Trust but Verify: Introducing AI Into Your Business Safely

The Reversals Are Now as Public as the Announcements

Companies that went all in on AI two years ago may have benefited from the announcement. However, their public reversals are now getting the same amount of attention. Michael Giorgio, co-founder of Imaginovation, spoke on the current state of AI, saying, “Right now, with where AI is, it's in a little bit of a testing phase. From my experience, it's like you're getting hit with all of this dopamine of fun, cool, amazing tools." 

The fun phase took many by storm. AI initially appeared to offer fast profits, easy cost-cutting, and a compelling way to attract new investors. Those who rushed to join the AI gold rush without fully evaluating the new tech are now retreating or proceeing with more caution, unsure of the true benefits. 

In Klarna’s case, they charged ahead, claiming that an AI chatbot could cover the work of 700 customer service agents. The initial strategy meant laying off workers and pausing hiring to save the company money and increase productivity. Soon after the move, it became clear that the quality of service was declining. Customers needed an easy way to speak with a real human who could understand the emotions and nuances of their inquiries. To address these shortcomings, Klarna began recruiting human customer service agents.

Most Companies Aren't Cutting Headcount — They're Patching Broken Processes

Headlines across mainstream media, YouTube, and other online platforms highlight the perception that AI is causing mass layoffs. While this is true in some instances, many businesses are turning to AI to repair inefficiencies that are taking up their staff’s time and resources. 

"It's more like they want to create more value within their company,” Giorgio says. “Or it could be just to fix their manual processes and to automate them, digitize them, and just allow them and their team, their employees to save time, money, and just make them more efficient holistically within the organization." 

These processes could be simple assignments, such as extracting information from invoices, categorizing support requests, searching internal documents, or helping employees with the first draft of an email. All of these actions create a more efficient workplace and allow employees to put their time toward more worthwhile tasks. 

It’s when organizations focus on pure numbers instead of the issues AI can solve that problems arise. Giorgio continues, “They're saying, 'Hey, we have 3,000 admins that's going to save us this amount of money'... So I think that's also happening on a much larger scale. And it's scary, man." 

MIT has research supporting Giorgio’s concern. Although the MIT Project NANDA report is preliminary and contested, its estimate that 95% of organizations didn’t receive any measurable return from GenAI investments shouldn’t be completely dismissed. The report identifies the pattern that companies were often investing where AI was the easiest to showcase instead of where it could remove the most significant operational bottlenecks.

Start With the Audit, Not the Tool

To make the most of AI implementation, companies must begin with an examination of their existing workflows. Finding a problem to solve and looking at the AI capabilities that can solve it is far more effective than deciding to adopt new tech and looking for ways to incorporate it into your current systems.

"We would do a little bit of an audit, kind of like a business process audit... So you got to make sure that you have some sort of audit in that environment... understand where the problems lie, what needs to be fixed, how it needs to be fixed,” Giorgio says. "First of all, understand the purpose and the why behind the pain points and challenges of what it's going to solve. Why the heck do you want to include AI in some of your systems, your infrastructure, your sales processes, your QA?"

Before implementing AI, make sure you know the intended outcome and how it will integrate into your company’s processes. The MIT researchers pointed to a “learning gap” as one cause of unsuccessful deployments. In their report, they noted that generic systems can stall because they don’t retain feedback, adapt to the environment, or integrate well with workflows. Having a sense of the big picture before implementation will allow AI to deliver the most value.

Go Zero to One, Not Zero to 100

After you identify a bottleneck that AI can solve, automating the entire process immediately can be a big mistake. Giorgio says, “If you go too quickly, if you're trying to incorporate AI within your culture, within your team... if you go 100% all in, nope, you're going to have a lot of consequences... You can't go from zero to a hundred, go to zero to one, one to two, and just step by step and just do it with intention and purpose." 

Even the most useful new technology can have unintended consequences. Starting with a limited rollout gives an organization time to establish a baseline, test the quality of AI-assisted work, set realistic expectations, and develop an effective change management strategy. 

Employees can also give feedback on the process before it extends to the rest of the organization. At Imaginovation, employees gain access to either OpenAI or Anthropic tools based on their responsibilities, strengths, and weaknesses of each model. Over time, teams can determine how it's affecting the workflow and productivity. 

There is no set pace for scaling AI adoption. MIT Project NANDA found that large enterprises take about nine months to move a pilot to scale while mid-market firms take closer to 90 days. This happens because smaller companies can evaluate and modify processes faster as they have fewer layers of approval and decision-making.

The Two Costs You Forgot to Add to Your Budget

With all of AI’s upside, it’s easy to overlook two major expenses that can quickly erase anticipated savings: uncontrolled consumption and ungoverned data exposure. Giorgio warns, "You can't just let AI run and do its thing once you think it's trained. That's the problem. And a lot of companies are making those mistakes where they just trust it too much."

Uber fell victim to uncontrolled consumption in early 2026. After encouraging employees to take advantage of AI coding tools as much as possible, the company used up its entire annual budget by April. Since then, the company has taken steps to correct the mistake by setting a monthly cap at $1,500 in token costs per employee for each AI coding tool.

The other hidden cost is security. When employees use unsanctioned tools, it can give rise to shadow AI. Shadow AI can expose uploaded intellectual property, customer information, or other sensitive data in ways that put the company at risk, making strong AI governance an essential practice. In 2025, IBM studied 600 organizations and found that shadow AI added about $670,000 to the average company’s breach cost. It also revealed that one in five organizations had experienced a security incident linked to shadow AI, and 97% of the organizations that experienced an AI-related security incident said they lacked proper AI access controls. 

What Stays Human

AI technology will continue to advance and accelerate parts of company workflows. However, humans remain an integral part of business operations. Responsibilities such as physical work, high-context system design, and accountability remain dependent on humans. Finding a balance between humans and AI remains the most productive path forward, and it’s important for corporate leaders not to jump into AI too soon.

An excellent example of this comes from METR, a nonprofit AI research organization, which conducted a randomized controlled trial in 2025. It reported that experienced open-source developers took 19% longer to complete selected tasks with AI tools from early 2025, even though these developers believed they were doing the work 20% faster. METR began a follow-up experiment with the latest tools in mid-2025. The follow-up found that 10 of the returning developers’ work had sped up by 18%, but the overall result wasn’t statistically conclusive.

No studies are perfect, and factors like AI tools, user behavior, and working methods are constantly in flux, making it difficult to draw definitive conclusions. METR’s study and revision make it clear that the technology and research are still in their infancy. Jumping head-on into full-scale AI adoption without human oversight could undermine productivity or produce unintended consequences that companies will have to resolve later. 
 

About the Author

Tim Condon Chief Revenue Officer at Clutch

Tim Condon is the Chief Revenue Officer at Clutch, the leading global marketplace of B2B service providers. Prior to Clutch, Tim served as the Chief Revenue Officer at Homesnap, the top-rated real estate app built for agents, which CoStar Group acquired in 2020. During his tenure, Homesnap grew its paying user base from 0 to over 80,000 clients. In addition, he previously served as the Director of New Ventures at The Washington Post. In this role, he built several new businesses, including The Capitol Deal, which became the third largest deal site in the DC metro area and was known for giving away 100,000 pizzas

See full profile