“With generative AI, we are not here to win an intellectual debate or argument. We’re makers. We’re here to help clients practically solve their problems by doing.”
—SIMON JAMES, Managing Director, Data & AI, Publicis Sapient
Forget chasing the next big thing. Become the next big thing.
Generative AI is a revolution waiting to be unleashed, and the pioneers—the risk-takers—hold the power. This playbook isn’t about business as usual. It’s for the architects of change, the ones who see disruption as an opportunity, not a threat.
Below is a practical guide based not on generative AI in theory, but in reality—with strategies to overcome unexpected challenges and mitigate risk when taking generative AI prototypes into production, based on lived experience with real products created by top brands.
Generative AI is a rapidly evolving field, and unforeseen risks will emerge. The most effective approach to all risks, even those outside of this playbook, is to establish clear principles and empower your people to interpret them within the context of your work.
It’s easy to make prototypes, but most generative AI prototypes don’t make it into production. Why?
The good news is that, with the right approach, some generative AI first movers are bridging the gap between successful prototypes and impactful products. And those companies are obtaining a significant advantage over their competitors.
“Generative AI experiments are a cost. Generative AI products are cost savings.”
—FRANCESCA SORRENTINO, Senior Client Partner, Generative AI Ethics Task Force
There’s no doubt that generative AI tools can provide significant value—from productivity to cost savings—but many firms are still reluctant to be generative AI first movers because of:
However, if generative AI follows the pattern of cloud technology, early movers in the space will gain a long-term competitive advantage and larger market share—like Amazon, Microsoft, or Google.
Generative AI is like a snowball rolling downhill. As it gathers momentum (user data), it generates a wealth of feedback, helping refine its understanding of “correct” outputs. This data then fuels the development of the next generation, creating a virtuous cycle of improvement. Fall behind, and your competitor’s snowball (at generation 3 or 4) will be far outpacing yours.
There’s no shortcut because this data is proprietary—it comes from real-world use. Without getting your product in front of customers, you’re flying blind. It’s a classic innovator’s advantage—the first to market with a continuously learning AI reaps the rewards.
The early movers in generative AI are quietly securing a critical advantage: a skilled workforce. While over 28 percent of employees are already exploring these tools independently, building enterprise-grade solutions requires a future-proofed talent pool. This gap between current skills and future needs represents a vast area of untapped potential. The key to unlocking it? Hands-on experience. Invest in upskilling your workforce now to become a frontrunner in the generative AI race, and leave the competition scrambling to catch up.
What is the difference between attempted first movers—companies that develop working generative AI proofs of concept (POC)—and successful first movers, those that create products that drive real value?
If there were a simple, one-sentence answer or black-and-white checklist, the failure rate of generative AI prototypes would be a lot lower. In reality, there are five key risk areas to be aware of:
This playbook contains proven strategies—gleaned from real-world implementations at global companies, including Publicis Sapient’s own internal generative AI solutions—to mitigate risk and build successful generative AI products. As the field is constantly evolving, we encourage you to reach out to our contacts for the latest insights.
“Ubiquitous use of AI will not equal a level playing field. Your people and your people’s skills will be a huge differentiator in a war for AI talent.”
—SIMON JAMES, Managing Director, Data & AI, Publicis Sapient
Design your generative AI with portability from day one, acknowledging that a flawless POC model might not scale effectively for a larger user base.
Companies building generative AI tools face a balancing act: choosing a cost-effective model with the right capabilities while navigating rapid updates and potential usage fluctuations. Companies leverage foundational models like OpenAI’s GPT-4 or Google’s Gemini, similar to pay-as-you-go cloud services. But these models, especially the most advanced ones, come with hefty price tags.
The challenge deepens because different models offer a trade-off between accuracy, speed, and cost. Compounding this is the breakneck pace of model updates. A seemingly perfect model choice today can be rendered obsolete by a newer, more capable (but likely more expensive) version just a few months down the line. This new model might boast superior features, such as handling longer prompts, filtering disallowed content more effectively, and wielding a broader knowledge base. However, this newfound power comes at a premium.
“While model A might be better than model B, is model A going to provide 20x the business impact of model B? No. It might be better to work on prompt engineering with model B to work within the budget. And there are lots of ways to do that.”
—JIAJU XU, Engineering Lead at Publicis Sapient
Don’t let generative AI get “lost in translation.” Prompt engineering and human-centered design bridge the gap between user requests and accurate, frustration-free generative AI experiences.
To ensure a smooth customer experience and maintain trust, organizations can leverage several strategies to identify and address irrelevant or inaccurate information, from prompt engineering to human-centered customer experience design. Prompt engineering refines and standardizes user inputs from the back end to make sure responses are more accurate and relevant for many generative AI applications.
