Choose off-the-shelf AI for generic tasks, build custom AI where your own data and processes create the value, and for most mid-sized organisations take the middle path: buy Claude API (or another model API) access and build the integration, extraction and approval layer around it. Five questions decide which: how specific the problem is, what it costs, what happens when AI is wrong, whether you have the data, and who maintains it.
Key takeaways
- Litmus test: "If a competitor subscribed to the same tool tomorrow, would they have our capability?" If yes, buy.
- Price the problem before evaluating solutions. Some problems are not worth automating.
- Match human oversight to the cost of a wrong answer.
- You can buy Claude API access directly from Anthropic or through Amazon Bedrock, Google Cloud Vertex AI and Microsoft's AI platform.
Why do build vs buy decisions go wrong?
Hypothetical scenario: a manufacturer buys an "AI quality control system" from an overseas vendor. The demo was impressive. Six months later it misses most defects that the old manual process caught. The vendor says the data is formatted wrongly. They sold a generic model and hoped the problem was generic too.
The reverse happens too: organisations spend heavily building custom AI for a problem an existing subscription already solves. Five questions prevent both.
Question 1: Is the problem generic or specific?
Generic problems:
- Summarising meeting notes
- Basic document OCR
- Email categorisation
- Simple FAQ responses
- Code completion
Buy for these. Microsoft 365 Copilot, Claude, ChatGPT Enterprise, Intercom and similar tools are trained for exactly these tasks.
Specific problems:
- Quoting from years of your own sales and margin history
- Defect detection for your products under your lighting
- Compliance checks against Australian regulation plus your internal policies
- Tender responses from your past winning bids
Build or hybrid for these. No off-the-shelf AI has your context, and the context is the value.
If your competitor subscribed to the same tool tomorrow, would they have the same capability?
- Yes, same capabilityBuy: use a market-leading product
- No, our context is uniqueBuild or hybrid: your data creates the advantage
Question 2: How much is the problem costing you?
Organisations sometimes spend months evaluating AI for a problem that costs little each year. Do the maths first.
| Factor | How to calculate |
|---|---|
| Hours per week on the task | Ask the team; they know |
| Fully loaded hourly cost | Salary ÷ working hours, uplifted for super, leave and overheads |
| Annual cost | Hours × hourly cost × working weeks |
| Error and rework cost | From historical issues |
Hypothetical example: an accounts payable team processes 200 invoices a month at eight minutes each, about 27 hours a month. At a hypothetical loaded cost of $85 an hour, that is roughly $27,000 a year, plus rework.
- Annual cost of the problemLabour plus rework
- Solution costBuild or subscription, plus running cost
- Expected reductionMeasured in a pilot
- PaybackSolution cost ÷ annual saving
If payback is comfortably inside your hurdle period, proceed. If the solution costs many times the annual problem, walk away. Our ROI calculator helps.
Question 3: What happens when it gets it wrong?
A chatbot giving a slightly wrong product suggestion is embarrassing. An AI paying a large invoice to the wrong supplier is serious.
| If the AI is wrong... | Impact | Recommendation |
|---|---|---|
| Someone gets slightly wrong general information | Low | Automate, monitor weekly |
| An internal team wastes time | Medium | Automate with spot checks |
| A financial transaction is wrong | High | A person approves each one |
| A regulatory or legal breach is possible | Critical | Person decides; AI assists only |
Under the Privacy Act and Australian Consumer Law, "the algorithm did it" is no defence. Design approval workflows accordingly.
Question 4: Do you have the data?
Many failed AI projects fail on data: scattered across spreadsheets, legacy systems and someone's inbox.
| Data requirement | Red flag | Green light |
|---|---|---|
| Historical records | "We'd have to export from the old system" | "It's in our ERP with API access" |
| Consistent format | "Every rep has their own template" | "Standard templates for years" |
| Volume | "About 50 a year" | "Hundreds a month" |
| Outcomes | "We don't track if it worked" | "We record win and loss on everything" |
Rule of thumb: if you cannot pull 12 months of clean, labelled data within a week, fix the data before building. See data quality and AI readiness.
Question 5: What is your maintenance reality?
AI systems drift. Models change, APIs are deprecated, regulations update and people leave.
| Item | Buy | Build |
|---|---|---|
| Licence or hosting | Subscription | Hosting and model usage |
| Vendor support | Included | Your team or partner |
| Internal maintenance | Light | Ongoing |
| Model review | Vendor's job | Scheduled, at least quarterly |
| Integration updates | Often included | Yours |
If nobody technical will be around in two years, avoid custom builds without a support arrangement.
| Assessment | Recommendation | |
|---|---|---|
| Generic problem, low annual cost | Low priority | Leave it manual for now |
| Generic problem, high annual cost | Evaluate products | Buy |
| Specific problem, moderate cost | Consider configuration | Buy and configure where possible |
| Specific problem, high cost, good data | High opportunity | Build or hybrid |
| Specific problem, high cost, poor data | Prep work needed | Fix data first, then build |
| Any problem, critical if wrong | High risk | Person approves every output |
What is the hybrid approach?
You often do not have to choose. Buy the intelligence; build the context.
Hypothetical example: a law firm where reviewing a commercial lease takes about six hours.
- Buy model access, such as the Claude API, for general comprehension.
- Build an extraction layer for the firm's clause types, with links back to the source text.
- Configure a review workflow so a lawyer checks every flagged clause.
If that cuts review to around 90 minutes for 20 leases a month, the firm recovers about 90 hours a month. See contract review AI.
- Model APIClaude, GPT or Gemini via API
- Your dataDocuments, history, rules
- Custom layerExtraction, integration, logic
- ApprovalA person reviews before action
How do you buy Claude API access in Australia?
Claude is Anthropic's model family. Organisations buy API access in four main ways:
| Route | How it works | Consider when |
|---|---|---|
| Anthropic API (Claude Console) | Create an organisation account, add billing, generate API keys, pay per token used | You want direct access and the newest features first |
| Amazon Bedrock | Enable Claude models in your AWS account, billed through AWS | You already run on AWS and want existing contracts, IAM and logging |
| Google Cloud Vertex AI | Enable Claude in Vertex AI Model Garden, billed through Google Cloud | You run on Google Cloud |
| Microsoft AI platform (Azure) | Claude models offered through Microsoft's AI platform, billed through Azure | You are standardised on Microsoft |
Before you buy, check:
- Data residency. Which Claude models are available in Australian regions on your chosen platform, and where prompts and outputs are processed and logged.
- Data use terms. Commercial API terms generally exclude training on your inputs by default; confirm in the agreement you sign.
- Rate limits and spend controls. Set usage tiers, budgets and alerts.
- Security. API keys in a secrets manager, access through your identity provider, logging of every call.
- Pricing. Per-token pricing varies by model and route; check the provider's current price list.
For sensitive workloads, organisations sometimes pair a commercial API with self-hosted open models on their own infrastructure. See OpenAI vs Claude vs Ollama and AI data sovereignty.
What do we recommend first?
When asked "build or buy?", our answer is often "neither, yet". First:
- Quantify the problem's cost.
- Audit the data.
- Define what "wrong" looks like and how much you can tolerate.
- Check whether an existing product gets you most of the way.
Many problems need an existing tool and a process change. The ones that need custom AI are where competitive advantage sits, provided the build is maintained. Our complete build vs buy TCO guide covers five-year costs.
How Solve8 can help
We help you price the problem, check the data and choose between buying, building and hybrid, then build the parts worth owning on your infrastructure in Australia with a person approving outputs. See our AI strategy service and custom AI development, or talk to us.
