Most companies are making the wrong AI mistake. It is not that they are moving too slowly; it is that they confuse “can chat” with “can deliver.” An AI agent can look impressive in a demo—searching, summarizing, drafting emails, even planning tasks. That does not mean it can reliably complete work end-to-end inside a real business. Logging into tools is not the same as finishing the job.
My blunt view for Asia-Pacific businesses is this: over the next 12 to 24 months, AI agents will not first replace people. They will first expose which firms have weak process discipline, poor data governance, and messy permissions. Gartner said in 2024 that more than 30% of generative AI projects will be abandoned after proof of concept, citing poor data quality, inadequate risk controls, and unclear business value. That is the real message. In most failed deployments, the bottleneck is not model intelligence. The bottleneck is that the company was never operationally ready for AI to act.
I use a practical framework to judge whether an agent initiative is worth funding: the Three Gates model—Can it connect? Can it be controlled? Can it be counted? Gate one is integration: can the agent connect to CRM (customer relationship management), ERP (enterprise resource planning), email, document repositories, and internal knowledge? Gate two is governance: can you restrict permissions, preserve logs, trace decisions, and insert human approval? Only gate three is financial: can you translate the deployment into labor hours saved, conversion uplift, faster response times, or fewer errors—into actual ROI (return on investment)? Most projects fail before they ever reach gate three.
Is your agent actually doing work—or just performing intelligence?
A lot of what the market calls “agents” today is really a polished layer over search, summarization, writing, and simple workflow automation. That is not useless. It is just a very different value proposition. If the agent only tells your staff what to do next, it is acting as an advisor. If it can update a ticket, assemble a quote, compare contract clauses, send a follow-up, and record the action back into a system, it starts to behave like execution capacity.
McKinsey estimated in 2023 that generative AI could affect activities that account for 60% to 70% of employees’ time in knowledge work. The overlooked word is activities—not whole jobs. Executives do not pay for an AI that saves eight minutes here and there. They pay for an AI that reliably connects three to five steps in an end-to-end process. That matters even more for SMEs in Hong Kong, Taiwan, and Southeast Asia, where small teams often rely on tribal knowledge rather than formal process design. Agents run into that wall immediately.
In real deployments, we repeatedly see the strongest early results in customer support, internal IT service, and sales follow-up—not because those functions are the most advanced, but because the task boundaries are clear. For example: classify a customer issue, check the FAQ, pull CRM status, draft a reply, then route to a human for approval. That is manageable. “Build me a fully autonomous business assistant” is usually not.
Gate One: If it cannot connect to your systems, it is just a chatbot
The real dividing line in enterprise agents is not model leaderboard performance. It is integration. Stanford’s 2024 AI Index highlighted data and application integration as a key barrier to enterprise adoption of generative AI. In Asia-Pacific, this challenge is often worse than in the US: many firms operate across SaaS subscriptions, Excel files, WhatsApp threads, legacy ERP, local storage, and department-owned folders. The problem is not lack of data. It is fragmentation, duplication, and version confusion.
That is why I do not advise mid-sized businesses to begin by chasing the “best model.” Start by identifying the minimum three systems that must be connected before value can happen. Usually the order is: 1) knowledge source, 2) transaction record, 3) action channel. In plain terms: the agent must first find the right information, then understand the live business state, and only then take action.
| Product / Approach | Approx. Pricing | Positioning | Strengths | Weaknesses / Risks | Best Fit |
|---|---|---|---|---|---|
| Microsoft Copilot for Microsoft 365 | About US$30/user/month | Productivity and collaboration agent in the Microsoft stack | Deep integration with Outlook, Teams, Word, Excel; relatively low deployment friction | Non-Microsoft actions still need extra integration; may remain stuck at personal productivity | Firms already standardized on Microsoft 365 |
| Salesforce Einstein Copilot / Agentforce | Enterprise pricing, usage-based | CRM-centric sales and service agent | Strong customer data and workflow integration; good for front-office automation | Higher cost and implementation complexity; value depends on CRM data quality | Mid-size and large firms with standardized customer processes |
| OpenAI API + custom-built agent | Token-based model usage plus development cost | Flexible bespoke approach | Deep customization across process, language, and tools; strong for cross-system orchestration | Governance, monitoring, permissions, and evaluation must be built internally | Companies with technical teams seeking differentiated workflows |
| Google Workspace + Gemini for Workspace | Tens of dollars per user/month depending on tier | Document, email, and meeting assistant | Strong drafting, summarization, meeting notes within Google ecosystem | Limited action depth across non-Google business systems | Service firms already running on Google Workspace |
The point is not that one vendor is universally superior. The point is that your core work already lives somewhere. If customer records, orders, knowledge, and approvals sit across four different systems, any agent platform is just the beginning—not the answer.
