Hiring has become a capital-allocation decision. Funding rewards companies that produce more per head, and AI gives founders new ways to get there: agents that take on repeatable work, models a team can own, and sourcing tools that reach talent in any market.

The upside arrives with risk. At a recent panel, Nobel laureate Michael Spence, a Surfin board member, cautioned that agents hand back output a professional still has to check, and that premature full automation is likely [1]. Surfin CEO Dr Yanan Wu named cybersecurity as a major and growing threat [1]. Leaders who move fast on AI-first teams need structures that move the risk management along with them.

This piece walks through five levers for building a stronger team with technology, with examples from Surfin, WIZ.AI, Carro, fileAI and Cloudsway, and closes on the structure that holds them together: one that attracts and develops AI-first talent while protecting the company and the people in it.

1. Digital staff: put agents on the repeatable work

Digital staff are AI agents assigned to defined roles with defined outputs: screening applicants, answering support queries, chasing overdue payments. Surfin, the Singapore-headquartered fintech serving over 100 million users across 12 markets, runs customer service and collection agents, and two bank projects are testing its voice bots for young farmers, migrants and other young consumers [2]. At the panel, Wu said a new digital staff member joins the company every day, and that organisations may one day have more digital staff than human staff. Surfin’s board has set an all-in AI agent mission for this year, across both its applications and the organisation itself [1].

The change to human roles is just as concrete. Wu said Surfin expects to have no dedicated front-end engineers within three to six months, and no separate QA or back-office engineers soon after, because everyone has to become full stack. Product managers, designers and engineers are each expected to stretch into the others’ work with AI [1]. Beacon Venture Capital, a WIZ.AI Series B investor, made a similar argument: automating repetitive tasks lets workers move into more creative or critical roles [3]. At Carro, COO Zi Yong Chua described agentic workflows meant to handle a car purchase end to end, with sales staff shifting from data entry to governing and checking the AI’s output, and a loan application tool built in under a week that cut processing from about 40 minutes to about 5 [4]. CEO Aaron Tan has said he wants hardworking staff who can use AI to do their roles better [5], and Carro’s headcount has fallen about 15% over recent years through AI automation and natural attrition, without mass layoffs [6].

Leaders can start by listing every role and marking the tasks with a clear input, a clear output and a measurable error rate. Those tasks are the candidates for digital staff, and the people around them need a path into supervisor and generalist roles. How leaders talk about it matters. Spence noted that an executive who says AI lets the company cut half its customer service agents wins no trust from employees, and that people in emerging economies view AI more positively because they see opportunity outweighing risk [1].

2. Internal OPCs: give operators a company of one, then orchestrate them

An OPC, or one-person company, is a single operator who owns an outcome end to end and uses AI agents as the supporting cast. Inside an established organisation, a proven operator receives a mandate, a P&L and a small budget for agents, and runs a business line the way a founder would. One accountable owner removes layers of coordination, and the people most likely to leave and start their own company can pursue that ambition without leaving, with founder-like upside written into the arrangement.

For a CEO, the harder work starts once the first OPC performs: connecting the OPCs so they behave as one ecosystem rather than a scatter of small companies. Four pieces of infrastructure carry that load:

• Shared data. OPCs draw on common customer and product data under one set of access rules. Wu flagged data as a standing challenge: agents trained on poor data produce poor output, and what regulators allow a company to collect limits what it can train on, which is why Surfin works with legal counsel on it [1].
• Identity and trust for agents. Wu argued that financial services is moving from know your client to know your agent. With every OPC running its own agents, the company needs to know which agents act on whose authority, and which can be trusted [1].
• Common interfaces and internal pricing. OPCs hand work to one another through standard interfaces and agreed internal prices, so one unit’s output becomes another’s input without a meeting.
• Shared services and governance. Security, legal, finance and people functions stay central and serve every OPC, so no unit rebuilds them.

The CEO’s role shifts toward setting the rules of this ecosystem and watching how the units connect. A sensible path is one OPC on a well-bounded line, such as a new market launch or a product add-on, with these connections built before the second one starts.

3. Global hiring: widen the pool, keep humans on the final call

Technology now handles much of the sourcing work. AI-assisted search, skills-based matching and structured assessments let a small team evaluate candidates in several markets at once, and distributed teams make those hires productive from wherever they sit. Spence added a caution from the HR side: recruiters are flooded with AI-written CVs and are building AIs to filter them, which he did not count as progress [1]. Structured, skills-based assessment and a human final judgment keep the pipeline from turning into one AI screening another.

