How to Choose an AI Agency in Malaysia: The Seven Questions That Separate Real Delivery from a Deck

Every AI agency in Malaysia will tell you they can build you an agent. Most of them are telling the truth, in the sense that a demo will exist and it will work in the meeting.

The difficulty is that the things which decide whether the project succeeds — who owns it at 2am, whether it can reach your systems, what happens when the model changes, whether you are locked in — are exactly the things a capability deck does not cover, and exactly the things a buyer has no obvious way to check.

Some of it can be checked, and the strongest check is one most buyers never make. Start there.

1. Ask for the Claude Certified Architect credential, and for proof of partner network membership

Start with two requests, because both are verifiable and most buyers ask for neither.

First, ask to see proof of Claude Partner Network membership — not the badge rendered on a website, which anyone can add to a footer, but the actual standing. We are a member of the Claude Partner Network and we will show you. A firm that hedges on this question is answering it.

Second, and more important, ask who on the team holds the Claude Certified Architect credential — and whether that person will be on your project.

This is the question that actually predicts whether you get a working system, and it is worth explaining why rather than treating it as a badge on a slide.

Anthropic's certifications are role-based and split by what you do with Claude. The Associate track certifies that someone can use Claude well. The Developer and Architect tracks certify that someone can build with it — the API, Claude Code, Agent Skills, and MCP, which are the components an agent is actually assembled from. All of them are assessed through proctored, identity-verified exams administered via Pearson, so unlike almost every other claim in a pitch, this one cannot be self-declared. Either the exam was sat and passed, or it was not.

Here is why that distinction matters more for agents than for most software.

There are two ways to arrive at something that looks like a working AI agent. One is to design it: decide what it is allowed to do, what it must never do, where a human approves, what happens when a tool call fails, how it behaves when the input is malformed or the model is having an off day. The other is to prompt at it until the demo passes, then ship the demo.

Both produce something that works in the meeting. Only one produces something that still works in month four, when a supplier changes their invoice format and nobody notices for a fortnight.

Certification is evidence of the first approach. It does not guarantee a good build — nothing does — but it establishes that the person holds the foundations of agent design rather than assembling it by feel. Vibe coding is a genuinely useful way to explore an idea quickly. It is not a method for putting something into production that your finance team depends on, and the difference between those two things is most of what you are paying an agency for.

So ask, in this order:

• How many people hold Claude Certified Architect?
• Will one of them work on our project, or are they the name in the pitch?
• Who is the most junior person who will touch this, and what do they do when they get stuck?

The third question predicts delivery quality and almost nobody asks it. Agencies staff pitches with their best people and projects with whoever is free.

To answer our own question rather than only posing it: seven of our engineers hold the Claude Certified Architect credential. That is the number worth asking any agency for, because it decides something practical — whether a certified architect is genuinely available to your project, or whether there is one in the company and they are already committed elsewhere. One certified name on a slide and seven across a team are different propositions when your build slips a fortnight.

One warning about the public partner directory, because you may find it and it will point you the wrong way. Anthropic publishes one, and it does show each partner's tier and deployments. But it is built for global systems integrators: the default view is Accenture, Bain, BCG, Capgemini, Cognizant, Deloitte, DXC, Infosys and KPMG, every one at Global Premier. The lowest published Services tier already requires at least 10 certified individuals and 2 joint production deployments, so the thresholds track company size, not fit for your job. Malaysian specialist teams generally do not appear there.

It is the right tool for shortlisting a global integrator on a multi-region rollout, and the wrong tool for choosing a supplier for a project in Klang Valley. Ask for the credential and the named people instead.

2. Ask what happens after go-live, and listen for a specific answer

This is the question that separates agencies fastest, and almost nobody asks it during selection.

An AI system is not a website. It does not sit still. The model gets updated, your data drifts, an edge case arrives that nobody scoped, and an agent that was 94% accurate in testing starts quietly making one particular kind of mistake in week seven.

Ask directly: who watches it, how would we know it had gone wrong, and what does that cost per month?

Answers that should concern you: "it's fully automated", "we hand over documentation", or a support retainer priced as a percentage of build cost with nothing specified about what it covers.

Answers that suggest real production experience: a named monitoring approach, an exception queue with a human in it, a stated review cadence, and a specific person or rotation. It is not glamorous, and it is where the actual cost of running AI lives. We put numbers on all three costs in what an AI agent actually costs to build and run in Malaysia.

3. Ask whether it can reach your systems

A surprising number of "AI projects" in Malaysia end up as a very expensive chat window. The team gets fluent, the demo is good, and then somebody still exports the CSV by hand every Monday because the assistant cannot see the order system.

