Today, artificial intelligence has become accessible to nearly everyone. Instead of relying on complex coding, AI can use simple, everyday language. But what exactly can we use it for, and how can we harness its potential effectively? Authors from the Harvard Business Review have explored this new era of human-machine collaboration, highlighting research that shows AI is poised to transform more than 40% of work activities soon.
Generative AI transforms jobs and the way we work
Generative artificial intelligence is set to revolutionize various jobs in the years ahead.
No longer exclusive to developers and IT professionals, AI tools empower anyone to provide instructions in plain, everyday language. Research from the Harvard Business Review authors reveals that generative AI has the potential to extend, automate, or reimagine most business functions and over 40% of all jobs in the United States. The most significant changes are expected in legal, banking, insurance, and capital markets, with additional impacts in retail, travel, healthcare, and energy.
In the future, many people will find that their success at work depends on their ability to get the best out of large language models like ChatGPT—and to learn and grow with them.
3 fusion skills to use AI effectively
HBR’s thought-provoking article highlights three fusion skills essential for anyone looking to use AI effectively and achieve meaningful results.
The ability to intelligently prompt involves instructing large language models in ways that produce better outcomes. This requires understanding AI’s language and leveraging its capabilities by breaking down processes into clear, step-by-step instructions or visualizing multiple possible solutions.
For instance, a customer service representative in financial services might use AI to resolve complex customer queries. A pharmaceutical researcher could rely on it to analyze drug compounds and molecular interactions. Meanwhile, a marketer might utilize AI to examine data sets and determine optimal retail pricing.
The integration of discernment involves incorporating human judgment—both professional and ethical—to ensure that the answers and solutions generated by AI are reliable and accurate.
Large language models often lack the technical or business background knowledge necessary to solve specific problems, and this gap is frequently filled with AI hallucinations. Identifying where the model requires further learning and determining credible sources to support this process is crucial, all while maintaining a strong commitment to data security.
Mutual learning involves supplementing artificial intelligence with authoritative knowledge bases when necessary.
This process is reciprocal: as generative AI learns from the organization’s data and expertise, the human training it gains the ability to leverage AI for increasingly complex challenges. This dynamic turns the machine into a valuable team member—a co-creator.
Let me know if you'd like further refinements!
Empirical research consistently shows that ad hoc instructions—commonly used by most AI model users today—often result in unreliable or poor outcomes, particularly for complex reasoning tasks. This applies across all functions, from customer service and marketing to logistics and R&D. Bringing greater rigor to using generative AI in the workplace is essential for achieving dependable results.
How to ask AI smart?
Always think step-by-step: when prompting AI, break down the process into its individual components and focus on optimizing each step. Visualize a chain of thought that leads to your desired outcome. Empirical studies show that their performance improves significantly when generative AI tools are guided to decompose reasoning tasks in this manner.
Even for the simplest tasks, including a “think step-by-step” prompt is beneficial. This encourages the AI to explain how it arrived at the result rather than merely providing an answer from a black box.
Mutual learning should also take place in stages
For complex tasks requiring expertise or business context in human-machine collaboration—such as law, medicine, R&D, or inventory management—you can achieve better results by integrating AI into the workflow gradually and in stages.
MIT researchers, for example, explored the feasibility of developing an “AI scientist” capable of integrating diverse experimental data and generating testable hypotheses. They discovered that ChatGPT 3.5-Turbo could be fine-tuned to learn the structural biophysics of DNA when the researchers divided this complex task into a series of manageable subtasks for the model to master.
This approach is equally practical in non-scientific areas, such as inventory management. Subtask stages can include forecasting demand, collecting inventory data, predicting reorders, evaluating order volumes, and assessing performance. We can leverage our expertise and information for each additional subtask to teach, test, and validate the model.
Make room for creativity
Many workflows, from strategy design to product development, are open-ended and iterative. To maximize human-AI interaction in these activities, it is beneficial to guide machines to explore multiple solution paths and respond less linearly and binaryally.
This approach to intelligent querying can also enhance AI’s ability to make accurate predictions about complex financial or political events. Research by Philipp Schoenegger and Philip Tetlock demonstrated this in an experiment where human analysts and forecasters were paired with GPT-4 assistants to create “super-forecasters.”
This collaboration aimed to assign probabilities and uncertainty intervals to possible outcomes while listing arguments for and against each scenario. The researchers found that predictions generated by the trained assistants—on topics ranging from the Dow Jones closing figures on a specific day to the number of migrants arriving in Europe via the Mediterranean in December 2023—were 43% more accurate than those produced by commonly used, untrained GPTs.
Critical thinking and judgement
Integrate the RAG
Not only can large language models be prone to hallucinations, but the information and data sets used to train them are often outdated. As a result, working with these models usually requires human judgment to determine how critical it is to ensure the outputs are reliable, relevant, and up-to-date. In domains where fresh and accurate information is essential—such as healthcare or finance—retrieval-augmented generation (RAG) can be a valuable tool. This method enables large language models to access and utilize external data sources not part of their original training.
Pay attention to data security
If your work with AI involves confidential data or proprietary information, avoid using open-source or public language models. Always rely on corporate-approved LLMs that operate securely behind corporate firewalls.
