Integration depth
Supported APIs are different from legacy exports, browser-mediated systems, identity integration, and bidirectional actions.
Commercial scope guide
What determines custom AI development cost in Hong Kong?
A credible estimate is driven less by the number of screens than by integration depth, data condition, evaluation, permissions, and the operating responsibility the system must carry.
By Onyx Devs Lab · Published · Updated
Scope before priceDefine the operating boundary
Risk before featuresTest expensive unknowns early
TCO, not demo costInclude operation and ownership
The main cost drivers
Two interfaces that look similar can have very different production requirements.
Evidence and evaluation
High-trust work requires representative test sets, citations, error analysis, regression checks, and human-review rules.
Operational controls
Queues, retries, permissions, audit logs, monitoring, fallbacks, and support often determine whether a prototype can go live.
What a useful estimate contains
A single headline price without assumptions cannot support a procurement decision.
Defined workflow
Named users, inputs, outputs, exceptions, integrations, and consequential actions.
Delivery stages
Separate discovery, prototype, pilot, production hardening, training, and ongoing operation.
Acceptance and change rules
State measurable acceptance criteria, excluded scope, dependencies, and how new evidence changes the plan.
Delivery method
Map one workflow
Avoid estimating an undefined enterprise-wide AI programme.
Expose the unknowns
List data, API, permission, accuracy, latency, and adoption risks.
Price a validation stage
Resolve assumptions that could invalidate the architecture or business case.
Estimate production ownership
Include security, evaluation, monitoring, support, and handover.
Frequently asked questions
Can you provide a fixed price?
A fixed price can be responsible after the workflow, dependencies, and acceptance criteria are bounded. Earlier work is usually better structured as a defined assessment or validation stage.
Why is an AI prototype cheaper than production deployment?
A prototype demonstrates a capability. Production deployment must manage real permissions, failures, edge cases, observability, data handling, and accountable human intervention.
How can a buyer reduce cost risk?
Start with one measurable workflow, give early access to representative data and operators, and agree on stop or redesign conditions before full implementation.
Start with one concrete operating problem.
We will first clarify the workflow, evidence, data boundary, and outcome standard before recommending the next step.
Contact Onyx