Choosing the wrong AI development partner can waste months and hundreds of thousands of dollars. This checklist — based on patterns we've seen in our own client projects — helps you evaluate vendors objectively.
The 12-Point Evaluation Checklist
Key criteria: (1) Senior engineer ratio, (2) AI-specific portfolio, (3) Transparent pricing model, (4) Direct tech lead access, (5) Code ownership from day one, (6) Weekly demo cadence, (7) SOC 2 or equivalent compliance, (8) Realistic timeline estimates, (9) Clear IP assignment, (10) Reference clients in your industry, (11) Defined escalation process, (12) Post-launch support plan.
Onshore, Nearshore or Offshore?
Where the team sits changes how you work with it as much as what it costs.
| Model | Example for a US buyer | Shared working hours | Typical trade-off |
|---|---|---|---|
| Onshore | Agency in your own country | Full day | Highest rates; easiest for workshops and on-site work |
| Nearshore | Latin America (for the US), Eastern Europe (for Western Europe) | Most of the day | Mid-range rates; real-time collaboration |
| Offshore | Vietnam, India, the Philippines | 1–3 hours of call time with the US; most of the day for Singapore, Japan and Australia | Lowest rates; needs written specs and an async-first workflow |
Choose by how often you need live decisions, not by habit. Discovery-heavy work with frequent stakeholder sessions favours overlap; a well-scoped build with weekly demos runs well offshore. For AI work, the constraint that matters most is engineers who have already shipped LLM or RAG systems to production — any of the three models can meet it or miss it, so judge the team you will actually get, not the country.
Operational Risks to Check Before You Sign
Most vendor failures are operational, not technical. Check five things in writing:
• Key-person risk. Who is the named tech lead, who else knows the codebase, and how quickly is a replacement onboarded if someone leaves? Ask for the notice and handover terms. • IP and code ownership. IP should transfer to you on payment, the code should live in your repository from day one, and open-source and third-party components should be listed with their licences. • Data access. Agree which environments the vendor can reach, whether production data is ever copied to their machines, and how access is revoked at the end. For personal data, sign a data processing agreement that names where the data may be accessed from. • Continuity. Ask what happens if the vendor's office, a key person or a cloud region becomes unavailable. A documented business-continuity plan (for example under ISO 22301) is a fair thing to request. • Exit. Agree what you receive at the end — source code, infrastructure-as-code, credentials, documentation and a handover period — before the first invoice, not after the last.
Red Flags to Watch For
Watch out for: vague 'AI expertise' claims without specific project examples, account manager buffers between you and engineers, unusually low rates that indicate junior-heavy teams, no code review process, and firms that promise delivery timelines under industry norms.
Questions to Ask Before Signing
Ask: 'Can I interview the engineers who will work on my project?' 'What happens if a key engineer leaves?' 'How do you handle scope changes?' 'What's your QA process for AI model outputs?' 'Can I see your CI/CD pipeline?' These separate serious partners from body shops.
Next step
Shortlisting AI vendors and want proof before signing a large contract? Book a free 30-minute scoping call for an AI Pilot (from USD 3,000, 3–4 weeks, on your own data) that solves one use case so you can judge the team on delivered work; you get a fixed-price proposal within 3 business days.
Senior engineers in Vietnam, a clearly defined scope, and a written proposal within 3 business days.
10 years of experience working with Japanese companies — two of those years based in Japan. Has run the business since 2018 (as NKKTech since 2025) with a senior-only engineering model.