Kuala Lumpur · serving Malaysia & Thailand
AI Adoption and Generative AI Consulting
Aqvantiq takes generative AI from pilot to production for organizations in Malaysia and Thailand. We design the pipeline around the model — data access, evaluation, guardrails, review, observability and cost per request — on AWS, Google Cloud, Huawei Cloud or Azure. We are a multi-cloud AI system integrator in Kuala Lumpur, so this is an engineering engagement, not a workshop.
Why AI pilots stall before production
Because the hard part is not the model. Pilots stall on what a demo never needs: a definition of correct, a real data path, identity and residency decisions, cost per outcome, an owner after go-live. Malaysian adoption figures show the shape — broad uptake, shallow usage.
Research commissioned by AWS and conducted by Strand Partners, reported by IT Brief Asia on 19 August 2026, put AI adoption among Malaysian businesses at 38%, up from 27% a year earlier. But 67% of those adopters stay on basic applications, and only 19% of them report a formal strategy for scaling AI across functions.
No definition of correct
Judged on demos. Without a labelled set and a threshold, nobody can approve it.
The data was hand-fed
Documents dropped into a folder by hand. Production needs a real path to the source.
Identity and data boundaries unresolved
Once the workflow touches customer or employee data, someone must answer what leaves the network.
Cost per outcome unmeasured
Token spend is invisible in a pilot and material at volume. Price one document.
No owner after go-live
Models, prompts and vendors change. Somebody owns regression testing and rollback, or it degrades.
No rollback path
The manual process was switched off too early. Every workflow keeps a way back.
Inside a production-grade generative AI pipeline
It looks like any other production system, with two additions: an evaluation harness and a review layer. Data access, prompts in version control, guardrails, tracing, cost attribution and rollback are the disciplines you already apply to services — applied to a probabilistic component.
Data access and retrieval
A real connection to the source system, permissions preserved. Indexing and refresh automated.
Prompts in version control
Prompts, model IDs and thresholds in the repository. A prompt change is a commit.
An evaluation harness
A labelled set from your own data, run on every change, reporting failure modes.
Guardrails and data handling
Input and output filtering, PII treatment decided explicitly, a written answer on retention.
Observability and cost per request
Traces per call, token cost attributed per feature, alerts on error rate and spend.
Fallbacks and rollback
Rate limits and outages handled deliberately: retry, degrade, or fall back to manual.
Where human-in-the-loop review belongs
Above a confidence threshold you set, on the outputs that carry consequence. High-confidence results flow straight through; everything below lands in a review queue with the source document beside it. Reviewer decisions become evaluation data and move the threshold as accuracy earns it.
- A threshold per output type, not one global setting.
- A review interface showing the output beside its source.
- An audit trail: input, model and prompt version, output, reviewer, timestamp.
- Corrections captured as labelled data, feeding the evaluation set.
- Reporting on what reviewers change most often.
- A measured review rate, so automation reduces effort rather than moving it.
Built exactly this way for a regional SI partner: extractions land in the spreadsheet format the client already used, below-threshold rows flag themselves for review beside a link to the source document, and corrections sync straight back — read the BillOps case study.
How AI adoption connects to your cloud platform
Tightly. An AI feature inherits your identity model, network egress rules, data classification and cost accounting. Teams that treat it as a separate initiative rebuild all four badly, then discover at the security review that the pilot cannot be promoted.
Decided up front
- Which provider processes which data class, in which region, with what retention.
- How the service authenticates: workload identity, not a key.
- Personal data handling under the Personal Data Protection Act 2010, with your privacy owner.
Boundaries come from cloud architecture, the shipping mechanism from DevOps as a Service, the integration surface from application modernization. Most stalled AI projects are blocked by a missing API.
How an AI adoption engagement runs
One workflow at a time, with a number at the end of it. You get a working system in production and evidence about whether to widen it, either way. We would rather deliver one automated workflow than a strategy nobody can execute.
- 1
Choose and measure
One workflow, quantified before-state.
- 2
Build the evaluation set
Labelled examples from your data.
- 3
Ship a thin slice
Production, behind a flag, human review.
- 4
Widen or stop
Compare on real volume, then decide.
What we will not do
- Quote an accuracy figure before seeing your data.
- Train a foundation model when a hosted model plus retrieval will do.
- Ship an AI workflow without a rollback to the manual process.
- Run a pilot with no evaluation set — that is a demo.
Delivery is managed by Anson Lau, the founder — a former Silverlake Axis cloud DevOps engineer — and done by the same group of engineers, with the engineer who scoped the work staying accountable for it.
Frequently asked questions
Can't find your answer?
Ask it directly and an engineer answers, usually the same working day.
Ask an engineer Why do AI pilots stall before production?
Because the hard part is not the model. Pilots stall on what a demo never needs: a definition of correct, a real data path, identity and residency decisions, cost per outcome, and an owner after go-live. Malaysian adoption figures show the shape — broad uptake, shallow usage: 38% of businesses use at least one AI tool, but only 19% of those adopters report a formal strategy for scaling it.
Do you build AI models, or integrate existing ones?
We integrate. Aqvantiq builds the pipeline around foundation models offered by the cloud platforms we already operate — Amazon Bedrock, Google Vertex AI, Huawei ModelArts — plus commercial APIs where they fit. Training a model from scratch is rarely the right answer for a business workflow, and we will say so rather than sell the more expensive project.
How do you keep a generative AI system accurate enough to trust?
With an evaluation harness built from your data and a human-in-the-loop review layer above a confidence threshold. Every change to a prompt, model or retrieval step is scored against the labelled set before it ships, and reviewer corrections feed back into that set. We do not publish an accuracy figure before seeing your data — anyone who does is guessing.
Where does our data go?
Wherever you decide, and it is written down before anything is built: which provider processes what, in which region, what is retained, and what never leaves your network. For workloads with residency constraints we keep the processing inside a cloud region you have approved — see cloud architecture for how those boundaries are designed and evidenced.
Can we change model providers later?
Yes, if the integration is built for it. We put the model call behind an interface in your code, so switching providers is a contained integration change rather than a rewrite. Amazon Bedrock, Google Vertex AI and Huawei ModelArts all give managed model access inside a cloud account you control, which keeps identity, networking and billing in one place.
Does human review ever go away?
Not where the output carries consequence. The threshold moves as measured accuracy earns it, so the share of items a person sees falls, and reviewer corrections feed the evaluation set that gates the next change. In regulated and finance-adjacent workflows human review is the control that makes automation approvable at all.
What is a realistic first project?
One workflow with a measurable before-state and a tolerant failure mode — document extraction, classification, triage, internal knowledge retrieval. We instrument the current process first, then ship a thin slice to production behind a flag with human review. That is a far better use of a quarter than a broad AI strategy nobody can execute.
Do you deliver AI adoption work in Thailand?
Yes. Aqvantiq delivers across Kuala Lumpur and Bangkok, and AI work sits on the same cloud platforms we build and run for Thai enterprises. Start at cloud consulting in Thailand.
Related services
Rolling the same workflow out to a Thai business unit? Cloud consulting and system integration in Thailand covers Kuala Lumpur and Bangkok.
For a pilot that will not reach production
Describe the workflow and where it is stuck — accuracy, data access, security review, cost — and you will get a direct answer. If the honest answer is that it should not be automated at all, that is exactly what you will hear.
Contact Aqvantiq