07-07-2025
5 Things Businesses Need To Know Before Automating Work With AI Agents
Harvey Hu is Founder & CTO at General Agency, building state-of-the-art AI agents to automate complex web workflows.
Duolingo's CEO recently emailed employees announcing the company will go 'AI First.' This means replacing most contractor work with AI and requiring every team to try automation before adding headcount. A few days later, Shopify's CEO told staff that using AI is now a 'fundamental expectation' for every role. They are not alone: McKinsey's latest global survey finds 78% of companies already apply AI in business functions. In 2025, effectively automating your work often means going beyond ChatGPT and utilizing agentic AI that closely integrates with your day-to-day workflow. As a business leader navigating the plurality of AI tools on the market today, there are five important considerations before you decide if/how your organization should join this 'AI-First' wave.
1. Predictive Vs. Generative AI—Understanding What's New About This Wave Of AI
The first order of business is understanding what's different about this wave of AI and why it may influence your business more than past ones.
Prior to 2023, artificial intelligence in industries was mostly predictive. Predictive AI typically uses historical data to forecast a single target. Often, this is a user behavior metric that the business cares about, such as Click Through Rate (CTR) in advertising or the probability that a user finishes watching a video on social media. In the past 10 years (since landmark publications like Deep & Cross Networks and Tensorflow by Google), big tech companies have leveraged these models to earn billions of dollars in revenue. Though powerful, these models have narrow applications. A single model, for example, is designed and trained specifically for Google Ad prediction and can not be used for anything else. Coupled with their high training costs, businesses outside of tech have found it hard to harness predictive AI's power.
Generative AI, on the other hand, creates new text, images or audio without a strict target in mind. This means a single model can be trained to simultaneously learn many things, from answering customers' questions to solving math problems. Though still expensive, a single model now becomes useful to many businesses at the same time, from customer support to education. The downside of a weaker supervision and the lack of a single training target is that generative models are harder to evaluate. They can more easily 'hallucinate,' for example, where they produce false answers confidently.
2. Horizontal Agents Vs. AI Workflows
When browsing through the many options for AI tools in the market, it helps to distinguish between AI workflows and AI agents. Most 'AI agents' creating business value today, outside of simple chatbots, are actually 'AI workflows.' They are sequences of steps hand-designed to solve specific problems in a certain vertical, often built upon common tools like Zapier and n8n. Identifying the right workflow and setting it up could be time-consuming, but seeing through the 'agent' facade and understanding the underlying technology could help. AI Agents, on the other hand, make decisions on what actions to take autonomously. They can learn and excel at a broad range of tasks. Horizontal AI agents are starting to create value in the market in early 2025. It's still the early days, but we could see significantly better user experiences offered by them compared to 'AI workflows' this year, especially with stronger reasoning models and better multimodal LLMs.
3. Target High-ROI, Repetitive Tasks First
A Zapier study shows 94% of knowledge workers spend part of every day on repetitive, time-consuming chores such as copying data or formatting reports. They follow clear patterns and are ripe for automation. Start with processes that look like assembly lines: data entry, report compilation, standard email replies, document classification. Freeing even one hour a day per employee yields ~250 hours a year—more than a month of recovered capacity. Not to mention, these low-value tasks drain morale and contribute to burnout.
4. Data Access, Security And Privacy—Treat The AI Like A New Hire
In 2023, a Samsung engineer pasted confidential source code into ChatGPT; weeks later, the company banned public chatbots internally to prevent future leaks. Apple, JPMorgan, Verizon and others enacted similar restrictions. Instead of reacting to incidents like this with drastic measures, think ahead about what data and systems it should see—no more, no less.
Use the 'intern test.' If you wouldn't hand a summer intern unrestricted access to a customer database, don't hand it to an AI. On the other hand, if an intern needs access to some of your internal tools to contribute effectively, AI agents probably need the same access as well.
5. Research What Human Supervision Is Available
AI agents still make errors. A U.S. judge fined two lawyers $5,000 after ChatGPT invented six non-existent cases they failed to verify. PwC advises firms to 'rigorously oversee GenAI.'
Depending on your business, keeping human in the loop with your AI automations could:
• Reduce business risk. Keep an eye on your new AI agents the same way you wouldn't let your new hire handle a client meeting alone.
• Build trust with AI Agents to unlock more capabilities. Models today have limitations. Working closely with them to understand their capabilities could unlock more potential.
Over time, you can relax oversight, but it's essential to understand what level of human supervision is possible with it when building new automations. At a minimum, you should be able to audit details of the AI Agent's work, ex post facto, in case something goes wrong.
Conclusion: Automate Boldly With Eyes Open
McKinsey estimates that generative AI could unlock up to $4.4 trillion in annual economic value—roughly the size of Germany's economy. The AI-first era has begun; steer into it with eyes open and hand on the wheel by understanding the technology, its limitations and what to look out for when making your choice.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
