The AI agency space in Southeast Asia right now looks a lot like the digital marketing agency space did in 2014.
Lots of entrants. Most rebranded from something adjacent (marketing agencies, consulting firms, software houses). A few are actually good. Most are selling confidence they can't deliver on.
If you're a business owner evaluating AI partners, you don't need a deck comparing five agencies. You need a sharp filter.
Here it is. Ten questions you can ask in a single 30-minute discovery call. If an agency can't answer them clearly, move on.
1. "Show me a live AI system you built for a client. Not a case study deck — the actual thing."
This is the first question, and it filters out 60% of agencies immediately.
A real AI agency will walk you through a working system. They'll show you the actual WhatsApp conversations a customer's AI is handling. They'll show you the dashboard, the hand-off logic, the metrics. You'll see outputs, not concepts.
A fake AI agency will show you:
- A slide deck with case study logos
- Screenshots of ChatGPT
- Generic "AI capability" diagrams
- A whitepaper on AI trends in SEA
If the answer to "show me the thing" is anything other than "sure, let me open my laptop" — they haven't built one.
2. "Who on your team will actually build this?"
Most agencies put their best salesperson in the first meeting. That person often can't build anything. The "AI specialist" you're talking to might be a strategist whose job ends after the proposal is signed.
The question is simple: who, specifically, will write the prompts, wire up the integrations, and maintain the system? Ask for names. Ask to meet them.
Red flags:
- "We have a team of AI engineers" (vague — how many, who)
- "We outsource the building to a partner in Vietnam/India"
- "Our senior consultant manages the delivery"
Green flags:
- "Alex will be your lead — here's their background, here are two systems they've built"
- "You'll have a single point of contact who's hands-on with the build"
- "We can arrange a technical call next week with the engineer"
You're buying the person, not the brand.
3. "How long does your typical project take, from kickoff to an AI that's live with real users?"
This question reveals whether they ship.
Honest answer: 3-8 weeks for a first AI use case. Sometimes longer for complex integrations, but the first working version should be live within 4-6 weeks for most SME use cases.
Bad answers:
- "12-16 weeks" (usually means lots of meetings, few decisions)
- "Depends on the client" (evasive — they don't have a track record to quote from)
- "We do a 3-month discovery phase first" (strategy shop in disguise)
If they can't commit to a rough timeline, they haven't done enough of these to know.
4. "What percentage of messages/tasks does your AI actually handle without human escalation, based on your existing clients?"
This is the gold question for AI agents specifically.
Honest numbers for well-built AI agents:
- Customer support AI: 65-85% handled without human help
- Lead qualification AI: 70-90%
- Document processing AI: 85-95%
Real agencies quote ranges backed by real deployments. They'll tell you "it's 72% for this client, 81% for that one, and here's why they differ."
Red flags:
- "Our AI handles 95%+ of everything" (too high — they're either counting messages where the AI replied but was wrong, or they haven't deployed enough to know)
- "It depends on your use case" (without any reference numbers — they don't have any)
- "We don't measure that" (run away)
5. "How do you handle the cases where AI gets it wrong?"
Every AI system makes mistakes. Good agencies have thought carefully about the failure modes. Bad agencies pretend mistakes don't happen until the first one costs you a customer.
Ask:
- How does the AI know when it's uncertain?
- When does it escalate to a human, and how?
- How do you catch and correct mistakes over time?
- Can the human who takes over see the full context of what the AI did?
Acceptable answer: "The AI has a confidence threshold — if it's not 80% sure of the right response, it tags a human. We review weekly logs to tune prompts and add edge cases to the knowledge base."
Unacceptable answer: "Our AI is very accurate." (Non-answer. They haven't deployed anything real yet.)
6. "What's your pricing structure, and what exactly does it include?"
This filters out agencies who price by feeling rather than cost.
Clear pricing structures look like:
- Setup fee (RM10k-30k): initial build, integration, first knowledge base, testing
- Monthly retainer (RM2k-8k): hosting, maintenance, ongoing tuning, small iterations
- Usage component (optional): API costs passed through at cost
The agency should be able to tell you:
- What changes trigger additional fees vs what's included
- What happens if usage spikes (seasonal businesses especially)
- What happens if you want to cancel — do you own the AI you paid for?
Red flags:
- Custom pricing that can't be broken down
- Vague "enterprise" pricing on day one without understanding your use case
- Inability to estimate monthly cost for a given message/task volume
7. "Who owns the AI, the prompts, and the data at the end of this?"
