The heightened interest in generative artificial intelligence (AI), specifically in AI agents and large language models (LLM), is related to their promising ability to transform companies by making them more efficient and increasing their productivity.

Many surveys have revealed that CEOs, COOs, and CFOs intend to invest in this type of innovation in the coming six to 12 months in order to reduce their operational costs through the automation of certain activities or business processes.

According to one survey published in 2023 by McKinsey¹, 75 percent of generative AI’s biggest impacts will be felt in four business areas: customer operations (service, support, and customer experience), marketing and sales, software engineering, and research and development. The subsequent improvement in productivity is estimated to be between 0.1 and 0.6 percent per year until 2040.

The enthusiasm surrounding AI is palpable, but investments are needed by many companies in order to realize their business aspirations. As with any such initiative, these aspirations cannot be realized without a good data and artificial intelligence strategy, not to mention proper execution.

This article will cover four major aspects that should be part of your strategy.

1. The selection of your use cases (relevance and impact on the productivity of the company or within the business field)
2. Your company’s capacity to integrate innovation into its processes
3. The robustness and reliability of your data life cycle management, from acquiring data through to their consumption
4. The presence of data and AI application governance

The operation of artificial intelligence (AI) agents

For those who are less familiar with the concept of an AI agent—also referred to as generative AI agents (LLM) or agentic AI—it is software or an application that can perform a task, an action, in an autonomous way. It uses instructions provided in a conversational manner through a large language model (LLM), such as ChatGPT, Gemini, or Claude, to understand and execute tasks. It can perform simple or complex tasks and its goal is mainly to automate these tasks. 

The anatomy of an agent includes: 

• a reasoning process;
• tools, which are features that enable agents to complete the requested tasks; and
• a model: ChatGPT, Gemini, Claude, or some other.

The relevancy of use cases

It is unrealistic to think that generative AI agents will resolve all issues related to marketing performance, sales, and customers. As of today, the most promising use cases are related to the following functions:

Customer experience

• Helping customers in self-serve mode (AI agent and chatbot)
• Summarizing customer complaints and comments
• Assisting customer service with resolving problems
• Improving the privacy of customer information (users and synthetic data)
• Personalizing customer experience (CX) and setting up advanced personalization for content, products, and services at every touchpoint

Marketing

• Creating ad and informational content
• Accelerating market analyses, summarizing consumer and market trends
• Improving analysis of structured (customer records and transactions) and unstructured (text, image, audio, and video) data
• Detecting trends in the data (customer understanding and segmentation) and recommending actions or tactics (data activation)

Obviously this is not an exhaustive list. Selecting the right use cases is just one piece of the puzzle—there is another cultural element that shouldn’t be overlooked: Is your company ready to integrate this innovation?

A culture of business experimentation and innovation

Generative AI and agents are in their initial stages. To allow companies to adapt them to their business processes, these tools need to be customized and contextualized via several techniques: prompt engineering, retrieval-augmented generation (RAG), fine-tuning, and model evaluation.

This process of customizing generative AI needs to occur through the testing and refining of algorithms, which is where the importance of having a culture of experimentation and innovation comes into play. Putting this type of technology into production without a trial, testing, and evaluation phase considerably increases the likelihood of not achieving your objectives.

Most technology vendors (Google, Salesforce, Snowflake, Databricks, Amazon, and Microsoft) will lure you with promises of how simple these solutions are and their positive (financial) impact on your organization. However, the reality is more of a mixed bag.

The art of integrating this type of technology into your business lies in the combination of many factors:

• availability and quality of data
• organization’s technological maturity level 
• organization’s technological expertise (talent that can make use of these innovations)
• relevance of the use cases these technologies will be supporting
• business expertise they will support (the capacity to create value)
• organization’s capacity to integrate these technologies into business processes quickly and with the least amount of risk (which relates to the importance of testing)

There is no AI without data. Understanding the life cycle of your marketing data is a crucial prerequisite for the eventual success of your initiative.

The robustness and reliability of data life cycle management

There are two categories of activity that are essential to the data life cycle and which exist in every company:

1. The collection and storing of data
2. The analysis and activation of data

The collection and storing of data

To be able to customize and roll out (in a marketing context) generative AI models and agents at an industrial level (productivity), you need to have quality data for every field (for example, for marketing). The data collection platform you employ—whether it’s Google Analytics, Adobe Analytics, Snowplow, Amplitude, or another customer data platform (CDP) from a different vendor—needs to be able to export your primary data to:

• cloud-based data warehouses: Google BigQuery, Snowflake, RedShift, Databricks or Azure Synapse; or
• data repositories in the cloud: Google Cloud Storage, Azure Blob Storage, or Amazon S3.  

