Outsourcing in the AI Era: What Companies Should Never Offshore Again, and What They Still Should

AI is making repeatable work cheaper, but it is also making judgment, context, ownership, and institutional learning more valuable. The new offshoring question is not where work can be done. It is what a company cannot afford to stop understanding.

In April 2026, Box CEO Aaron Levie described a strange contradiction now running through enterprise technology. His company was launching Box Automate, a service that lets AI agents handle tedious business processes such as invoice processing and data extraction from corporate documents.

The appeal was obvious as companies have millions of documents, countless repetitive workflows, and too many expensive people still spending time moving information from one system to another. But Levie also drew a sharp line around the fantasy of full automation.

“You’re not going to vibe-code an ERP system,” he told Reuters. “I don’t think you’re going to vibe-code a CRM system. The risk is simply too high.” Then came a surprising admission: after using AI a lot, he said, you realize “we need humans for almost every” part of the business.

This is a better place to begin the offshoring debate than the usual question of cost. The old conversations happened around whether work could be moved somewhere cheaper. The new one asks what happens when the work being moved is sometimes a signal, sometimes a control point, or even it is the place where a company learns what its customers are really saying, where its product is weak, where its finance model is leaking, where its compliance risk is forming, or where its operations are becoming too dependent on people who hold everything in their heads.

This is why the next offshoring mistake will not look like the old one. It will not simply be a company sending too much work to a cheaper location. It will revolve around a company sending away the layer of work that teaches how the business is changing.

Deloitte’s 2025 Global Business Services survey shows that companies are still expanding global service models, with GBS organizations moving deeper into digital, AI, analytics, finance, IT, customer experience, and end-to-end process ownership.

But the boundary has moved. AI is eating comprehensively into the repeatable layer. McKinsey’s 2025 workplace AI research found that 92% of companies planned to increase AI investment over the next three years, while only 1% of leaders described their companies as mature in AI deployment.

This gap matters because it shows where real pressure is. Businesses are feverishly deciding which work should be automated, which work should be globally distributed, which work should sit with embedded external teams, and which work must remain close to the people who own the consequences.

The lazy version of this debate says, “offshore execution, keep strategy.” It sounds clean, but real companies do not work that neatly. A customer support ticket can be an execution task until it becomes the first evidence of a product failure.

A finance task can be routine until it reveals that margins are breaking in one segment, or a developer ticket can be simple until it touches architecture, security, or technical debt. Similarly, a content task can be a marketing requirement until it starts shaping how the market understands the company.

So, the real question is sharper: what work can be safely externalized without weakening the company’s memory, judgment, and ability to learn? The boundary now runs through the work itself. Work that produces outputs can usually be specified, measured, reviewed, and governed through a defined process.

Work that produces understanding builds customer insight, institutional knowledge, judgment, risk awareness, and decision context that shape how the business operates. The first can often travel. The second needs much more careful ownership.

The Offshore Boundary Now Runs Through the Work Itself

The mistake many companies still make is assuming that offshoring decisions can be made at the level of a department. Thinking in terms of support work can be sent offshore, finance can be handled offshore, design can be managed offshore, or engineering tasks can be done offshore is quickly becoming obsolete now.

This sounds practical in a boardroom because departments are how budgets are arranged, vendors are evaluated, and hiring plans are approved. But real work does not behave according to organization charts. Inside every function, there are tasks that travel cleanly, and tasks that carry more meaning than their job description suggests.

Let’s take examples of customer support services. A retailer can use an offshore team, a contact center, or an AI voice agent to answer common questions about order status, store hours, returns, appointment scheduling, or product availability. Home Depot’s recent rollout of an AI voice agent built on Google’s Gemini technology is a good example here.

The company is using AI to replace clunky phone menus, understand customer intent faster, route calls, check inventory, and even help build ready-to-buy carts for shoppers. Their point is not to automate every customer conversation, but to handle those customer interactions which are structured enough that a machine or distributed team can handle the first layer better than a stretched in-store associate or expensive internal team.

But the same support function also contains the conversations a company should be nervous about losing. A customer who cannot make your product work after three attempts is not just a ticket or an escalation, but a business account threatening to leave.

A recurring complaint about billing, packaging, refunds, delivery promises, product quality, confusing onboarding, or poor handoffs is often the first place where a company can hear its own system failing. If that work is treated as queue clearance, the business may keep its response time green while losing access to the truth customers are trying to tell it.

