Real World AI

1. AI is transforming businesses, but responsible implementation is crucial

Responsible AI isn't just good for your business; it's good for society.

AI is revolutionizing industries, but its power comes with responsibilities. Companies must consider ethical implications and potential biases when implementing AI solutions. For example, Amazon's AI-powered recruiting tool showed bias against women, while the COMPAS criminal risk assessment tool exhibited racial bias. These cases highlight the importance of careful design and testing.

Responsible AI development involves:

• Identifying and mitigating biases in training data
• Ensuring transparency in decision-making processes
• Considering the societal impact of AI applications
• Implementing robust security measures to protect sensitive data

By prioritizing responsible AI practices, companies can harness the technology's potential while avoiding pitfalls that could harm their reputation and users.

2. Develop a clear AI strategy aligned with business goals

AI isn't a goal in and of itself. It's a very powerful tool and often transformational, but the strategy you develop should pursue a business goal.

Start with business objectives, not technology. A successful AI strategy begins by identifying specific business problems that AI can solve. For instance, Autodesk focused on reducing customer support resolution time by automating password reset requests, which led to a significant improvement in customer experience and operational efficiency.

When developing an AI strategy:

• Clearly define the business problem and desired outcomes
• Assess the potential impact and ROI of AI solutions
• Align AI initiatives with overall company goals and values
• Involve stakeholders from various departments to ensure buy-in

Remember that AI is a means to an end, not the end itself. By focusing on business goals, companies can avoid the pitfall of implementing AI for its own sake and instead drive meaningful improvements in their operations and customer experiences.

3. Choose the right "Goldilocks" problem to start your AI journey

If you can solve the first problem you attack and prove the impact AI can have, you'll have a much easier time getting support and resources to tackle the next 10 problems.

Start small, but impactful. The ideal first AI project, or "Goldilocks" problem, should be manageable enough to solve quickly while still delivering clear business value. This approach builds momentum and confidence in AI within the organization.

Characteristics of a good Goldilocks problem:

• Narrow in scope and easily definable
• Supported by sufficient high-quality historical data
• Capable of delivering quick wins and measurable impact
• Aligned with broader business objectives

For example, Autodesk's focus on automating password reset requests was a perfect Goldilocks problem. It was specific, had a clear business impact, and paved the way for automating 60 other use cases, significantly improving their customer service efficiency.

4. High-quality, diverse data is the foundation of successful AI

Garbage in, garbage out.

Data quality is paramount. The success of AI models depends heavily on the quality, quantity, and diversity of the data used to train them. Poor or biased data can lead to inaccurate or unfair outcomes, as seen in the case of the Apple Card's gender bias issue.

Key considerations for data:

• Ensure data is representative and covers all relevant use cases
• Implement rigorous data cleaning and preprocessing procedures
• Address potential biases in data collection and annotation
• Establish a sustainable pipeline for ongoing data acquisition and updates

Companies should invest in building robust data infrastructure and processes to support their AI initiatives. This includes developing clear guidelines for data annotation, implementing quality control measures, and continuously monitoring data quality throughout the AI lifecycle.

5. Build cross-functional teams to drive AI success

Machine learning products can't be developed by a team of data scientists alone. They require a team effort, and you need the team to work for the AI to work.

Diverse expertise is crucial. Successful AI implementation requires collaboration between various disciplines, including data science, engineering, product management, and domain experts. For instance, Figure Eight's reorganization into cross-functional teams led to a significant increase in product launches and improvements.

Key roles in a cross-functional AI team:

• Data scientists and machine learning engineers
• Software developers and DevOps specialists
• Product managers and business analysts
• Domain experts from relevant departments
• UX designers and user researchers

By bringing together diverse perspectives and skills, cross-functional teams can better identify AI opportunities, develop practical solutions, and ensure successful integration into existing business processes.

6. Create a successful pilot before scaling to production

A great pilot is intentionally planned and carefully scaled, like Autodesk's. Parameters are clearly defined: the pilot is limited in time, scale, and scope, and run in a controlled environment.

Prove value before scaling. A well-designed pilot project helps validate the AI solution's effectiveness and potential impact before committing significant resources to full-scale implementation. This approach allows for iterative improvements and builds confidence in the technology.

Key elements of a successful AI pilot:

• Clear, measurable objectives aligned with business goals
• Realistic timeline and resource allocation
• Defined success criteria and evaluation metrics
• Plan for scaling to production if successful

For example, OmniEarth's approach to helping California water districts reduce consumption started with a single county before expanding statewide. This allowed them to refine their model and prove its value before scaling up.

7. Adapt and secure your AI solutions for long-term success

Machine learning technology inherently changes over time as the data training it changes; you have to be able to adapt to deal with it.

Prepare for change and challenges. As AI solutions move from pilot to production, they often encounter new scenarios and potential security threats. Companies must be prepared to adapt their models and implement robust security measures to ensure long-term success.

Considerations for production AI:

• Implement monitoring systems to detect performance issues or anomalies
• Develop processes for model updates and retraining
• Establish security protocols to protect against adversarial attacks
• Create incident response plans for potential AI-related issues

For instance, Google's content moderation efforts for YouTube required continuous adaptation and improvement, combining machine learning with human review to address evolving challenges in identifying inappropriate content.

8. Foster an AI-driven culture across the organization

Every department's AI use cases will be as different as the departments are themselves, so every leader has to be trained to identify problems in their own department.

Cultivate AI literacy company-wide. To truly lead with AI, organizations must develop a culture where every department understands and embraces AI's potential. This involves educating leaders across the company to identify AI opportunities within their domains.

Steps to foster an AI-driven culture:

• Provide AI literacy training for all employees, especially leaders
• Encourage cross-departmental collaboration on AI initiatives
• Align incentives to promote AI adoption and innovation
• Share AI success stories and learnings across the organization

Companies like Amazon and The New York Times have successfully embedded AI throughout their operations by cultivating this AI-driven mindset across all levels of the organization.

9. Implement robust data governance and quality control measures

Data, as we mentioned earlier, is the new IP. It's an incredibly important asset for your company, and its use has to be managed accordingly.

Govern data like a valuable asset. As AI becomes more integral to business operations, effective data governance becomes crucial. This ensures data quality, security, and compliance with regulations, while also enabling efficient use of data across the organization.

Key aspects of data governance:

• Establish clear policies for data collection, storage, and usage
• Implement access controls and security measures
• Ensure compliance with relevant regulations (e.g., GDPR)
• Document data lineage and transformations
• Create processes for data quality assurance

For example, British Airways' insufficient data governance led to a massive fine for a data breach. Implementing strong governance practices can help companies avoid such pitfalls and maximize the value of their data assets.

10. Continuously monitor and update AI models to prevent drift

It's a good idea to refresh models at least monthly (if not more, depending on your use case; some models are updated as often as every day) to account for data drift.

Adapt to changing realities. AI models can become less accurate over time as the real-world conditions they were trained on change. This "model drift" can lead to poor performance or even harmful outcomes if not addressed.

Strategies to combat model drift:

• Implement regular model performance monitoring
• Establish processes for model retraining and updates
• Create alerts for significant changes in input data or model outputs
• Conduct periodic reviews of model assumptions and relevance

Facebook's content moderation challenges with Facebook Live demonstrate the importance of continuously updating AI models to address new scenarios and changing user behavior. By proactively monitoring and updating AI systems, companies can ensure their solutions remain effective and relevant over time.

Last updated: August 11, 2024