Short answer
An AI project has three costs: people to assess and build, usage-based model fees, and support after go-live. People cost most, priced as person-days times a day rate; the Thai government benchmark for IT staff with 7 to 10 years' experience is about THB 6,048 to 7,573 per person-day (PDMO, March 2026). Pilot one measurable task before investing fully.
Key points
- A quote should separate people, model fees and support; if you get one number, ask for person-days per phase.
- Divide each quote back to a day rate and compare it with the PDMO benchmark of THB 6,048 to 7,573.
- Model fees are the only cost that grows with volume every month, so measure them on a pilot before budgeting a year.
- If a person checks every output, train the team; if the task repeats hundreds of times or needs company data, build a system.
- Run four stages (assess, pilot, measure, scale), each ending in a go or stop decision.
Where does the money go in an AI project?
Most of the money pays for people, not for the AI: the people who make it work on your processes: watching the work, connecting to company data, testing, and setting data rules. Model fees, which worry people most, are usually smaller, but they are the only cost that grows with volume every month.
When you read a quote, separate the three. If a vendor gives one number, ask for the person-days per phase and a monthly model-fee estimate at your volume.
| Cost | How it is priced | Reference point |
|---|---|---|
| People: assessment, build, training | Person-days × day rate | Thai government benchmark of about THB 6,048 (7 years) to 7,573 (10 years) per person-day, PDMO March 2026 |
| AI model usage | Per token sent and received | Varies several-fold between large and small models; see per-million-token prices in the model guide below |
| Support after go-live | Monthly, or a monthly block of hours | No public benchmark; scope it in the contract (fixes, model-version updates, periodic output checks) |
AI models and effort levels: how to choose without overpaying
How do you use the government benchmark to read an AI quote?
Use it as one yardstick for the day rate inside every quote. The Public Debt Management Office (PDMO, Ministry of Finance) consultant-fee criteria, March 2026 edition, set a base salary for IT staff with a bachelor's degree of THB 50,400 a month at 7 years and THB 63,110 at 10 years. Multiply by the 2.640 cost factor and divide by 22 working days, and you get about THB 6,048 and 7,573 per person-day.
It is not the private market rate: freelancers usually quote lower, specialists higher. Divide each quote back to a day rate; if one is unusually cheap, ask why the days are fewer. Often testing on real data or the data rules were left out.
Base salary × 2.640 ÷ 22 working days gives about THB 6,048 at 7 years and THB 7,573 at 10 years.
Source: PDMO consultant-fee criteria, March 2026 edition, bachelor's degree holders.
Should you train the team, build a system, or buy a tool?
If a person uses the AI and checks every output, training the team is usually enough. If the task repeats hundreds of times, needs company data, or needs an audit trail, it should be a system. Getting this split right early is where most of the saving is: a system for a task three people do once a week rarely pays back, and training does little for a task done 500 times a day.
- Train: drafting email and documents, summarising reports, research, anything a person reads before using.
- Build: sorting incoming documents, answering customers from your own knowledge base, extracting invoice data into accounting; high-volume work tied to existing systems.
- Buy: work every company does the same way, such as meeting transcripts and notes.
- Not yet: work that must be right every time with no one checking, or that a process change would fix.
What are the four stages of an AI project?
Assess, pilot, measure, scale. Each stage ends in a go or stop decision instead of rolling on automatically, so the large spend happens only after real results. You keep the assessment even if you stop after it, and a pilot that misses its agreed measures is stopped or adjusted, not scaled.
- Assess real workMeasure time and errors; pick the one to three tasks that pay most.
- PilotOne task on real data, small scope, one user group, measures agreed up front.
- MeasureCompare time, errors and model fees at real volume; stop or adjust if it misses.
- ScaleMore users and tasks, support, training, and periodic checks as models change.
What should you get back from an AI consultant?
You should get deliverables that let you carry on without depending on the same vendor forever. At SyncEdge, engineers work with AI assistants and are accountable for what they hand over: the code, the tests and the documentation.
- An assessment: which tasks use AI and which do not, with reasons and numbers.
- Data rules: what may go to outside services, what may not, and how identifying data is removed first, in line with PDPA.
- Source code, prompts and configuration with documentation, with the contract stating clearly who holds what, and model-provider accounts in your company name.
- A test set of real examples with correct answers, re-run whenever the model or system changes.
- A pilot report with before-and-after figures and monthly model fees at real volume.
- User and administrator manuals.
Consultant, freelancer, or an off-the-shelf AI tool?
It depends on how similar your task is to everyone else's and how much company data it touches. For generic work such as meeting transcripts, a packaged tool beats a custom build.
| Option | Fits | Watch for |
|---|---|---|
| Off-the-shelf AI tool | Generic tasks; start today, pay monthly or per use | Your data sits with the provider; read its data terms; limited fit to your process |
| Freelancer | Small, well-defined, self-contained jobs, such as adding a chatbot to a web page | Continuity after handover, model-version updates, documentation |
| Consultant or firm | Work that needs assessment, touches company systems and data, spans departments or needs years of support | Higher price; pay by phase with a point where you can stop |
An off-the-shelf example: SyncEdge Scripta, AI meeting notes and transcripts
Checklist Before Hiring an AI Consultant
- Does the quote separate people, model fees and support?
- Are person-days given per phase, so you can derive a day rate?
- Does the pilot have agreed success measures and a stop point?
- Are model-provider accounts in your company name?
- Does the contract state which source code, prompts, configuration and tests are handed over?
- Is there a written rule on which personal data may go to outside services?
- After go-live, who checks outputs, and who pays when the model version changes?
Frequently asked questions
How much does AI consulting cost in Thailand?
It depends on person-days more than anything. The public benchmark is the PDMO rate of about THB 6,048 to 7,573 per person-day for IT staff with 7 to 10 years' experience (March 2026 edition); private rates vary around it. Add usage-based model fees measured on a pilot, and monthly support if you need it.
Can we just use ChatGPT or a general AI tool ourselves?
Yes, and many tasks should start there. Hiring helps when the work needs company data, runs at high volume, connects to existing systems, or needs customer-data rules the whole organisation follows. Without those conditions, training the team to use the tools well is usually the better spend.
What is an AI solution?
An AI solution puts an AI model to work on one specific business task, rather than just switching a tool on. Most of the work is connecting the model to the company's data and systems, testing on real cases, and setting rules for what data may leave the company. The model is only one part of the system.
How do we estimate monthly model fees?
Multiply monthly volume by the text size per task and the model's price per token. The reliable number comes from a pilot, which measures tokens per task on real work. Prices differ several-fold between model tiers, so simple tasks should not run on the most expensive one.
What if the pilot doesn't work? Is the money wasted?
You lose only the assessment and pilot, which is why payment should follow the stages. What you get back is evidence of why it didn't work and an assessment you can still act on, which is far cheaper than finding out after building the whole system.
Is sending customer data to an AI service a PDPA breach?
Not automatically, but it needs a lawful basis, and the model provider acts as a data processor that needs an agreement in place. Remove identifying data before sending where you can, choose services that do not train on your data, and make it a rule for the whole team.