AI IT Outsourcing Offshore Development

AI Development Cost: What Drives It and How Vietnam Compares

9 min read
AI Development Cost: What Drives It and How Vietnam Compares

AI development cost is hard to pin down because most published numbers mix three different bills: the one-time build, the monthly cost of running the system, and the team that keeps it accurate after launch. Guides that rank for this question quote ranges from a few thousand dollars to well over half a million, and both ends can be honest. The spread comes from what is being built and what the buyer counts.

This guide separates those bills, explains what moves each one, and shows how the team cost changes when the engineers work from Vietnam, using public salary data. We build AI systems from Hanoi. We do not publish prices, and the figures below are market data, not a quote from us.

Why published AI price ranges disagree

Put two price guides side by side and one says a chatbot costs $15,000 while the other says $130,000. Both can be right. One means a support bot on a hosted model that answers from a short FAQ. The other means a bot that reads customer records, books appointments, hands off to a human, and has passed a security review. The word "chatbot" covers both.

Three questions explain most of the gap:

  1. What does the system touch? An assistant that only reads public documents is a small job. One that writes to your CRM or billing system needs access control, audit logs and tests for every tool call.
  2. How good must it be, and who decides? "Good enough for an internal tool" and "accurate enough to answer customers unsupervised" are different projects.
  3. What is included? Many estimates stop at launch. Token bills, monitoring and the engineer who retunes prompts every month arrive later and are rarely in the first number.

The three bills: build, run and team

Three bills that make up AI development cost: the one-time build, the monthly run cost, and the team that keeps the system accurate

The build is what most quotes describe: discovery, data work, the model approach, integrations, evaluation and security. You pay it once, though a large change in scope or model can reopen it.

The run cost is what the system costs each month in use: model API calls or GPU hosting, vector storage, logging and monitoring. It grows with traffic, so a successful launch raises it.

Then there is the team that keeps the system useful. Models change versions, your documents change, and users find questions nobody tested. Somebody has to re-run the evaluation, fix the failures and watch the bill. For most business AI systems this is the largest of the three over a few years, and it is the one buyers forget.

What drives the build cost

Data readiness

The model is rarely the slow part. Getting the right data, in a usable form, with permission to use it, usually is. A retrieval assistant over a clean, current document library can start quickly. The same assistant over ten years of scanned PDFs, with duplicates and no owner, needs OCR, cleanup and decisions about which version is true before any AI work begins. When we scope a project, the first estimate we write is for the data work, because it sets the floor for everything else.

The model approach

There are four broad ways to build, and each step up costs more to build and maintain:

  1. Call a hosted model through its API with well-written prompts. Lowest build cost. Run cost follows usage.
  2. Add retrieval (RAG) so the model answers from your own documents. This adds indexing, a vector store and a pipeline to keep the index current.
  3. Fine-tune a model on your examples. This adds training data preparation, training runs and a way to compare each new version with the last.
  4. Train a custom model. Business software rarely needs this. It makes sense for specific vision or prediction tasks with large proprietary datasets.

Most of the projects we see land on the first two. Start at the lowest step that meets the accuracy target, and move up only when the evaluation shows you need to.

Four model approaches from hosted API to custom model, with what each adds to build and run cost

Integrations and tool access

An AI feature that only answers questions is one thing. An agent that calls your APIs, such as creating a ticket, refunding an order or updating a record, is closer to a new integration layer, with permissions, retries, logging and a human approval step for risky actions. Each system the AI must read from or write to adds design, testing and security work. In agent projects, integration effort often outweighs the model work. Our page on AI agent development services covers how we scope tool access.

Evaluation and security

A model that "seems fine" in a demo has not been tested. You need an evaluation set: real questions with known good answers, documents with known fields, conversations with known intents. Writing it takes your domain experts' time, and that time is a real cost on your side as well as the vendor's.

Security adds its own work once the system can see internal data or act on it. The OWASP Top 10 for LLM Applications lists the main risks, starting with prompt injection and sensitive information disclosure. For regulated buyers, the NIST AI Risk Management Framework gives a shared vocabulary (Govern, Map, Measure, Manage) for the controls an auditor will ask about. A system that talks to customers needs both, and both belong in the build estimate.

What drives the run cost

For systems built on hosted models, the main run cost is tokens. OpenAI's API pricing shows the usual structure: prices are per million tokens, output tokens cost several times more than input tokens, and cached input costs a small fraction of the normal input price. Three things follow for anyone budgeting a system:

  • Long prompts cost money on every call. A retrieval system that stuffs twenty documents into each prompt pays for all twenty, every time.
  • Verbose answers cost more than short ones, because output is the expensive side.
  • You can cache repeated context, such as the same system instructions on every call, and at volume that changes the bill noticeably.