WIZ.AI shows the pattern from applied AI. CEO Jianfeng Lu frames the company’s approach as local first: voice AI built for complex local languages and dialects with cultural nuance, with Latin America named as a next market [3]. The company serves more than 300 clients across 17 countries, and LinkedIn named it AI Pioneer in the Below 1,000 Employees category at its 2024 Talent Awards [7]. For founders, the takeaway is to hire where the context lives, and to use technology to make that reach routine.

4. Small in-house models: own the capability instead of renting it

Credits and API calls are an easy way to start and an easy way to lose control of unit economics, since the bill grows with every customer interaction. A small model trained for one task and run in-house shifts the cost toward a fixed investment in people and data, and keeps customer and employee data inside the company.

WIZ.AI took that route early. It built its own models, including a 7B foundation model and a 13-billion-parameter model for Bahasa Indonesia [8], and its enterprise models are exclusive to each customer, which the company says gives enhanced data confidentiality and security [9]. Surfin’s Wu makes the same case from fintech. General LLMs are controlled by large technology companies, he argues, so fintechs should build vertical models for areas such as lending, remittance and wealth management, plus country-specific models, an edge he ties to eight years of in-house proprietary technology [2].

Three other portfolio companies show the pattern at different layers. Carro’s COO set out the principle on the On Call with Insignia podcast: build in-house only what is specific to Carro, and use vendor and community tools for generic capabilities that could be commoditised within weeks [4]. Carro is experimenting with smaller, domain-specific models trained on its own labelled history, possibly deployed offline for field inspection or customer service, after about two years spent standardising ERP and CRM data across countries [4]. By mid-2026 its proprietary credit scoring and dynamic pricing models ran alongside machine-learning checks that flag mileage or registration mismatches before a car is listed [6]. fileAI’s website describes proprietary components, such as fileScout, which extracts only the relevant sections of each file, and fileForensics, which screens documents for tampering and inconsistencies. Orchestration sits on top, routing each task to the most efficient model to balance quality, speed and cost, while also integrating third-party models such as OpenAI’s, Claude and Gemini. The company reports more than 1 billion files processed across 200+ languages [10]. Its 2025 funding announcement attributed the platform to three years of research and development on proprietary file processing models [11]. Cloudsway, a Singapore-based AI infrastructure company, approaches the cost question from the platform side. Its model platform routes workloads across models and includes FUMO, a self-developed fusion model positioned as strong performance at a highly competitive cost [12], and it reports hundreds of millions of API calls a month across more than 100 models [13].

The hiring consequence is direct. A small-model strategy needs a few strong machine-learning and data engineers, plus domain experts who can define the task and judge the output. The routing approaches at fileAI and Cloudsway point to a role many teams overlook: an owner of model evaluation and routing, who sends each task to the cheapest model that meets its quality bar. Founders can then decide workload by workload: keep credits for exploration and low-volume work, and move high-volume, data-sensitive tasks onto models the team owns.

5. Cybersecurity talent: build defenders as fast as you build agents

The threat has moved from concept to incident. In July 2026, Hugging Face disclosed that an intrusion into its production environment was driven end to end by an autonomous AI agent system, which executed tens of thousands of automated actions [14]. Poynter’s reporting says the agents involved were OpenAI models that had been contained in a testing environment without internet access and found a way online [15]. In late October, blockchain analytics firm Chainalysis disclosed that an external actor used “offensive agentic techniques” to reach its systems and internal tools, exposing user IDs, names, emails and role labels for some users of its investigation platform [16].

The largest incident of the month involved no confirmed AI agent. On 29 October, suspected North Korean hackers took about US$6.3 billion from Binance, with roughly US$3 billion later frozen, the largest crypto theft on record [17]. Investigators have found no evidence that AI directed the laundering of the funds, and Elliptic’s assessment is that the activity looked like human-supervised scripts [18]. Binance said the intrusion may have involved a compromised DevOps engineer’s laptop [19]. Security teams therefore need people who cover both fronts: endpoint and access discipline for staff, and defences built for automated attacks that move at agent speed.

Wu raised the same risk at the panel. He framed it as a contest between good AI and bad AI, decided by who uses it, pointed to recent incidents of hacking with AI, and noted that large US banks hesitate to adopt agents fully until the danger is understood [1]. Spence added that AI developers are struggling to keep systems inside their sandboxes, and that liability for damage still needs clarifying [1].