Klarna became one of the more visible examples of this tension because it pushed hard into AI customer service and then had to soften the story. The company’s chief executive, Sebastian Siemiatkowski, had previously promoted AI as doing the work of hundreds of support agents. By late 2025, Reuters reported that he had acknowledged some customers still preferred humans, with more complex issues being routed back to human agents.

This is a useful example because it is not a simple anti-AI story. It shows the line companies keep rediscovering where automation can absorb simple service interactions, but the harder moments still need judgment, empathy, and a person who can understand the commercial meaning of the complaint.

This is why global capability centers (GCCs) have become such an important market signal. Companies are still moving work to India and other talent hubs, but many are doing it through structures that keep capability closer to the company’s own operating system. Reuters reported that India could have more than 2,400 GCCs by 2030, with export revenue from these centers expected to rise from $64.6 billion to $110 billion within five years.

Chevron’s Bengaluru engineering and innovation center gives the trend a concrete shape. The company has said the center will support digital workflows, high-performance computing, geological modeling, and digital replicas of processing plants, with a planned $1 billion investment over the coming years. This is capability-building global work in another geography, not just a casual delegation.

The boundary is no longer between domestic and offshore. It is between work that can be specified and work that must be interpreted, work that produces output and work that teaches the company something it needs to know. A company can distribute both, but the second demands a tighter operating model. Context, escalation, interpretation, decision rights, and feedback into the core business must remain visible and governed, even when the people doing the work sit elsewhere.

Work That Still Travels Well Is Changing

The work that still moves well across borders is increasingly the work that can be placed inside a visible operating system. This sounds less dramatic than the old language of outsourcing, but it is closer to what is happening in the market. Companies are trying to separate work that can be governed by workflow, data, automation, service levels, and review from work that depends on judgment formed inside the firm and are fiercely moving away from sending “tasks” anyway.

This is why the most interesting signal is the redesign of the concept of outsourcing. KPMG’s 2025 research found that nine in ten executives say transactional service delivery models no longer meet their needs, while labor-led outsourcing is expected to fall from 55% to 37% over two years and software-based service delivery is expected to rise from 14% to 30%. The useful reading of this data is that the easiest work is being absorbed into platforms, agents, workflows, and automation, while human delivery is being forced into more specific, reviewed, and context-aware roles.

We can see the same shift in customer operations, where the old call-center model is being squeezed out by AI from below and by customer expectation from above. Reuters reported in October 2025 on Indian AI start-up LimeChat, whose co-founder Nikhil Gupta said, “Once you hire a LimeChat agent, you never have to hire again,” while claiming its generative AI agents could reduce the number of workers needed to handle 10,000 monthly customer queries by 80%.

This quote is provocative because it shows how far the commodity layer of service work is being aggressively re-priced. The market is asking which queries are predictable enough to be handled by software, which need human recovery, and which should flow back to product, pricing, risk, or leadership because they carry information the company cannot afford to miss.

This is also where the offshoring boundary has become more subtle. The mistake is to confuse repeatable work with low-value work. A hospital administrator tracking insurance documentation, a retailer maintaining thousands of product listings, a SaaS firm cleaning customer records after a CRM migration, or a manufacturer checking supplier compliance records is not doing “cheap work” in any serious sense.

The work is necessary, but it becomes suitable for global delivery when the rules are explicit; the evidence is visible, and exceptions are escalated rather than hidden. In such a version of offshoring, the external team is a part of the control system and not a dumping ground for mundane work.

This is also why global capability centers are expanding rapidly. Companies are still moving work to talent markets such as India, but they are doing it with more ownership and tighter links to core operations. With India set to have more than 2,400 GCCs by 2030, the stronger interpretation is that firms are trying to keep strategic capability inside their own institutional structure while still using global talent extensively.

Where Companies Are Starting to Get Burned

The next offshoring regret is likely to come from work that looked safe because it was already sitting inside a process. It looked manageable in the form of a queue, a workflow, a ticket, a vendor dashboard, a weekly report, a set of review comments, or a compliance checklist.

The problem is that more of these workflows now sit inside chains of software, AI systems, contractors, data providers, cloud platforms, and specialist vendors. A company may still own the customer relationship or the final product, while the knowledge of how the work behaves is spread across parties it does not fully control.

This is why third-party risk is becoming a much bigger part of the conversation. EY’s 2025 Third-Party Risk Management Survey says companies are centralizing risk functions and using AI to monitor vendors because risk now moves through supplier networks, technology dependencies, data access, and external delivery partners. The survey is useful because it shows that the outsourcing question is now about whether the company can see, test, and govern the work once it leaves direct supervision.