The preferred infrastructure for generative AI models is the data lake, for storing unstructured data, and vector databases. There is also the recent appearance of the data lakehouse concept (the open table format), which is a hybrid between a data warehouse and a data lake, with some examples being Google BigLake, Databricks Delta Lake, and Snowflake with its Open Catalog.

Data collection is critical to obtaining quality data, since these data will in turn influence the quality and performance of the AI models. Data centralization by business domain (data mesh) is also essential. Agents and large language models perform much better when they are contextualized and trained with data from a specific field, such as marketing.

Why every company should have a data strategy

Data analysis and activation

Depending on the use case, simple or advanced data transformation will be necessary—while data transformation is not covered in detail in this article, this step is an important one. This transformation can be employed to create data models for supporting knowledge graphs, which enable better performance by LLMs.

Data analysis is supported by the algorithms within the tools employed, such as ChatGPT, Claude, Gemini, etc. AI agents can also analyze data based on their features. RAG (retrieval-augmented generation) technologies are used to ingest external data, originating from your company databases, into different tools. Prompt engineering and fine-tuning refine the degree of customization for your generative AI model.

The strength of AI agents lies in their capacity to perform an action in addition to an analysis. If you have the internal skills, you can build your own AI agent. You could also resort to using technology vendors (Google, Amazon, Salesforce, and Snowflake), who offer agents that are already integrated into their applications to facilitate productivity gains by their users. Building an AI agent from scratch, by writing code, is no small feat: You’ll need to have that internal expertise available or take on a partner.

There are a few vendors who have facilitated the use and creation of DIY agents that do not need advanced expertise, such as Salesforce Agentforce, Google Agentspace, and Snowflake Intelligence (data agents). Many of these products are new to the marketplace and at an experimental stage. 

The goal of these applications is to promote productivity by reducing costs and increasing operational efficiency. These agents execute several activities, such as:

• Making recommendations based on data from a company or business sector
• Analyzing complaints or comments and offering solutions, whether directly to the customer or to the customer support team
• Helping customers, staff, or students with their requests for information by customizing responses based on the information provided during the conversation
• Facilitating the onboarding of customers, staff, or students
• Supporting marketing teams (whether media or marketing automation) in the creation of campaigns or campaign briefs by suggesting audiences and combinations of strategies and metrics for tracking
• Analyzing marketing campaign performance and recommending tactics based on their insights
• Helping customers or prospects discover products and services on a website or mobile application through conversational interactions
• Personalizing the conversion stage based on information provided by the customer:  shopping cart, appointment form, information request form, etc.

Customization and specialization of the agent based on the field of activity are critical if you are hoping for actual productivity gains, hence the importance of exploiting your data by business sector, from data collection through to consumption.

The final section of this article is highly important: How can you ensure that these AI agents don’t become a risk to your company?

Data and AI application governance

Good data and AI application governance is required for the success of your initiatives.

Data governance needs to ensure the high quality and accessibility of data while also making them secure. Collecting data on individuals must be performed transparently and with their consent. Every measure to secure both personal and company data must be put in place. Processes and technologies for validating and tracking data quality are required before these data can be used by AI systems.

AI agents must produce reliable output for users. Processes to evaluate output quality must be continually implemented. Biases and hallucinations must be detected quickly and followed by mitigation measures. Tracking the performance of models and agents is required to ensure staff and customers adopt this innovation. As with data, AI agents and models must be protected against internal and external attacks.

Keeping pace with legislation and regulations is a requirement for every company that intends to invest in this type of initiative.

Conclusion

While the promise of generative AI and AI agents is attractive to executive teams, a good data foundation (collection, storing, and engineering) remains a challenge for many companies. Companies that invest today in acquiring and transforming unique, high-quality data will obtain a better return on investment (ROI) on the deployment of AI agents for marketing. 

If you would like to discuss a data and AI strategy for the purposes of strengthening your customer-centric marketing and increasing the efficiency of your marketing operations, don’t hesitate to get in touch with us.

References

¹  Chui, M., Hazan, E., Roberts, R., Singla A., Smaje, K., Sukharevsky, A., Yee L. and  Zemmel R. (2023). The economic potential of generative AI: The next productivity frontier.

²  Incorrect, inappropriate, or fictitious content that is generated by generative artificial intelligence and presented in a factual way, as if the information were genuine or consistent with the original query (translation of definition provided by the Office québécois de la langue française).