMEs launch meaningful pilots under A$60k. Start small, prove ROI, then scale.

Talk to a real human about your AI project

Digital One Agency — AI SaaS Tool Builds • AI Bot/Agent Builds • Bespoke AI Software
We design and deliver AI that pays its way: grounded answers, safe automations with approvals, and dashboards that keep quality and cost on track.

👉 Book a Free 1‑Hour Strategy Session: https://digitaloneagency.com.au/book-a-meeting/
📞 Call us (24/7): +61 2 9057 9079
📧 Email: contact@digitaloneagency.com.au

We work globally across Australia, the USA, the UK, and New Zealand. We’ll scope your project in one working session and provide a clear plan, budget, and run‑cost model you can take straight to your board.