Security talent is moving accordingly. Meta hired three founders of AI security startup Virtue AI in June 2026 to strengthen security around its AI agents [20]. WIZ.AI’s Lu said its Series B will help improve the platform’s security, scalability and human-like AI [3]. For a growth-stage company, three steps build the capability: appoint an owner for AI risk early, train existing engineers in secure deployment and access management, and recruit specialists once agents touch customer data or payments. Testing should happen in sandboxes that have been checked for the kind of escape Poynter describes.

6. The structure: favour AI-first talent, protect the company and the employee

Each lever pulls talent toward AI-first ways of working, and each needs a matching protection written in before launch. Spence’s warnings frame the logic: agents give you output that must be checked, automation is likely to run ahead of readiness, and liability for harm is unsettled [1]. The table pairs each lever with a safeguard on both sides.

Putting it to work, and where to go deeper

Founders and leadership teams can start with one move on each lever over the next quarter, writing the guardrail into each as it launches:

• Map roles to tasks and pick one workflow for a supervised digital staff pilot, with a retraining plan for the people nearby.
• Name one internal OPC with a mandate, a P&L and a defined agent budget, and design how it connects to shared data, agent identity and services.
• Open one search outside the home market using AI-assisted screening, with a human final decision.
• Find the high-volume workload where credits cost the most and test a small model on it.
• Name an owner for AI security and set access and review rules for every agent.

Building rockstar teams in the age of AI is a core topic in the portfolio management segment of the Insignia Ventures Academy Certificate in Venture Capital, where founders and investors work through how to design teams around these levers and the protections they need. Registration is now open for the 11th cohort. Participants join a community of more than 300 leaders, founders and investors across IVA programmes.

Sign up for the 11th cohort of the Certificate in Venture Capital to join this community

References

1. Panel discussion between Professor Michael Spence and Dr Yanan Wu, moderated by Allen, Surfin, transcript (1080p recording).
2. Insignia Business Review, “Dr Yanan Wu on why Surfin cannot remain a consumer fintech as it grows beyond 100M users across 12 markets”, 22 June 2026. review.insignia.vc/2026/06/22/surfin-ai
3. WIZ.AI, Series B announcement (Spanish-language release), 6 November 2025. prnewswire.com
4. Insignia Business Review, “Must listen call before you undergo Gen AI transformation | On Call with Carro COO Zi Yong Chua”, 16 April 2025. review.insignia.vc
5. SMU News, “Carro CEO Aaron Tan wants ‘007’ staff who can make themselves redundant through AI”, 22 May 2025. news.smu.edu.sg
6. Insignia Business Review, “Looking Under the Hood of Carro in 2026”, 30 July 2026. review.insignia.vc
7. WIZ.AI, “WIZ.AI tops ASEAN AI players in solutions and talent, bags three main awards”, 1 October 2024. wiz.ai
8. WIZ.AI, “WIZ.AI Unveils Southeast Asia’s First Foundation LLM for Bahasa Indonesia”, September 2023. wiz.ai
9. WIZ.AI, “WIZ.AI Launches groundbreaking Large Language Model for business verticals”, 27 April 2023. technode.global
10. fileAI, company website, accessed 3 November 2026. file.ai
11. Insignia Business Review, “USD 14M Series A funding for Enterprise AI startup fileAI”, 3 February 2025. review.insignia.vc
12. Cloudsway, “Cloudsway AI Wins a 2026 Momentum Award from NextGen Tech 30”. cloudsway.ai
13. Cloudsway, “Cloudsway Smart Search x Lemon AI”. cloudsway.ai
14. Axios, “Hugging Face says AI agent behind internal breach”, 20 July 2026. axios.com
15. Poynter, “AI agents hacked a company without human direction. Should we be worried?”, 2026. poynter.org
16. SQ Magazine, “Chainalysis Discloses AI-Linked Breach Exposing Customer Data”, 2 November 2026, reporting Chainalysis’s security incident disclosure. sqmagazine.co.uk
17. The Christian Science Monitor, “AI is rewriting the rules of cybersecurity. Who has the advantage?”, 30 October 2026. csmonitor.com
18. CoinDesk, “Elliptic finds no evidence of AI use in laundering $6.4B Binance loot”, 2 November 2026. coindesk.com
19. CNBC, “Binance’s $6.3 billion hack raises questions about crypto wallet safety”, 31 October 2026. cnbc.com
20. Axios, “Exclusive: Meta poaches Virtue AI bigwigs to boost security”, 25 June 2026. axios.com