# How Custom AI Development Works: My Process > How custom AI development actually works — the six-stage process, what happens at each step, and what you contribute. By a developer, not a consultant. Canonical: https://www.aimakers.co/custom-ai/process/ Last updated: 2026-09-24 ## How custom AI development actually works. [Custom AI](https://www.aimakers.co/custom-ai/)/Process Six stages, working software from week 1, fixed price after a one-page spec, weekly Friday demos. No theatre, no committee, no surprise invoices. [See the timeline →](https://www.aimakers.co/custom-ai/timeline/) ## Five principles, applied to every project These aren’t marketing values. They’re the things that have made our projects ship. ### Start from working code, not a blank page Every custom AI I ship starts from the OpenClaw foundation — a battle-tested AI agent runtime with 60+ integrations already wired in. I extend it for your business instead of writing from scratch. That collapses what used to be a 6–12 month build into 3–12 weeks. ### Show working software every Friday You see a live demo every week from week 2 onwards. No 12-week silent build then a big reveal. If the AI is heading the wrong direction, you see it in week 3 — not week 11. ### One developer, one decision-maker I work directly with your point of contact. No account managers, no PMs translating between you and the engineer. Faster decisions, fewer broken games of telephone. ### Fixed price, signed scope You sign off on a written technical spec before any code is written. Scope creep is your decision, not a surprise invoice. Every change-request is quoted in writing. ### Ship to production, not a demo Prototypes that look great in a sandbox and fall over in production are the most expensive thing in AI. I test against real data, real edge cases, real users — before launch. ## The six stages Same process whether the project is small or large — only the depth changes. Each stage has clear inputs from your side and concrete outputs from mine. ### Discovery Understand the workflow. Before scoping anything I sit with you (call, screen-share or in-person where a site visit makes sense) and watch the actual work happening. The point is to understand the messy reality, not the pitch-deck version. I finish discovery with a one-page process diagram showing exactly where AI fits. - Your time (1–2 hours) - Read access to relevant systems - A volunteer who does the work today - Process diagram - Scope draft - Honest go / no-go recommendation ### Technical design Decide what to build, in writing. I write a 3–8 page technical spec that nails down: what the AI does, what tools it has, what data it can access, how the admin panel works, what success looks like. You sign it. Quote becomes fixed price. - Sign-off on scope - Access to one test environment - A point person on your side - Written technical spec - Fixed-price quote - Project kick-off date ### Build (with weekly demos) AI logic + integrations + admin UI. The bulk of the project. I build in vertical slices: by end of week 1 of build there's a thing you can poke at, even if it's 10% of the final scope. Every Friday: live demo, screen-recording, list of next-week priorities. You shape the AI as it grows. - ~2 hours/week of your time for demos - Decisions on edge cases as they come up - Working software, deployed to staging - Friday demos & recordings - Audit log + evaluation harness ### Test & harden Prove it survives reality. Edge case testing, security review, real users running real tasks. I pre-build an evaluation harness — a set of test cases — so future changes don't silently break old behaviour. This phase is where mediocre AI projects ship and good ones don't. - Real test data - 3–5 user volunteers for UAT - Security/IT contact for review - Edge-case test report - Security review sign-off - UAT user feedback addressed ### Launch Production deploy + team training. Cut over to production — your server or my managed cloud at nora.work. Recorded training session for your team. Documentation for admins. A WhatsApp group / Slack channel for the first 30 days so questions get answered fast. - Production credentials - Team availability for training session - Live in production - Recorded training - Admin docs - 30-day support channel ### Steady-state Iterate, monitor, evolve. Custom AI is not "done" at launch — it gets better as your team uses it. Most clients move onto a managed plan for hosting, monitoring, model upgrades and ongoing tweaks. Or you self-host and call us for changes. - Real usage - Feedback loop - Monthly improvements - Model upgrades as new ones ship - Optional new tools / integrations ## Readiness checklist Six questions to answer before kicking off a custom AI project. If most are “yes”, you’re ready. If most are “not sure”, a short scoping chat is exactly where to start. ### 01.Do we have a clearly-defined process this AI will improve? Why it matters: Vague problems make for vague AI. "Reduce admin time on invoices" is workable. "Use AI to be more efficient" is not. ### 02.Is the data the AI needs accessible? Why it matters: If your data lives in a 1990s system with no API, building the connector might be 50% of the work. ### 03.Who decides what the AI should do — and who has authority to change it? Why it matters: Single decision-maker = fast project. 4-person committee = slow project. ### 04.What does success look like, in numbers? Why it matters: Hours saved, error rate, conversion, ticket deflection — pick one or two and I'll measure it from week 1. ### 05.Where will it be hosted? Why it matters: Managed cloud is fastest. On-premise is slower but gives you full data control. ### 06.What is the budget envelope, ballpark? Why it matters: Helps me scope realistically. A small budget and a large one are both fine — they buy different things. ## Process FAQ ### How is custom AI development different from a software project? Custom AI uses LLMs and AI agents at the core, but most of the work is still software engineering — integrations, admin panels, error handling, deployment. The "AI part" is typically 20–30% of the codebase. The rest is the plumbing that turns a prompt into a reliable production system. ### Do you train custom models? Almost never. For 95% of business use cases, frontier models (Claude, Gemini, GPT) with the right tools, memory and prompting outperform a custom-trained model by a wide margin — and cost a fraction. I do fine-tune when there's a clear case (proprietary terminology, regulated outputs, narrow classification tasks). Most of the time, agentic systems beat fine-tunes. ### What happens if it doesn't work? I catch this in week 1–2 of build with a working prototype. If it's not viable, you get an honest "this won't work and here's why" — not 8 weeks of dressed-up failure. I have walked away from projects in week 2; that's better than charging for a build that won't ship. ### Can we move suppliers later? Yes. Everything I ship runs on open standards — OpenClaw is open source, integrations use standard APIs, the model layer is provider-agnostic. You're not locked in. Your code is your code. ### What does ongoing support look like? Two flavours. Managed — I host, monitor, update, fix, for a fixed monthly fee quoted with the build. Self-hosted — you run it; I'm available on a retainer or per-change basis. Either way, no surprise vendor lock-in. ## Related - [Custom AI development](https://www.aimakers.co/custom-ai/)