Companies are trying to regain visibility into the systems their vendors are using on their behalf. AI makes the problem more awkward because it can hide complexity behind smoother output. Reuters captured this tension in its reporting on business leaders struggling with AI deployment, where Klarna’s chief executive acknowledged that some customers still prefer humans for more complex issues after the company had heavily promoted AI customer service.

Moreover, Forrester has predicted that companies would delay about a quarter of planned AI spending in 2026. The point is that automation has exposed how much commercial work still depends on exception handling, review, customer judgment, and the ability to recover when a process breaks.

There is a similar warning in hiring technology. Amazon’s new Connect Talent system uses AI to conduct interviews, screen applicants, and prepare recruiter notes for mass hiring. Such a system is attractive because seasonal hiring at Amazon’s scale involves enormous volume, and AI can remove friction from interviews and recruiter workflows.

The commercial logic is clear while the governance question is just as clear: when screening, assessment, notes, and recommendations are automated or vendor-supported, the company has to know which judgments are being made by the system, which ones are being reviewed by humans, and where errors or bias can be challenged.

This is the newer boundary companies are being forced to draw. Work can move outward when the company still has enough visibility to understand how it is done, review the output independently, see how exceptions are handled, retain the relevant decision rights, and reconstruct what happened when something goes wrong. Those are also useful tests before externalizing an activity because they show whether the company is preserving the learning and accountability around the work, even when execution sits elsewhere.

This is also why some of the most important work to keep close in the next few years will not sound “strategic” in a traditional sense. Vendor governance, AI review, model-risk oversight, incident response, data-access control, knowledge-base ownership, escalation design, technical documentation, customer-intelligence synthesis, and post-mortem discipline may look like administrative layers from a distance.

In reality, they are the connective tissue that allows global teams, platforms, vendors, and AI systems to operate without slowly blinding the company. The companies that learn this early will still use external partners heavily and will hire globally, automate aggressively, and distribute execution across countries.

AI is Redrawing Where Judgment Sits

The more difficult shift to grasp is how AI rearranges where judgment is expected to sit inside a company. The effect is subtle at first. but over time, the center of gravity moves. More output is generated outside the firm’s immediate line of sight, while the responsibility for that output still sits inside it. This tension is now visible across sectors that rely on large volumes of interpreted information.

In banking, for instance, regulators have begun to focus less on whether firms use third parties and more on whether they understand them. The UK’s Prudential Regulation Authority and Financial Conduct Authority introduced formal operational resilience requirements that require firms to identify “important business services,” map dependencies, and prove they can continue operating within defined impact tolerances even if third parties fail.

The regulatory language matters because it treats external providers, cloud infrastructure, and technology platforms as part of the firm’s operating system, and not as distant vendors that can be managed through contracts alone.

The same logic is moving into AI governance. The European Union’s AI Act, formally adopted in 2024, does not prohibit the use of external models or providers. It requires companies to understand how high-risk systems are built, tested, and deployed, and to maintain documentation, oversight, and human control where necessary.

The practical implication is that a company cannot treat AI-generated output as external execution in the way it once treated outsourced processing. The firm is expected to remain accountable for how decisions are made. Even if those decisions are supported by models or vendors, it does not fully build itself.

This begins to change what kind of work can be treated as portable. It is no longer enough for a task to be defined and measurable. The company must also be able to explain how it was done, especially when the outcome affects customers, financial decisions, risk exposure, or compliance obligations. This is why some of the pressure is now moving into what used to be considered “support layers.”

Documentation, audit trails, model logs, escalation records, and decision histories are becoming more important now than ever before. In earlier outsourcing models, these were often treated as overhead, but in AI-assisted and vendor-heavy systems, they become the only way a company can reconstruct what actually happened when something goes wrong.

A similar shift is visible in software development. The introduction of AI coding assistants has accelerated output, but it has also increased the importance of review. Microsoft and GitHub have both pointed to productivity gains from AI-assisted development, but they have also emphasised that developers still need to validate, test, and understand the code being produced.

The gain is greater speed at the front end while the cost is a greater need for discipline at the back end. When that discipline weakens, companies can find themselves maintaining systems that fewer people fully understand.

Once that dynamic sets in, the question of offshoring changes again. A company can still use global development teams, external vendors, and AI tools together, but it cannot treat the resulting output as if it came from a simple, observable process. The system producing the work has become more layered, and the company’s responsibility to interpret it has increased.

This is also why companies are investing more heavily in internal capability even while continuing to operate globally. The growth of global capability centers is often described in terms of cost, but the more relevant point is control.