Gate Two: Without governance, labor savings quickly become operational risk
Companies that overestimate agent intelligence usually underestimate governance at the same time. Deloitte’s 2024 enterprise AI surveys consistently found that data privacy, hallucination (confidently wrong outputs), regulatory exposure, and brand risk remain top executive concerns. That is not legal being overly cautious. Once an agent can take action, the risk moves from “bad content” to “bad process” and “bad transaction.”
This is especially important in Greater China and cross-border contexts. A Hong Kong company may handle mainland China and international customer data in parallel. A Taiwan manufacturer may be dealing with supplier documents, export compliance, and sensitive IP. A Singapore regional HQ may be orchestrating workflows across multiple jurisdictions. If you do not separate permissions into read, recommend, simulate, execute, and submit, your agent will end up either too weak to matter or too dangerous to trust.
In practice, I recommend a four-level permission ladder: level one can only retrieve; level two can draft; level three can simulate actions in a sandbox (an isolated test environment); level four can write back to production systems only after human approval. Many companies rush toward “full autonomy” and then freeze the whole program after one wrong email, one bad order, or one incorrect price update. Mature deployment is not about removing humans. It is about placing humans at the points where risk is highest and judgment matters most.
Gate Three: If you cannot calculate ROI, your agent becomes another IT cost
IDC has repeatedly shown that global AI spending is growing rapidly and shifting from experimentation toward use cases and business outcomes. Translated into boardroom language: leadership no longer asks whether you are using AI. They ask how much more revenue, efficiency, or resilience it creates.
Do not fund agents with vague claims about “transforming the enterprise.” Build the business case around one process. The cleanest ROI cases usually come from three buckets. First, labor savings: for example, average handle time in customer service falls by 20% to 40%. Second, revenue uplift: faster sales follow-up improves lead conversion. Third, error and leakage reduction: fewer quoting mistakes, fewer missed customer responses, fewer document-version errors. Platform vendors and Forrester-style Total Economic Impact studies often showcase strong results, but executives should ask a harder question: can that outcome be replicated inside our process, with our data quality and our operating habits?
APAC SMEs in particular need financial discipline. If your annual AI budget is smaller than the total cost of one experienced operations manager, do not pretend you can run ten agent initiatives at once. Pick one high-frequency, labor-intensive, measurable process with manageable downside risk. Prove the first return. Then reuse the data structures, prompts, permissions, and monitoring patterns elsewhere.
Are multi-agent systems always better? Usually, no
The market loves the idea of multi-agent systems—different agents specializing in research, planning, execution, and review. In some complex tasks, that architecture is genuinely useful, especially for heavy analysis, long decision chains, or software engineering. But that does not mean it is the right starting point for enterprise operations.
The more agents you add, the longer the error chain becomes, the harder accountability gets, and the more token cost and latency can rise. a16z and many developer communities noted throughout 2024 that in production, observability (the ability to see what the system is doing), stability, and evaluation matter more than making the system “feel more human.” In plain business terms: a single-agent workflow that reliably triages refund requests, fills in missing fields, drafts a response, and routes exceptions is often more valuable than a flashy multi-agent architecture that no one can control.
This is not an argument against advanced architectures. It is an argument about sequence. Prove the task with a simpler design first. Add multi-agent complexity only when it demonstrably improves coverage, precision, or throughput.
Key takeaways: pass the three gates before you talk about autonomy
If you are an owner or executive, use the Three Gates model as a funding filter. Gate one: Can it connect? At minimum, connect one knowledge source, one business system, and one action channel. Gate two: Can it be controlled? You need permission tiers, logs, human approval, and exception handling. Gate three: Can it be counted? Track two or three hard metrics within 90 days—labor hours, response speed, conversion rate, error rate, or churn.
The companies most likely to win are not the ones that buy tools earliest. They are the ones that standardize process earliest, clean their data earlier, and define responsibility boundaries clearly. The core of an AI agent is not that it acts like a person. It is that it can complete a bounded piece of work more predictably than a person.
Self-check
- Are we trying to automate a clearly defined process—or a vaguely defined role?
- If the agent makes one bad decision tomorrow, what is the worst operational, compliance, or reputational damage?
- Can we prove within 90 days that it improves at least one core KPI (key performance indicator) with real numbers?