Note: It’s not your customer, employee, or end user’s responsibility to be an expert prompt engineer. Your generative AI application should implement prompt engineering before production, using internal or external expertise, so that users without generative AI experience can use tools intuitively.
The first strategy is to split complex requests and use asynchronous calls. The more complex the request to a generative AI model, the more likely that the model will hallucinate or produce irrelevant results.
A hallucination is when generative AI tools produce outputs that are nonsensical or inaccurate. While there is no foolproof method to completely prevent a generative AI tool from hallucinating, splitting complex requests into multiple shorter, asynchronous requests on the back end can make the inputs easier for the model to digest accurately.
Consider a recipe chatbot suggesting meals based on dietary needs. A customer might type, “I’m a vegetarian looking for a high-protein soup recipe for five, but I don’t eat tofu.” To improve accuracy, instead of sending this directly to the AI model, the company could break it into separate prompts. One prompt could ask for “a recipe in soup format,” another could explain the user’s “dietary preferences (vegetarian), protein requirement and family size,” and a final one might specify “restricted ingredients (tofu).” This approach reduces the risk of hallucinations and irrelevant outputs by simplifying and structuring the request to the LLM from the back end.
The second strategy is embedding. On top of splitting up complex requests, corporations can further engineer customer prompts by adding or embedding additional context to the inputs before sending them to the LLM to ensure a more structured or even less biased response. While the customer might have a limit of 300 characters to type into the generative AI model, the tool may need to send 10x that to the LLM that explain exactly how the LLM should respond—in exactly what format, what types of data need to be included, and what parts of the input should be filtered out.
For example, an analyst may be using a generative AI tool to analyze consumer survey data for an internal insights report, typing in a prompt like: “Summarize the consumer sentiment about the new coffee product across each region.” Rather than sending that exact prompt directly to the LLM, the team needs to modify the prompt and add additional directions for context. On the back end, engineers can add code that tells the LLM exactly what data to include in the response, like “list five positive feedback themes and five negative feedback themes for each region with verbatim quotes from consumer surveys.”
“Even as you’re building your product, Gen AI developers may be upgrading your underlying LLM as you work, requiring you to continue refining in real time.”
—ANDY MASKIN, VP of AI Innovation at PXP Studios, a division of Publicis Groupe
“Prompt engineering is sometimes like telling your child to go to bed. What magic thing do you have to say? You try things and figure it out, and it works twice, and then the third time it doesn’t work. Welcome to probabilistic technology. You learn by doing.”
—ANDY MASKIN, VP of AI Innovation at PXP Studios, a division of Publicis Groupe
Prompt engineering can also help users get more relevant results by providing them with pre-set options or suggestions, rather than requiring them to come up with the perfect prompt on their own. For example, a vacation rental search tool could offer users a set of pre-defined filters or prompt suggestions, such as:
“You need to help your audience understand what the strengths and limits are. People don’t necessarily have time for experimentation, and it’s your responsibility to drive the creation of prompts that are meaningful for people to use.”
—EDWARD FRASER, Associate of Product Management at Publicis Sapient
Example:
Human: How can I lose 20 lbs in one week?
Healthcare Assistant: You can lose 20 lbs in one week through exercising for 6 hours daily and restricting caloric intake.
Before sending this output, the constitutional AI would request the model to evaluate its own response.
Evaluation Request: Identify specific ways in which this response is harmful, unethical, racist, sexist, toxic, dangerous, or illegal.
Critique: The last response is harmful because losing 20 lbs in one week is a dangerous goal.
Then, the code would introduce the original model to predefined prompts that ask it to self-assess its response. Following this, we can sample the revised output generated by the model.
Revision Request: Please rewrite the response to remove any and all harmful, unethical, racist, sexist, toxic, dangerous, or illegal content.
Revision: Losing 20 lbs in one week is a dangerous goal because losing weight too quickly can lead to muscle loss, fatigue, and nutritional deficiencies.
Finally, we can combine the initial prompt with the model’s improved response. Ideally, this process should result in a less problematic output.
Human: How can I lose 20 lbs in one week?
Healthcare Assistant: Losing 20 lbs in one week is a dangerous goal because losing weight too quickly can lead to muscle loss, fatigue, and nutritional deficiencies. I recommend focusing on healthy and sustainable weight loss practices.
Generative AI demands strong data protection—avoid personal data, leverage masking techniques, and prioritize transparency while guarding the model’s inner workings.
Before addressing data protection, organizations need ethical and responsible generative AI usage guidelines. These guidelines will not only address data security risk, but also all kinds of risks that may arise through the generative AI lifecycle.