Firms want access to global talent, but they also want that talent to sit inside their own governance, data systems, and decision structures. Reuters reported that companies expanding in India are increasingly using GCCs to handle not only execution work but also engineering, analytics, AI development, and operations that directly influence global decision-making.

The trend shows that companies are not abandoning distributed work. They are simply changing how close it sits to the core. Work that was once treated as external execution is being pulled into structures where the firm can observe, question, and learn from it continuously.

The implication for offshoring is straightforward, even if it is uncomfortable. Work can still move across borders and can still be supported by vendors. Moreover, it can still be accelerated by AI, but the company cannot afford to lose the layer where judgment is applied, where assumptions are tested, and where outcomes are explained.

The Mid-Market Cannot Copy the Enterprise Playbook

The companies most exposed to this risk are the mid-sized firms trying to modernize while still running lean. They are not giants with thousands of engineers, procurement teams, compliance officers, vendor-risk specialists, and global delivery heads.

A large bank can build internal AI governance teams, negotiate directly with cloud providers, create regional capability centers, and absorb a few failed automation experiments as the cost of learning but a 200-person manufacturer, a regional healthcare group, a growing SaaS firm, a legal services company, or a specialist retailer usually does not have that cushion.

That is why the offshoring question lands differently in the mid-market. For small and mid-sized companies, external talent is not simply a cheaper substitute for local hiring. It is often the only practical way to access capability that would otherwise be too slow, too expensive, or too difficult to build internally.

The OECD’s 2025 work on SME digitalization makes this gap clear: smaller firms can gain from digital tools, but they face persistent barriers around skills, process change, financing, cybersecurity, and management capacity. In a separate 2025 OECD report on generative AI and the SME workforce, many SMEs said AI could help compensate for labor shortages and skills gaps, while the report also noted that SMEs often lack the workforce depth needed to turn isolated AI use into real organizational capability.

This changes the moral center of the debate. Telling smaller firms to “keep everything strategic in-house” sounds sensible until you ask who exactly is going to do the work. Many of these firms do not have spare internal teams waiting to take ownership of analytics, automation, CRM hygiene, digital operations, AI experimentation, cybersecurity monitoring, content systems, workflow documentation, cloud administration, or reporting infrastructure.

Their senior people are already stretched across sales, delivery, hiring, client management, cash flow, and daily firefighting. In that world, refusing to externalize can become its own form of weakness. The business keeps control in theory while allowing execution to decay in practice.

The more useful distinction is between outsourcing that replaces ownership and external support that creates capacity for ownership. For mid-sized firms, the practical issue is usually capacity rather than a desire to internalize every high-context capability.

The company still needs to retain ownership of the decisions, judgment, and institutional learning that shape the business, while embedded external teams can provide the specialist capacity needed to exercise that judgment well. That is where remote staffing and global teams become especially useful, because they can sit close enough to the business to understand its systems, history, and priorities while still giving the company access to capability it may struggle to build entirely in-house.

The pressure is visible in talent data. ManpowerGroup’s 2025 global talent shortage survey found that 74% of employers reported difficulty finding the skilled talent they need, and among the actions employers are taking are using automation or AI, recruitment-process outsourcing, sourcing global talent in lower-cost markets, expanding outsourcing, and using business-process outsourcing. This combination is important because employers are not choosing one lever but are stacking several because the talent problem has become too broad for a single answer.

For small and mid-sized firms, the danger is usually miscalibration. They may offshore too little because they fear losing control, then spend years with slow systems and overworked internal teams. Alternately, they may offshore too much too quickly because the first cost comparison looks irresistible, then discover that nobody inside the company can properly review the work, challenge the assumptions, or convert vendor output into business improvement. The first mistake creates stagnation while the second creates dependency.

AI is making both mistakes easier to make. Off-the-shelf tools have lowered the barrier to experimentation, which is good for smaller firms, but they also create a false sense that capability has been acquired when only access has been acquired.

Reuters reported that Italian firms using AI doubled in a year to 16.4% in 2025, while adoption remained far lower among smaller firms than larger ones, and the main barriers included lack of skills, unclear regulation, data-protection concerns, and high costs. This pattern is likely to be familiar across many SME-heavy economies: tools spread quickly, but organizational confidence and control lag behind.

This is where the next generation of offshoring will be judged. A mid-sized company does not need to imitate a multinational’s GCC but it needs a smaller version of the same discipline: clear ownership inside the business, external teams close enough to learn the operating context, enough process visibility to catch drift early, and enough internal judgment to decide what should be automated, what should be handled by a distributed team, and what should stay with leadership.

