Data Economy AI agent Strategy Options
Jarkko Moilanen (Ph.D.)
Data Platform Product Head, Open (Source) Data Product Specification maintainer, Books authored: Terraforming Data Product Governance, AI-Powered Data Products, Deliver Value in the Data Economy, and API Economy 101.
We all probably have played with ChatGPT. Some of us have adopted it in daily life and take advantage of it in various tasks. ChatGPT is nice and depends on human actor activity. We can label ChatGPT and alike as AI copilots augmenting human capabilities. The AI agents which do not require constant interaction with humans are autonomous AI agents. Next to the AI agents we have humans as well. How are those different and how to use them? For this purpose, I created a simple model in which we have 4 actors:
Actors?—?building?blocks
Human actors are obviously us homo sapiens doing the work we did before the rise of AI. Copilot actor refers to AI/ML- powered assistants which augment human capabilities.
Autonomous agent-agent actors are independently task-performing AI agents which for example enable more flexible and dynamic automation at the system level for example data pipeline creation. The difference compared to traditional automation is that AI agent is capable of defining the needed integrations on their own and learning in order to be more accurate in the future.
The last actor is the autonomous human-agent actor in which case the agent is capable of doing more than just automation in integrations etc. It can be given an objective, which it can break down into tasks to execute, and communicate with other agents. This latter takes input from humans as objective, but after that begins autonomous task completion and possibly uses other autonomous agent-agent actors in the process. In the following, we take a closer look at the strategies in depth.
Augmentation-driven strategy
In the first option which we have named “Augmentation driven” the humans are given copilot agents, which augment the human capabilities. The interaction requires human input and almost nothing happens without it. This is what we have now entered to the markets as solutions. The most prominent example is ChatGPT. It does not do anything unless we humans tell (ask) the agent to generate text, write code, or analyze a file.
This is the most light-weight and easy to start with strategy. The beginning of the book “AI-Powered Data Products ” discusses this option in great detail. In this option, you keep the foundations intact and bring a ’helping hand’ into the tool stack. This option causes the least changes in operations and in architecture as well. You might need to bring some new technology to your stack to convert supporting code-level functionality to copilots but that should not be too hard in most cases.
Scale-driven strategy
In the second option, we have 3 actors: humans, copilots, and autonomous agent-agent actors in the backend. In this case, the aim is to increase the level of automation to the next level instead of just doing rules-based automation. Also, the copilots are there to remove repetitive tasks from humans with the help of automation. In this option, humans and AI agents divide the driver's seat. Copilots actors execute tasks again only if the human acts first. The autonomous AI agents in the backend instead act autonomously for example on behalf of the copilot and thus increase the power of the copilot.
Scaling is something we all look for. You can’t have a good business without scaling. In the world of AI one of the biggest issues relies now on Large Language Models (LLMs). It is said that training of ChatGPT 4 costs $600 Million and that Openai inc might not be able to maintain the development model in the future. New models to handle LLMs must be found and one strong candidate is decentralization of it. Yet there are issues to solve before we get the full value of decentralization.
For data teams looking to enhance their operations, the idea of utilizing AI copilots powered by decentralized Large Language Models (LLMs), has enormous promise. Data teams may drastically shorten the time it takes to deliver data products and use less effort in their production by including these AI copilots into their workflow. This is accomplished by automating certain tasks, offering suggestions for solutions, and generally assisting AI copilots throughout the process of creating data products.
The ability of AI copilots to grasp and produce conversational language is one of the most notable benefits. In order to capture complex technological requirements through conversational exchanges done in plain language, this promotes seamless connections with clients. In addition to simplifying the process for customers, this departure from conventional forms and paperwork improves the accuracy of demand gathering, creating a strong foundation for the succeeding stages of product development.
Working with AI copilots is iterative, which adds even another level of creativity to the procedure. AI copilots can instantly create previews or prototypes of the desired data products during these natural language dialogues. Customers can see in real-time how their needs are being transformed into concrete data products thanks to this dynamic preview method. As a result, any necessary corrections, explanations, or revisions may be quickly implemented, encouraging a cooperative and effective method of polishing the finished output.
In summary, the addition of AI copilots represents a significant improvement in the data team’s toolkit and will help them achieve more efficacy, productivity, and customer happiness.
领英推荐
Now we have tackled the scaling part and it’s time we take another leap and turn our focus on business objectives.