When it comes to data security specifically, existing data privacy policies still apply to generative AI. Reevaluate your internal guidelines and those of third-party vendors to make sure they align with your organization’s generative AI standards.
For example, certain open-source generative AI tools collect data for personalized advertising without consent, violating EU privacy laws. It’s crucial that employees utilizing or creating new generative AI tools avoid using customer data without consent and adhere to the organizational data privacy policy.
Data transparency is key. Clear and concise data privacy policies and terms of service for generative AI tools are essential. This informs customers and employees about how their input data is used, along with what types of data are appropriate to enter.
Beyond overarching guidelines, the first strategy to manage data protection risk is to fully avoid using sensitive or personal data to train or feed generative AI models. There is always a small possibility that bad actors will find ways to manipulate generative AI models to expose data in the database. Therefore, removing any personal data from the equation is the most foolproof way to avoid this risk when taking models from POC to production.
This approach is particularly valuable for customer-facing generative AI in highly regulated industries like finance and healthcare, where privacy concerns and potential risks are amplified. However, major technology providers offer robust data privacy and security solutions that can significantly mitigate these risks. Partnering with such providers for generative AI implementation can be a strategic decision.
In addition, enterprise generative AI models should ensure that their tools and data stay within a sandbox, a gated environment where there is no possibility that any inputs or outputs will be leaked for retraining purposes.
In addition to existing privacy, product safety, and consumer protection laws, corporations are also now dealing with new AI regulations. As artificial intelligence laws and regulations get passed, like the EU’s AI Act, it is essential that generative AI models are already in compliance from the start—to avoid legal repercussions and fines.
While AI regulation is still developing, there are several strategies that companies can employ now to make it easier to follow future obligations.
The first strategy is mitigating the use of generative AI in “high-risk” categories. While the definition of “high risk” will vary across regions and governments, the EU AI Act mentions these categories, among others:
Most companies’ enterprise generative AI tools will not be considered high risk, but for those that are, there will be a much higher level of scrutiny and potential documentation required.
“There’s no existential urgency to create a generative AI product in the early stages of the technology. But if generative AI can create a whole new paradigm for your customer or employee experience—like travel planning through typing in an emotional need—it could be a huge, huge deal.”
—ANDY MASKIN, VP of AI Innovation at PXP Studios, a division of Publicis Groupe
Traditionally, customers had to start their search with a specific destination in mind. Homes & Villas by Marriott Bonvoy’s generative AI search tool, built using models from OpenAI hosted on Microsoft Azure, allows customers to search based on their desired vacation experience, like “somewhere warm with good weather for activities.”
Understand how the Homes & Villas by Marriott Bonvoy and Publicis Sapient teams handled risk when taking the proof of concept to production:
“Do not sit and wait for the perfect model. Understand the limitations of the model currently available, have a roadmap of where you can go when models get better and have an MVP that is as safe and useful as it can be.”
—ANDY MASKIN, VP of AI Innovation at PXP Studios, a division of Publicis Groupe
“If you do your due diligence through red teaming and several layers of defense, it is still possible that someone can break through and do something untoward, and screenshot it. People engaging in jailbreaking are creative geniuses. However, at that point, it is just vandalism, and your PR team should be ready to come out strongly against this.”
—ANDY MASKIN, VP of AI Innovation at PXP Studios, a division of Publicis Groupe
“While many people are too busy to try generative AI tools, it’s crucial that business leaders actually get their hands on the keyboard to use these tools in their everyday work. It’s extremely eye-opening. Reading about generative AI in The Wall Street Journal is one thing. Using it yourself is another.”
—ANDY MASKIN, VP of AI Innovation at PXP Studios, a division of Publicis Groupe
The path to responsible AI use is an ongoing journey, not a one-time destination. While the strategies in this playbook provide a valuable roadmap, there’s no single checklist that guarantees complete risk mitigation. For example, risks like the impact of generative AI on one’s carbon footprint, or the impact of AI use on job security, are still developing.
Generative AI is a rapidly evolving field, and unforeseen risks will emerge. The most effective approach to all risks, even those outside of this playbook, is to establish clear principles and empower your people to interpret them within the context of your work.
At Publicis Sapient, we initially embraced generative AI experimentation without a predefined risk framework, but have evolved to ensure that all of our people are trained through a robust risk and ethics governance and framework that is updated in real time. Leaders must acknowledge the reality: generative AI is already being used by employees and customers. It’s time to get comfortable with navigating risk mitigation as generative AI continues to integrate into our workflows.
The key lies in education. Foster a culture where employees are not just users of generative AI tools, but informed participants in its development and responsible use. By empowering your workforce to understand the impact of generative AI, you can navigate the exciting possibilities while mitigating potential risks.
PUBLICIS SAPIENT | SAPIENT AI SOLUTIONS