For most of 2024 and 2025, the enterprise conversation about AI agents centred on capability: what agents could automate, how many calls or tickets they could absorb, how quickly a pilot could be stood up. In 2026, a second conversation has caught up with the first, and it is about what happens when those agents are let loose on real customers, real systems, and real data.
The numbers have moved quickly. A Cloud Security Alliance survey released in April 2026 found that 82% of enterprises had AI agents running in their environments that IT did not know about, and 65% had experienced an AI agent-related incident in the previous twelve months. Among those incidents, 61% involved data exposure, 43% operational disruption, and 35% financial losses [1]. A VentureBeat survey in September found that only 14% of enterprises could shut off a misbehaving agent within minutes [2].
These findings extend a pattern IBM first quantified in its 2025 Cost of a Data Breach Report. Thirteen percent of organizations reported breaches of AI models or applications, and of those, 97% said they lacked proper AI access controls [3]. Organizations with high levels of shadow AI, meaning employees using unapproved tools, paid an average of USD 670,000 more per breach [4].
Analysts have drawn the commercial conclusion. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls [5]. MIT’s NANDA initiative, studying generative AI deployments in 2025, found that 95% of organizations saw no measurable return, and traced the gap less to model quality than to systems that do not retain feedback, adapt to context, or improve over time [6].
Taken together, these studies describe one problem from two sides. Enterprises that deploy agents without supervision take on risks they cannot see, and enterprises that deploy agents without a learning loop get tools that stop improving the day they go live. Both failures point to the same missing piece: people who stay responsible for the agent after launch.
Most of the work happens after launch
WIZ.AI, the Singapore-headquartered voice AI agent company, has built its operating model around that missing piece. Its recent messaging on LinkedIn puts it plainly: going live is only one part of the process. Once a solution is deployed, real customer interactions surface situations testing did not cover, and reviewing what works and what does not is how the solution gets refined over time [7].
That framing sounds modest, but it inverts how many AI projects are scoped. A typical pilot treats deployment as the finish line and measures success at launch. WIZ.AI treats deployment as the point where the most useful data starts arriving. Its stated vision is to build AI agents that “stay reliable in every conversation” and improve daily, refined from each interaction without disrupting what already works [8].
The scale of that feedback loop matters. WIZ.AI reports serving more than 300 enterprise clients across 17 countries and handling over 100 million AI calls a month in more than 16 languages, including Singlish, Bahasa, Thai, Tagalog, and Vietnamese [9]. At that volume, edge cases stop being edge cases. An accent the model mishears, a regulatory disclosure phrased slightly wrong, or a customer who asks for something outside the script will appear thousands of times a day. Whether those moments get caught, reviewed, and fed back into the system determines whether the agent improves or quietly accumulates errors.
Bounding the problem before building the agent
The human-led approach begins before any agent is built. In another recent post, WIZ.AI argued that good AI starts with good questions: what challenges customers are actually facing, where AI can make a meaningful difference, and how a business will know whether it is delivering value. Those questions, the company says, shape its approach from identifying needs through to measuring outcomes [10].
This is also where the risk conversation starts. Singapore’s Model AI Governance Framework for Agentic AI, launched by the Infocomm Media Development Authority in January 2026 as the first framework of its kind, recommends that organizations assess and bound risks upfront by selecting appropriate use cases and limiting how much autonomy a system has [11]. An agent scoped to send payment reminders, confirm appointments, or answer a defined set of inbound queries has a far smaller blast radius than one given open-ended access to a customer database.
WIZ.AI’s product lines reflect that kind of bounding. Talkbot handles outbound calls such as reminders and telesales, Wizlynn manages inbound conversations as a multi-agent service, and a separate Quality Management product reviews calls for compliance [8]. Each has a defined job. The company has also moved beyond voice agents into helping enterprises identify where AI adds value and rework workflows around it, with the stated aim of moving clients from experimenting with AI to working with it [8].
Where humans sit in the loop
The IMDA framework’s second recommendation is to define checkpoints where human approval is required, particularly for irreversible actions, and it warns specifically about automation bias, the tendency to over-trust a system that has performed reliably in the past [12]. That last point is easy to overlook. The more dependable an agent becomes, the more tempting it is to stop watching it, which is precisely when an unusual failure does the most damage.
