From: redpointai
The rapid evolution of generative AI (GenAI) has led to diverse models for enterprise adoption, ranging from consulting-led deployments to out-of-the-box products. While the space is rapidly evolving, certain AI product market fit use cases are demonstrating clear success and market fit today [01:08:00].
Current Successful Enterprise AI Applications
Currently, the most successful enterprise use cases for GenAI fall into general categories, with some specific vertical applications:
Healthcare Automation
In healthcare, a significant application focuses on making note-taking and form-filling easier for doctors [04:59:00]. This involves using passive listening microphones to capture doctor-patient interactions and pre-populate forms, significantly reducing the time doctors spend on administrative tasks [05:07:00].
Customer Support
Customer support is a widespread area where the technology is ready, and the need is very much present [05:24:00]. This application is moving quickly across various verticals, including telco, healthcare, and financial services [05:31:00].
Research Augmentation
Another prominent use case involves augmenting humans with AI agents that can perform extensive research tasks [05:47:00]. These agents can conduct a month’s worth of research in just an hour or two [05:51:00]. For instance, wealth managers in a bank could use an agent to quickly research and develop strategies to hedge against geopolitical events, providing robust research with citations for auditing by human experts [06:04:00]. This capability can make knowledge work tenfold more effective and productive [06:57:00].
Driving Factors for AI Adoption
The Role of Reasoning Models
The advent of reasoning models, like those available since Q4, has been a significant unlock [07:09:00]. These models can spend varying amounts of energy on problems of different complexities, making them far more capable than non-reasoning models [08:00:00]. Tasks that were previously impossible for models to accomplish with sufficient accuracy are now consistently successful with reasoning capabilities [16:24:00]. The ability to reflect, understand failures, and find alternative paths is a crucial advancement [16:47:00].
Data and Custom Models
While general models are increasingly powerful, custom models remain important for domains with proprietary or niche data not available on the public web [10:55:00]. For example, manufacturing data, customer transactions, or detailed personal health records are not typically found online [11:29:00]. Companies like Coher partner with organizations to create custom models for these specific domains, accessible only to them [11:43:00].
The process of training models has shifted significantly, with an overwhelming majority of data now being synthetic [15:15:00]. While human data is still necessary for evaluation and for generating small, trustworthy pools of data in specific domains (e.g., from doctors), synthetic data generation, enabled by the “chitchat” capabilities of initial models, allows for massive scaling [13:03:00]. In verifiable domains like code and math, results can be checked to filter out low-quality synthetic data [14:49:00].
AI Integration and Product Development
Effective enterprise AI adoption requires significant support due to the technology’s complexity [01:34:00]. AI agents need extensive access to company data (emails, chat, calls, CRM, ERP, HR software) to drive automation, which raises significant privacy concerns [01:58:00]. Each company also uses a unique “tapestry” of software, necessitating custom setups to integrate all relevant context into the model [02:43:00].
Coher’s “North” agent platform aims to solve the problem of custom application building by offering a consumer-grade product experience that is highly customizable for enterprises [19:01:00]. This platform allows customization of UI, branding, data connections, and tool integration, and even enables plugging in custom fine-tuned models like Llama [19:17:00]. This approach can accelerate a company’s product roadmap by 12 to 18 months [19:43:00].
Companies that build both the underlying models and the applications gain a significant advantage through vertical integration [20:55:00]. This allows them to optimize the models for specific customer use cases, leading to higher product quality [21:10:00].
Emerging and Future Use Cases
Back Office Automation
Looking ahead, more mundane back-office tasks are expected to come online as infrastructure like Coher’s North platform gets installed in enterprises [24:19:00]. This includes tasks in finance, legal, and especially sales [25:04:00]. For sellers, AI can provide intelligence briefings before meetings, compiling research on companies, strategic imperatives, key leaders, and internal conversation history [25:14:00].
Long-Term Consolidation
Initially, enterprises may adopt a scattered approach, with different teams purchasing their own specialized AI applications [26:37:00]. However, this will likely lead to an unsustainable maintenance burden due to disparate data source connections [27:03:00]. The long-term trend is expected to be consolidation towards a single platform plugged into all company data sources, capable of supporting diverse automation objectives [27:09:00].
Model Improvement Trajectories
The “scale is all you need” hypothesis is breaking, as diminishing returns from simply increasing capital and compute are observed [09:21:00]. The next step in AI development will require smarter and more creative approaches [09:33:00].
A crucial missing capability in current models is the ability to learn from experience, akin to humans who start as novices and become experts over time [08:46:00]. Models should be able to learn from interactions and feedback from users, remembering past conversations and preferences [09:00:00]. While this capability doesn’t fully exist yet, experiments are underway, likely involving integrating interaction history into a queryable database that provides context to the model [45:39:00]. This would create a more personalized and effective AI assistant that grows with the user [46:07:00].
! aspects
- Human Involvement: Humans are still essential for evaluating model usefulness and providing initial high-quality data [13:03:00].
- Architectural Diversity: While transformers have shown surprising longevity, the field is actively exploring new architectures beyond the current dominant ones [35:34:00].
- Global Access: For AI to be truly useful globally, models must understand and speak diverse languages and cultures, not just English [30:49:00]. Efforts like Coher’s Aya project involve thousands of native speakers contributing data in over a hundred languages [30:32:00].
- Specialized Foundation Models: Beyond general language models, future foundation models are anticipated for specialized domains like biology (e.g., cancer data), chemistry, and material science, requiring significant investment in data collection and model development [32:13:00].
AI Risk and Societal Impact
While short-term risks like misuse by bad actors or job displacement (if retraining infrastructure is insufficient) are concerns, existential “doomsday” scenarios are considered less pressing than near-term and mid-term challenges [48:39:00]. AI is viewed as an augmenter that will increase human productivity and supply in a demand-constrained world, rather than a mass displacer of jobs [47:55:00].