Making it reach your systems is integration work. The current mechanism is MCP, a protocol that lets a model read from your own databases and tools with your permissions attached. It is a genuinely different engagement from prompt work or training, with a different budget. We explain the mechanics in connecting AI to your business systems.

The question to ask is not "can you integrate with our ERP". Everyone says yes. Ask which specific systems you have connected before, what the authentication looked like, and what broke. The last part is the tell. An agency that has actually done it will have an answer about permissions, rate limits, or a vendor API that behaves badly. An agency that has not will describe the happy path.

This matters enough in Malaysian buying that it shows up directly in what people search: "best ai integration software agency partner in malaysia" is a real query we see.

4. Ask about lock-in before you like them

Buyers worry about this more than agencies acknowledge. One of the highest-intent queries we see in Malaysia is, almost verbatim, whether there are agencies that work without long-term lock-in contracts. Somebody typed that because they had been burned.

Three practical things to settle in writing, before the relationship gets friendly enough that raising them feels awkward:

• Who owns the code, the prompts and the configuration? "We own the IP and licence it to you" is a legitimate model, but it is a different purchase and it should be priced differently.
• Where does it run? If it runs only in the agency's cloud account, on their API keys, you cannot leave without a rebuild. Your own accounts, your own keys, is the cleaner arrangement.
• What does exit look like? Ask for the offboarding process specifically. A firm that has one written down has thought about it; a firm that improvises an answer has not.

None of this requires distrust. It requires the terms to be explicit while everyone is still cheerful.

5. Ask whether they will train your team, and whether that is claimable

An agency that builds you something and leaves is solving half the problem. The other half is that your team has to use it, and adoption is where most Malaysian AI projects quietly stall.

This is another thing buyers search for directly: "ai implementation partner in malaysia that also trains our team". It sits at the top of our results for a reason — it is a real and specific need, and it is not what most agencies sell.

There is a concrete funding angle here that many buyers miss. If you employ ten or more Malaysian employees you are already paying an HRD Corp levy of 1% of wages, and registration is compulsory. Employers with five to nine may register optionally at 0.5%. That money has already left your payroll. Training delivered by an HRD Corp approved provider is how you recover some of it, provided the course runs at least four hours and the grant is approved before it starts.

So "do you train" is only half the question. The other half is "is your training claimable, and will you file the e-TRiS application?" — which needs a quotation, a training schedule and a trainer profile. We set out the detail in our guide to what Claude training costs in Malaysia and which team to send first.

6. Ask for one project that went badly

Ask it plainly: tell me about a project that did not go well, and what you changed afterwards.

The answer tells you two things: whether they have done enough work to have one, and whether they are willing to be straight with you when there is nothing to gain from it. A supplier who cannot name a single failure is either inexperienced or managing you, and you will find out which during delivery rather than during selection.

The related question is scope honesty: what would you tell us not to do? Any agency worth hiring has opinions about which of your ideas is a bad fit for AI. If everything you propose is met with enthusiasm, you are talking to a sales process, not an advisor.

7. Check they are solving your problem, not demonstrating their technology

The last one is the least technical and the most predictive.

A good discovery conversation spends most of its time on your workflow — who does what, how often, where it breaks, what a mistake costs. A weak one spends most of its time on their stack.

The practical test: after the first meeting, could they describe your process back to you accurately enough that one of your own staff would recognise it? If not, whatever gets built will be a solution to a problem nobody at your company actually has.

A note on shortlists

You will probably start with a search, and the results will be a mix of genuine specialists, general digital agencies that added an AI page, and government and directory pages that rank on authority rather than relevance. Treat page one as a starting list, not a ranking of competence — including where we appear on it.

For what it is worth: in the four weeks to 21 August 2026, our own pages drew 424 impressions across 27 Malaysian AI-agency queries and earned zero clicks. We are on page two for most of them. That is a fact about our search performance, not about our delivery, and the same is true in reverse for whoever is currently at the top.

The short version

1. Ask for proof of partner membership and for the Claude Certified Architect on your project.
2. Ask who watches it after go-live, and what that costs monthly.
3. Ask which systems they have connected, and what broke.
4. Settle IP, hosting accounts and exit in writing, early.
5. Ask whether training is included and HRD Corp claimable.
6. Ask for a project that went badly.
7. Check they can describe your workflow back to you.

If you want to run those seven questions at us, that is a conversation we are glad to have — and if the answer is that you need something we do not do well, we would rather say so at the start.