WIZ.AI’s design places people at three points. First, at escalation: Wizlynn’s routing hands complex cases off to live agents rather than forcing the AI to resolve everything [9]. Second, at review: the Quality Management product offers 100% QA coverage of calls, replacing the sample-based audits that contact centres have traditionally relied on [9]. Third, at refinement: the post-launch review cycle described above, in which WIZ.AI’s teams and its clients study what real conversations reveal and adjust accordingly [7].
The underlying infrastructure carries the certifications regulated buyers expect, including SOC 2 Type II, PCI DSS, and ISO 27001, 27017, and 27018 [9]. Certifications on their own do not prevent an agent from saying the wrong thing to a customer, which is why the review layer matters. But they address the access-control gap that IBM identified as the common thread in AI-related breaches [3], and they are the entry ticket to the banks, insurers, and telcos that make up WIZ.AI’s core market.
The company’s own tagline from its appearance at the Contact Centre Association of Singapore’s Regional CX Symposium in November 2026 summarises the stance: AI-powered, human-led, on the view that customer experience needs both [13].
Why regulated buyers are leaning in
The clearest signal that this approach resonates with risk-sensitive enterprises is who has chosen to back it. WIZ.AI’s Series B, announced in November 2025, was led by SMBC Asia Rising Fund, the corporate venture arm of Sumitomo Mitsui Banking Corporation, with Kasikorn Bank’s Beacon Venture Capital and SMIC SG Holdings joining existing investors Singtel Innov8 and Granite Asia [14]. Several of these investors are also users. SMBC said it plans to use the solution across Asia Pacific, and Singtel Innov8 said WIZ.AI’s voice AI had shown good outcomes across use cases within the Singtel Group [14]. Founder and CEO Jianfeng Lu said the funding would go toward delivering “even more secure, scalable, and human-like AI solutions” [14].
Banks and telcos are among the most heavily regulated and most heavily audited deployers of customer-facing AI. They answer to central banks, data protection authorities, and telecom regulators, and a single badly handled call can become a compliance matter. When institutions in that position commit both capital and production traffic, it suggests the supervised model is clearing a bar that more autonomous approaches struggle with.
The commercial results are consistent with that. WIZ.AI reported revenue growth of more than 100% over 2024 and 2025 [14]. One example is Monee (formerly SeaMoney), which saw product activation rates rise 40 to 50% after deploying WIZ.AI’s agents as its user base grew from one million to 15 million [9]. The case is one instance of a broader point: in regulated customer engagement, the deployments that scale are the ones that operators trust enough to keep expanding.
Localization as a risk control
There is a further, less obvious way WIZ.AI’s model reduces risk. Much of the agent-security discussion focuses on data leaks and unauthorized actions, but in voice AI a large share of failures begin with something simpler: the agent misunderstands what the customer said. In Southeast Asia, where a single call can mix English, Malay, and Hokkien, or Tagalog and English, a model tuned on mainstream accents will misread intent more often, and every misread is a chance to give wrong information, mishandle a request, or trigger a complaint.
WIZ.AI has spent more than six years building localized voice AI for banking, insurance, and telecoms across the region [15], and Lu has described the company’s strategy as “local-first” [14]. SMBC’s Keiji Matsunaga pointed to the system’s ability to adapt to diverse local languages, regional nuances, and accents as part of the bank’s rationale [14]. Accurate understanding is a quality feature, but in a regulated call it is also a control: fewer misunderstandings mean fewer incorrect disclosures and fewer escalations that should have happened and did not.
Trust compounds
The risks now being documented across enterprise AI are unlikely to shrink. As agents gain access to more systems and take on more consequential tasks, the surface area for data exposure and unintended action grows with them. The IMDA framework’s emphasis that humans remain ultimately accountable for what AI agents do [11] is likely to become the baseline expectation for regulators in the region, and enterprises will increasingly ask vendors to show how that accountability works in practice.