Model choice matters too. A smaller model often handles classification or extraction as well as a large one at a fraction of the price, so many production systems route easy requests to a cheap model and hard ones to an expensive one.

Around the model sit the usual running costs of software: hosting, a vector database, logs, monitoring and backups. They behave like the operating cost of any system, which we compare for in-house and offshore teams in software maintenance cost: in-house vs offshore.

The team: where the long-term money goes

People are the largest cost in most AI projects, during the build and after it. Where the team works therefore changes the total more than any choice of model.

Here is what public salary data shows (market data, salaries only):

  • In the US, the median annual wage for data scientists was $120,230 in May 2025, and for software developers $135,980, according to the Bureau of Labor Statistics (data scientists, software developers).
  • In Vietnam, the ITviec IT Salary and Recruitment Market Report 2025-2026 puts data analysts and data scientists at 40.65 million VND a month, and data engineers at 56.9 million VND at the three-to-four-year level. The survey covered 1,839 respondents in 2025.

At BIDV's selling rate of 25,800 VND per US dollar on 11 November 2026 (VietnamPlus), 40.65 million VND a month comes to about $18,900 a year. That is roughly one sixth of the US median for data scientists.

Market salary data: US median wages for data scientists and software developers (BLS, May 2025) next to Vietnam data roles (ITviec 2025-2026)

Read that ratio with care. It compares salaries, and a vendor's rate covers far more than salary: benefits, project management, QA, infrastructure, office, recruiting, bench time and margin. A US company hiring in-house also pays more than salary, plus the recruiting time, which is long for AI roles in most markets. Our article on hiring AI engineers in Vietnam covers that side, and software development cost in Vietnam vs the US, UK and Australia works through the full cost of a team. The salary gap is still the main reason the same AI system costs less to build and maintain with an offshore team.

How to read an AI development quote

When a quote reaches you, from us or anyone else, these are the parts worth reading before the total:

  • The data assumptions. Which data does the estimate assume is ready, and in what form? Every wrong assumption here becomes a change request.
  • The model approach. Does the quote say which approach it uses and why? Fine-tuning where retrieval would do is expensive. Prompts alone, when the task needs your data, will miss the accuracy target.
  • The evaluation plan. Who writes the evaluation set, how big is it, and what score counts as done?
  • Integrations. Which systems does the AI read from and write to, and are permissions and audit logs included?
  • Run cost estimate. Does the quote estimate monthly token and hosting spend at your expected traffic, or stop at launch?
  • After launch. Who re-runs the evaluation when the model version changes, and how is that work billed?

A quote that answers all six is easier to trust than a lower one that answers two.

How we keep AI costs in check at MOHA

Our AI team works from Hanoi, and a few habits from shipped projects keep budgets honest.

Retrieval comes before generation. Our Test-case AI pulls from more than a hundred ISTQB-aligned test templates before the model writes anything, so the output stays tied to approved material and the prompts stay short.

People handle the hard cases. The support chatbot we built for StarHub classifies each message and passes it to a human agent when needed, so the model does not have to be perfect to be useful.

Small jobs stay small. Adding generative AI to the DIG-AgriTech mobile app, to cut cooperative members' paperwork, ran from September to October 2025 at 4.66 man-months, built with NodeJS, Dify, n8n and LangChain.

If you need an AI feature built, or an existing one brought under control, that is the work of our AI development services team.

FAQ

Is it cheaper to use an API model than to train our own?

For most business use cases, yes. An API model has almost no build cost for the model itself, and you pay as you use it. Training your own makes sense when you have a large, specific dataset and a task that general models handle poorly, or when data cannot leave your environment. Even then, an open model hosted in your own cloud usually costs less than training from scratch.

Why does my AI system cost more each month after launch?

Usually because people use it more, prompts have grown, or answers have become longer. Check the average tokens per request and the share of traffic going to the most expensive model. Caching repeated context and routing simple requests to a smaller model are the two fixes that change the bill most.

How do we budget for an AI project when the accuracy is uncertain?

Fund a short first stage that builds the smallest working version on your real data and measures it against an evaluation set. The result tells you which model approach the full build needs, and that makes the rest of the estimate reliable.

Talk to us about your AI project

If you know what the AI should do and which data it needs, you have enough to start a useful conversation about cost. Contact us to discuss your project.

Contact us

Emily
Related Articles
AI Digital Transformation Software
IT Outsourcing Offshore Development Software
We got your back! Share your idea with us and get a free quote