The automation paradox: Why shrinking your way to success with AI is often not a winning strategy

AI, pundits say, is on a path to becoming “invisible as electricity”—so deeply woven into daily workflows that we hardly even notice it. At the same time, AI agents may soon automate a steadily growing share of repetitive, tedious tasks while also extending the productivity of workers. One AI-enabled worker can “do what used to take a team of dozens,” said Deloitte’s Chief Futurist Mike Bechtel. In fact, Deloitte’s Tech Trends 2025 predicts this shift could revive Apple’s 2009-era mantra: “There’s an app for that” could morph into “There’s an agent for that.”

Yet that doesn’t mean organizations can simply shrink their way to success.

The perils of ‘shrinking’ your way to success with AI

As AI seeps into the enterprise, the automation paradox looms—yes, AI can lighten the workload at the front end, but it can also ramp up complexity behind the scenes. The automation paradox necessitates a more nuanced approach than assuming automation inherently yields efficiency gains across the board. Mature AI deployments often introduce architectural complexity that requires sophisticated human oversight and expertise — and they thrive on strategic data infrastructure that often require specialized expertise to build.

Deloitte has observed an evolution in how large companies and governments shift how they think about AI implementations. “In 2023, many talked about using AI—both traditional machine learning and Gen AI—to do the same work faster or cheaper,” Bechtel said. That often implies fewer people. “But as we enter 2024, more organizations realize you can’t shrink your way to success. Sure, you can optimize costs, but that doesn’t drive growth,” says Bechtel. “Instead, they’re saying, ‘We can do today’s work better’ by giving employees a ‘super suit’ or ‘force multiplier.’ Then looking beyond 2025, they’re asking, ‘What new work can we do that didn’t exist before?’ It’s reminiscent of the late ’90s dot-com era: companies that just added ‘.com’ to their name didn’t do so well, but those that used the internet to create entirely new offerings succeeded.”

From tedium to transformation

As Bechtel put it referring to a hypothetical employee, “If you automate Toby’s workload, you don’t necessarily replace Toby,” Bechtel continues. “You free Toby up for strategic work that was on the backlog. It’s about elevating people, not just replacing them.”

But adopting the latest innovations alone doesn’t guarantee success. Organizations must take stock of the lessons history has to offer, Bechtel says. “Another thing we’ve learned in Tech Trends is that futurists are secret historians. Most ‘new’ things are just the next page in a longer story.” As a case in point, he points to the early 1990s when even a simple website required buying your own servers—an often expensive ordeal. The cloud and SaaS has dramatically shifted the landscape. Today, that same progression is propelling us into the age of AI agents, yet the fundamental challenges of cost, talent, and adoption remain as relevant as ever.

When less is more

But specialized AI, not catch-all chatbots, may be the smarter bet. As Bechtel puts it: “Innovation loves constraints.” He cites domain-specific AI offerings like Salesforce’s AgentForce or ServiceNow’s Xanadu—agents laser-focused on key business tasks. “A giant monolithic chatbot can do anything, but that’s too broad,” he says. “People don’t know where to start. If you have a smaller, domain-specific AI that says, ‘I’m here for supply chain help,’ it narrows your questions. That helps both you and the AI.” Plus, not every task needs a large language model; sometimes a lightweight tool will do. Success hinges on addressing a real problem rather than chasing tech for tech’s sake.

‘Garbage in, garbage squared’

Data is the bedrock of effective AI, which is why “bad inputs lead to worse outputs—in other words, garbage in, garbage squared,” as Deloitte’s 2024 State of Generative AI in the Enterprise Q3 report observes. Fully 75% of surveyed organizations have stepped up data-life-cycle investments because of AI. Layer a well-designed data framework beneath AI, and you might see near-magic; rely on half-baked or biased data, and you risk chaos. As a case in point, Vancouver-based LIFT Impact Partners fine-tuned its AI assistants on focused, domain-specific data to help Canadian immigrants process paperwork—a far cry from scraping the open internet and hoping for the best.

“A major theme of our Tech Trends research is that trust precedes tech,” Bechtel warns. “You can’t just build something cool and then retrofit governance. It’s a losing strategy. Many organizations say, ‘We want Gen AI,’ but first they need a solid data strategy. Finally, it’s about training AI on data that represents the future you want, not just the past. Otherwise, you inadvertently codify old biases.”

When leading with real problems can lead to real solutions

That attention to trust and ethics also underscores a broader shift in IT’s evolving role. Tech Trends dubs it “IT Amplified”: as AI systems mature, IT leaders are the first to don these “super suits,” tackling real, tangible problems—like reducing customer wait times by 60%—rather than trying to drag the organization through forced transformations. “If you lead with real problems, you welcome new tools because they solve something tangible,” Bechtel notes. “If you’re clinging to old processes, these changes feel forced.”

In other words, investing in AI isn’t just about cutting-edge systems or lofty ambition; it’s fundamentally about rethinking the people, processes, and trust that allow any new technology—be it a web server in the ’90s or a domain-specific AI agent in 2025—to realize its fullest potential.