Custom AI projects rarely fail because the model isn't clever enough. They fail because the problem was vague, the data wasn't ready, nobody defined what "good" looked like, or the system was never looked after once it launched. A clear process fixes all four.
This article is part of AI for Business: 15 Practical Use Cases Across Different Industries — our overview of where AI delivers real value for businesses.
Step 1: Define one problem precisely
"We want AI" isn't a brief. "Support staff spend hours a day answering the same returns questions; we want 24/7 answers from our policy documents, with a handover to a person when needed" is. Write down:
- who will use it and for what task;
- ten questions or inputs it must handle perfectly;
- what it must never do (guess prices, give legal advice, share one customer's data with another); and
- how you'll measure success — time saved, response time, accuracy on test cases.
Not sure which problem to pick? Start with 15 practical AI use cases.
Step 2: Audit your data
The AI can only be as good as the information behind it. List the documents and systems it needs, then check: are they current? Are there duplicates or conflicting versions? Are some scanned images that need OCR? Who is allowed to see what? This step often reveals that tidying the content is half the project.
Step 3: Check feasibility with a short discovery
If you're unsure whether AI can do the job well, run a fixed-price discovery of one to two weeks. It should produce an architecture recommendation, a data readiness assessment, a realistic scope and cost, and ideally a small working prototype on a subset of your documents. It's far cheaper to learn "not yet" here than halfway through a build.
Step 4: Choose the architecture
| Decision | Typical options | What drives the choice |
|---|---|---|
| How the AI knows your business | RAG (retrieval from your documents), fine-tuning, or both | Whether answers must be current and cited — usually RAG. See RAG vs fine-tuning. |
| Language model | GPT-4o, Claude, Gemini, or open-source such as Mistral and Llama | Answer quality in your domain, cost per query, data rules |
| Answer or act? | Assistant/chatbot, or an agent that takes actions | Risk and complexity — see chatbots vs agents |
| Hosting | Cloud API, your own cloud account, or on-premise | Where your data is allowed to go |
| Interface | Web widget, portal, Slack/Teams, or an API into your own app | Where users already work |
| Integrations | CRM, helpdesk, database, store, document storage | What the AI must read from or write to — see AI integration |
Step 5: Build in small, testable pieces
Build the data pipeline first, then the AI backend, then the interface and admin tools. Test each piece against your real examples as you go, rather than all at the end. Our typical sequence for a custom AI assistant:
- Week 1 — Discovery: domain, data sources, success criteria.
- Week 2 — Architecture: models, search index, hosting, integrations.
- Weeks 3–5 — Build: ingestion, AI backend, interface, admin panel.
- Week 6 — Review and tune: accuracy testing, prompt improvements, edge cases.
- Weeks 7–8 — Launch: staging and user acceptance testing with your team, then production and monitoring.
Smaller AI features added to an existing product can ship in 3–4 weeks; multi-module platforms take 10–12.
Step 6: Test like a sceptic
- Run every real example and record whether the output was correct, partly correct or wrong.
- Try to break it: ambiguous questions, questions it shouldn't answer, documents it doesn't have.
- Have your subject expert read outputs — they'll catch answers that are "correct" but unhelpful.
- Confirm it says "I couldn't find that" instead of guessing.
Step 7: Launch, monitor, improve
Launch to a small group first. Log unanswered and low-confidence cases, review them weekly, and fix the content or prompts behind them. Keep an eye on running costs as usage grows. Plan who owns the system after launch: your team, a support retainer, or an ongoing product partnership.
We don't just build AI systems; we operate our own, including MyLegalWare and ClaimResolveAI. Running a live product means regulations and documents change, users ask things nobody predicted, and the system has to keep up. Plan for that from the start.
Next steps
Still deciding whether to build at all? Read custom AI vs off-the-shelf AI tools. Working out a budget? See how much an AI solution costs. Ready to scope yours? Talk to our team — we sign an NDA before any documents are shared.