Objective-driven strategy
The third option describes a more radical shift in thinking and in the role of AI agents. Humans might tell the objective and autonomous agents to start executing it. In the process of executing agents take advantage of other agents described above. One might call the human-agent AI actor a Chief Agent orchestrating other AI agents in order to achieve the human-given objective. This third option is intriguing since it offers vast opportunities in business, and is not focused on data pipelines or simple market analysis. In this option, the abstraction level is high and so are the requirements for AI agents.
Another thing to keep in mind is that since the goal is given as a business objective, the result must be actionable. In other words, it is not acceptable to fall back on simple market analysis and nice dashboards. The resulting package delivered to the human (decision-maker) must be something that requires a decision - go or no go. If the human-given objective would be for example: "increase revenue by 20% during the next 12 months", the response from Chief Agent must contain at least:
Needless to say building such an autonomous AI agent probably is not feasible as one monolith. Most likely taking a ’microservices’ alike approach is more likely to succeed and split the package into multiple autonomous AI agents some of which are in the ’backend’ operating under the guidance of a Master AI agent interacting with the CEO. The autonomous agent at this level would be a multidisciplinary swarm of AI agents.
In one customer case, I proposed this approach to a CEO and he summarized the expected outcome from an autonomous agent:
”I would like to understand where the +20% should come from. Which customer segment, why do they need your data, how much are they willing to pay, how much sales do you need to get then? This leads to the necessary sales and marketing investment and thus to resources, the number of activities, and the accumulation of results.”
Another interesting aspect was raised regarding where new revenues should probably come from:
”For example, with brand marketing, nobody buys anything, you have to focus on customer acquisition. Likewise, sales must hold on to existing customers because, for everyone who leaves, two more must be acquired.”
This assumption and standpoint are valid since according to research it is 3,5 times more expensive to gain more sales via new customers than to utilize the existing customer base for add-on sales. Thus the primary guideline for autonomous AI agent building a playbook for a 20% increase should be default start from add-on sales rather than new customer acquisition unless ordered otherwise. Typically, these customers are recognized for the challenges they present in their purchasing behaviors. This highlights the importance of customer data and the potential it holds. By identifying unknown data about customers and then gathering it if needed (or purchasing it). Again empathy and knowing your customers, their processes, and their goals are needed in order to make value-based selling.
Also, goal setting is something to think about:
”So every time I see some goals ’to grow +X% something’ I get the feeling that we don’t know what we are doing. A more correct goal would be to ’increase the number of buying customers by +20 %’ which ultimately leads to +20% turnover, but that way you can demand specific actions from sales and marketing.”
From these couple of discussion snippets, you can see that it’s not a simple thing to build an autonomous AI agent to fulfill the needs of the CEO. What is fundamentally important to understand here is that AI is given a goal, an objective, and based on that it must be able to derive needed data to draw insights, must understand economic formulas, know the cost structure of technology to be used, and identify suitable marketing strategy with needed marketing tools and costs to mention a few.
The requirements above are exhausting but you should not feel overwhelmed with it. You can and should take the iterative approach and start from one corner (delivering the most value at the moment) and expand it as you go further.
Another thing worth mentioning is that given that we go further with the ChatGPT alike experience and take that more profoundly as part of the human-computer interaction principle, we need to train people in this skill. It is no surprise that the web is now filled with prompting tips, guides, and even commercial training. The speed AI development is now advancing is breathtaking even for the technical specialists in the area. You as businesspeople must stay on board and build an understanding of it in order to gain value and advantage from it. That is the key challenge we have
Co-Steward of CKAN Project | Executive Director at Link Digital
1 年Agents need agency and prior to that we’d aim to trust them to act in good faith, practice or measure. I’d put agency to one side entirely and first look at the intersubjective embeddings and learned patterns a model holds and be sure that these can be known and trusted in the context of where that model may be given agency. The story surrounding AI and its applications is great but the science matters way too much to be ignored. The social science, the political science, the philosophy of technology and our entanglements with it as people and communities. Before we chase down hallways and open every door, or as we do so, let’s try to keep balanced and cover AI as it applies to society, a pattern of technological development and also as one ‘other way’ of creating a reference model of intersubjective sentiments or referenced yet still intersubjectively contextualise facts. So my thought is let’s make sure AI agent strategy options are what comes only after trust in AI agents.