Watch out for distortions
Pay close attention to distortions that may arise in your prompts. For example, suppose a financial analyst asks an LLM to explain how yesterday’s quarterly report indicates that the company is ready for a five-year growth cycle. In that case, the response may exhibit recency bias—the tendency to overestimate recent information when predicting future events.
Large operators are actively developing solutions to help users address these challenges. Microsoft and Google are introducing features in their services to give users better control over harmful prompts and responses. Meanwhile, Salesforce is building an AI architecture that safeguards confidential customer data by capturing it during prompt creation, preventing its sharing with third-party LLMs. This system also evaluates outputs for risks such as toxicity, bias, or privacy violations and gathers feedback to refine prompt templates.
Always investigate suspicious answers
Hallucinations and errors are inevitable, even with careful planning, as current research shows. Therefore, always approach AI responses with caution and scrutinize them for errors or suspicious signs.
When users encounter an incorrect output, they often reflexively prompt the model to try again repeatedly, which can gradually degrade the quality of the response, as UC Berkeley researchers Jinwoo Ahn and Kyuseung Shin have shown. Instead, the researchers recommend identifying the specific step where the AI made a mistake and using another large language model to address that step.
This involves breaking the problem into smaller, individual components and using the new output to fine-tune the original model’s response. For example, imagine a scientist using OpenAI’s ChatGPT to develop a new polymer through stepwise computations. If an error occurs in the chain, they could turn to Anthropic Claude to break down the problematic step into smaller subproblems and provide an explanation.
This refined information can then be fed into ChatGPT to improve the answer. By applying chain-of-thought principles, this technique enhances the quality of outputs that initially appear incorrect.
You and AI: Master and Disciple
As LLMs grow in size and complexity, they may develop advanced new skills that were not explicitly part of their training but emerge after being customized with contextual data or knowledge. To foster this development, consider the following steps:
Provide AI with patterns of thinking
Before assigning a problem to a large language model, you can train it to adopt a specific way of thinking. For example, you can teach it the “least to most” approach, which involves breaking down a complex challenge into smaller, more straightforward tasks. Solve the least difficult task first, then use that solution as a foundation for tackling the next challenge, and so on.
Denny Zhou and his team at Google DeepMind demonstrated that this “least to most” method improved AI output accuracy from 16% to 99%.
Take, for example, a marketing manager at a sportswear brand seeking AI assistance to plan a new collection. Here’s how the problem can be broken down for AI:
- Target Audience: Identify fitness enthusiasts who could be potential customers—a relatively straightforward task, especially for a model trained on the company’s customer data.
- Messages: Develop messages that emphasize performance, comfort, and style—this is a more complex and creative task that builds on the identified target audience.
- Channels: Select social media platforms, fitness-specific blogs, and influencer collaborations to effectively deliver these messages to the target audience.
- Resources: Allocate the budget—often the most debated aspect in any organization—according to the chosen channels.
Teach AI new processes
You can delegate task execution to AI by guiding it with examples within the context of your prompts.
For instance, researchers reported in Nature that they trained large language models to summarize medical information by providing examples such as radiology reports, patient queries, status updates, and doctor-patient dialogues. Their findings revealed that 81% of the summaries generated by the models were equivalent to or better than those produced by humans.
You can also train AI by providing contextual information and guiding it through questions until it solves the problem.
Take two software companies, both aiming to increase their sales. At the first company, the sales team struggles to forecast software license demand effectively. Their manager starts by supplying historical sales data to the AI and asking about the expected demand for the next quarter. Next, they provide details on customers’ software upgrades and annual budgets, inquiring about the effects of seasonality. Finally, they add detailed statistics from CRM systems and marketing reports, asking how marketing campaigns impact sales.
At the second company, the sales team focuses on improving customer selection. Their manager provides specific financial data and asks the AI to rank customers by spending. They then proceed with follow-up questions about geography, technical expertise, and other relevant criteria.
At every step, both managers teach the AI and fine-tune its ability to perform tasks within their company’s sales strategy. Incorporating organizational and industry expertise into their interactions enables the AI to adapt. The AI gains more experience with each company’s unique sales processes, producing increasingly accurate and valuable responses.
Learning new fusion skills
The widespread adoption of generative AI skills will require significant organizational investment as well as individual initiative, learning, and hard work. While some companies already offer relevant training, most have yet to develop widely accessible programs.
HBR’s research underscores this point: in their 2024 survey of 7,000 professionals, 94% of respondents expressed willingness to learn new skills to work with generative AI, yet only 5% reported that their employers were actively providing training in this area.
As a result, many individuals will need to take the initiative and stay ahead of the rapid advancements in AI.
The AI revolution is no longer on the horizon—it’s here. Leading companies leverage this transformative technology to rethink how industries, functions, and jobs are performed.
Generative AI has significantly raised the stakes, demanding that humans collaborate, act consciously to ensure trust, and continuously refine the technology and themselves to achieve better outcomes. No other major innovation in history has spread at such a rapid pace. Work will evolve faster and more profoundly than many of us can imagine. Get ready: the future of business won’t be shaped by AI alone but by the people who can harness its potential most effectively.