That shift favours providers who have treated oversight as part of the product from the start. A deployment model built on bounded use cases, human escalation, full review coverage, and continuous refinement costs more effort upfront than switching on an autonomous agent and walking away. It is also the model that produces an auditable record of what the agent did, a team that knows why, and a system that is measurably better each month than it was the month before. As more of the hidden costs of unsupervised AI come to light, that record is what enterprises will be paying for.
References
[1] Cloud Security Alliance and Token Security, “Autonomous but Not Controlled: AI Agent Incidents Now Common in Enterprises,” Cloud Security Alliance, April 20, 2026. https://cloudsecurityalliance.org/artifacts/autonomous-but-not-controlled-ai-agent-incidents-now-common-in-enterprises
[2] VentureBeat, “VB Pulse Agentic Security and Identity Report,” VentureBeat, September 2026. https://venturebeat.com/vb-pulse-agentic-security-and-identity-report-september-2026
[3] IBM, “IBM Report: 13% of Organizations Reported Breaches of AI Models or Applications, 97% of Which Reported Lacking Proper AI Access Controls,” IBM Newsroom, July 30, 2025. https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls
[4] IBM, “2025 Cost of a Data Breach Report: Navigating the AI Rush Without Sidelining Security,” IBM Think, 2025. https://www.ibm.com/think/x-force/2025-cost-of-a-data-breach-navigating-ai
[5] Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” Gartner Newsroom, June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
[6] MIT NANDA, “The GenAI Divide: State of AI in Business 2025,” as reported by Storyboard18, September 29, 2025. https://www.storyboard18.com/amp/digital/study-finds-95-of-companies-see-no-return-on-ai-investments-due-to-learning-gap-81696.htm
[7] WIZ.AI, “Going live is only one part of the process,” LinkedIn, November 2026. https://www.linkedin.com/posts/wiz-ai_wizai-enterpriseai-aiimplementation-activity-7526927747370110977-_AAc
[8] WIZ.AI, “Becoming AI-native, and helping enterprises do the same,” WIZ.AI. https://www.wiz.ai/pages/locations.html
[9] WIZ.AI, company website (platform metrics, product descriptions, certifications, and Monee case study). https://www.wiz.ai/
[10] WIZ.AI, “Good AI Starts with Good Questions,” LinkedIn, November 2026. https://www.linkedin.com/posts/wiz-ai_wizai-enterpriseai-activity-7522922322152964097-VD73
[11] Fintech News Singapore, “Singapore Launches Model AI Governance Framework for Agentic AI,” Fintech News Singapore, January 2026. https://fintechnews.sg/125071/ai/singapore-agentic-ai-framework/
[12] Allen & Gledhill, “Singapore Launches New Model AI Governance Framework for Agentic AI,” Allen & Gledhill, February 13, 2026. https://www.allenandgledhill.com/sg/publication/articles/32314/launches-new-model-ai-governance-framework-for-agentic-ai
[13] WIZ.AI, “AI-powered. Human-led. The future of customer experience needs both,” LinkedIn, November 2026. https://www.linkedin.com/posts/wiz-ai_ccas-customerexperience-voiceai-activity-7524348118423732225-O2Wu
[14] TechNode Global, “Singapore’s WIZ.AI Raises Tens of Millions in Series B to Scale Enterprise AGI Solution Globally,” TechNode Global, November 5, 2025. https://technode.global/2025/11/05/singapores-wiz-ai-raises-tens-of-millions-in-series-b-to-scale-enterprise-agi-solution-globally/
[15] Agora and WIZ.AI, “Agora and WIZ.AI Partner to Deliver Enterprise-Ready AI Agent Solutions,” PR Newswire via AAP, June 18, 2025. https://www.aap.com.au/aapreleases/cision20250618ae12745
Paulo Joquiño is a writer and content producer for tech companies, and co-author of the book Navigating ASEANnovation. He is currently Editor of Insignia Business Review, the official publication of Insignia Ventures Partners, and senior content strategist for the venture capital firm, where he started right after graduation. As a university student, he took up multiple work opportunities in content and marketing for startups in Asia. These included interning as an associate at G3 Partners, a Seoul-based marketing agency for tech startups, running tech community engagements at coworking space and business community, ASPACE Philippines, and interning at workspace marketplace FlySpaces. He graduated with a BS Management Engineering at Ateneo de Manila University in 2019.