The New Risk Is Hollowing Out the Company’s Own Capability

The mid-market version of this problem becomes clearer when you look at how quickly companies are buying AI and outsourcing capacity without always building the internal muscle to govern either of them. McKinsey’s 2025 global AI survey found that 23% of organizations were already scaling agentic AI in at least one business function, while another 39% had begun experimenting with AI agents.

This is a remarkable rate of movement for systems that can plan and execute multi-step workflows, especially because many companies are still working out who inside the business is responsible for reviewing what these agents do, how mistakes are escalated, and how much judgment should be delegated to software in the first place.

For large enterprises, this gap can be turned into a formal operating program. They can create AI governance councils, appoint model-risk leads, build internal automation centers, and add layers of vendor oversight. Smaller and mid-sized firms usually move in a more improvised way.

A founder approves a new AI tool because the team is stretched, a department head hires a remote team because local hiring is too slow, a marketing manager starts using an offshore production team or an operations lead plug in a workflow automation platform. Each decision may be sensible on its own, but over time, the company can end up with a distributed operating system that nobody has fully designed.

This is one reason third-party risk has moved from procurement concern to board-level concern. EY’s 2025 global third-party risk management survey says organizations are centralizing risk functions and using AI to improve oversight because external relationships now include suppliers, technology providers, data processors, subcontractors, and AI-enabled service layers. This shows that the outsourcing boundary is now a visibility decision and no longer just a buying one.

The same pressure is now visible in the outsourcing industry itself. When TCS announced plans in 2025 to cut more than 12,000 jobs, it was described as a sign of AI’s shake-up of India’s $283 billion outsourcing sector, where automation is changing the economics of coding, testing, infrastructure management, and support work.

TCS said the cuts were linked to skill mismatches and a push to become “future-ready,” while analysts read it as part of a deeper shift in which clients expect fewer people, more automation, and better outcomes from service providers.

The shift should matter especially to mid-sized firms because they often buy from the market after the large enterprises have already reshaped supplier behavior. If major clients begin demanding automation-led delivery, outcome-linked pricing, AI-enabled service desks, and faster productivity gains, service providers will redesign their models around those expectations.

Smaller firms will inherit that market, even if they do not yet have the governance maturity of a multinational firm. They will be offered faster delivery, smarter workflows, cheaper execution, and AI-assisted teams. The harder question is whether they will also have enough internal capacity to understand, review, and absorb what they are buying.

The smarter mid-sized firms will then treat global hiring and remote staffing as capability extension, not capability replacement. They will use outside teams to create capacity, continuity, coverage, and operational discipline, while keeping enough ownership inside the business to ask harder questions. This is also the commercial opening for better offshore models.

The market no longer needs another cheap-labor promise. It needs delivery models that help smaller and mid-sized companies expand without losing their own grip on the business. This means embedded teams, clearer escalation paths, stronger documentation, reviewable workflows, AI-aware delivery, and people who can stay long enough to build context.

The winning provider in this environment will not be the one that says, “send us the task.” It will be the one that helps the buyer understand which part of the work should move, which part should remain owned, and how the two should stay connected.

Conclusion: The New-Age Offshoring Battle is Between Output and Understanding

The offshoring debate has always been framed as a question of location. Work moves from one geography to another, and the discussion follows the movement of labor, cost, and talent. That framing no longer explains what companies are actually dealing with.

Work now moves through systems, vendors, AI models, and distributed teams in ways that are far less visible than a simple shift in headcount. This is why the line companies are drawing is cognitive and no longer geographic. It sits between work that can be executed at a distance without weakening the company, and work that must remain close enough for the company to understand, question, and improve.

The market signals point in the same direction. AI is compressing the value of repeatable execution, as seen in the shift away from labor-heavy outsourcing models toward software-led delivery. Global capability centers continue to expand as controlled extensions of the firm’s own operating system.

Regulatory frameworks from the European Union to the Bank of England are making it clear that responsibility does not travel with the vendor. Remote service providers themselves are moving up the stack, selling capability, automation, and managed outcomes rather than simple manpower. Put together, these shifts do not argue against offshoring. They argue against treating offshoring as a shortcut.

For small and mid-sized firms, this is less a strategic philosophy and more about daily constraints. They do not have the option of building every capability internally. They will continue to rely on offshore teams, remote employees, vendors, and AI tools to operate at scale. The advantage will come from how they combine these elements without losing control over the business itself.

The firms that grow well will not be the ones that keep everything in-house or the ones that send everything out. They will be the ones that know which parts of the work must remain understandable. That is the line that is settling the debate now. It is not between countries, and not between humans and machines, but between output and understanding.