Most clients don't think to ask this until it's too late.
Some agencies design lock-in on purpose. Their platform, their prompts, their integrations — you can never leave without starting from scratch somewhere else.
Ask:
- If I cancel the retainer, can I export the prompts and system configuration?
- Who owns the customer data the AI is processing?
- Is my AI running on your platform or a major cloud provider I could take over?
- What happens to the AI if your agency shuts down?
Green flags:
- "You own everything — prompts, data, integrations. We'll hand it all over if you leave."
- "We deploy on your cloud account or a shared one you control"
- "Here's our offboarding process"
Red flags:
- "Our platform is proprietary"
- "The prompts are our IP"
- "The integration only works on our infrastructure"
If they won't let you own what you paid for, don't pay them.
8. "What happens in the first 30 days after launch?"
A deployed AI system on day one is not the same system on day 60. It needs tuning — real customer messages reveal gaps, edge cases, and prompt improvements nobody anticipated during the build.
Ask what the first 30 days look like:
- How often are they reviewing the AI's output?
- How do they capture edge cases and improve the system?
- What's the expected accuracy improvement from week 1 to week 4?
- Who handles the tuning — the client or the agency?
Acceptable answer: "We review logs daily for the first two weeks, weekly after. We'll tune prompts, expand the knowledge base, and report on accuracy trends in a bi-weekly call."
Bad answer: "Once it's live, it runs." (Nothing runs untouched. Reality: the AI gets worse over time if it's not maintained.)
9. "Can I talk to two of your current clients?"
This is the reference check. Do it.
A real agency has client relationships they're proud of, and those clients are willing to take a 15-minute call with a prospect.
Questions to ask the references:
- How long has the AI been live?
- How often does it break or need intervention?
- Did the agency deliver on time? On budget?
- Would you hire them again?
- What surprised you — good and bad?
Red flags:
- No references available ("confidentiality" — usually not a real blocker for a short call)
- References that are all C-suite introductions ("my CEO friend") not operators who use the AI daily
- References that sound rehearsed or vague
If the agency can't produce two clients who'll talk, they either don't have many clients or their clients aren't happy.
10. "If we decided to hire you, what's the very first thing you'd build?"
This is the closer. A good agency can answer concretely, even if they've only spent 30 minutes talking to you.
A good answer sounds like:
"Based on what you described, I'd start with a WhatsApp AI agent for your customer enquiries. First 2 weeks: connect to WhatsApp Business, extract your FAQ knowledge from your team, build the initial knowledge base. Week 3-4: internal testing, tune prompts, set up human escalation. Week 5 onwards: live with real customers, daily review and tuning. I'd expect 60-70% autonomous handling in week 6, rising to 80% by week 10. That's the smallest useful thing to ship first."
A bad answer sounds like:
"Well, it depends on the discovery phase. We'd need to do a full audit first. There's a lot of variables. After our assessment we can put together a roadmap with multiple workstreams."
The first answer is someone who builds. The second is someone who sells decks.
The Meta-Point
Reading back through the 10 questions, a pattern emerges.
The filter is concreteness. Real AI agencies talk in specifics — names, weeks, percentages, client examples, tools, costs. Fake AI agencies talk in abstractions — strategies, roadmaps, frameworks, assessments.
You don't need to be technical to run this filter. You don't need to know what a prompt is or how a model works. You just need to listen for whether the agency is describing a thing they've actually done or a process they plan to run on you.
If it's the first, you're talking to a builder. If it's the second, you're talking to a consultant with an AI label.
One More Filter: The "Walk Away" Test
Here's the final question, and it's the one we use ourselves when deciding whether to take on a client.
If the agency never contacts you again after this call, what are you left with?
If the answer is "a deck with a proposal" — you're paying for hope.
If the answer is "a specific understanding of the first thing we'd build, the rough timeline, and what it would cost me" — you're in the right conversation.
A good agency's first call should leave you with the shape of the project even if you never sign the contract. They should be able to tell you what to do — including the honest admission that maybe AI isn't the right answer for your specific use case.
That's the sign of a partner who's trying to solve your problem. Which, ironically, is the behaviour that makes you most want to hire them.
Evaluating AI agencies? Start by asking us these ten questions. Book a free AI audit and we'll walk you through our answers — and give you an honest assessment of whether AI actually fits your business. Sometimes the right answer is "not yet", and we'll tell you that.