Enterprise Intelligence is the new definition of “Capital”
We, in the industrial and business world, are well past the labor arbitrage and cost arbitrage business cases. It’s a fabulously disruptive world of connected opportunities of “Algorithmic Arbitrage” enabled by AI-augmented Human Intelligence.
While the basics of the classic Adam Smith model still hold good, the definitions of Land-Labor-Capital is fundamentally changing in this 4th industrial revolution:
- Land is now meant for AI data centers and manufacturing, “CfX” (Chips for Everything), plus power generation infra to fulfill the tremendous hunger for energy of the DCs
- Labor means new skills, from creative design to algorithms and agentic robotics technology developments.
- Capital has moved beyond the flow of money, it has become the flow of Intelligence, in fully connected network economies, from physical goods to services supply chains to crypto blockchains and NFT.
With intelligence emerging as the new capital, the Trigent Tech Forum session 2 discussions spun around the realities of the operating model of Enterprise Intelligence, where the rubber hits the road. Leaders who have been early movers and disruptors in this next-digital business world, shared practical examples and challenges, from their own experience and industries.
Stay grounded on AI: Avoid Gen-Washing and Agent-Washing everything AI
The practical implementation of Agentic AI in terms of realistic use cases and value, were the anchoring elements of the Trigent Tech Forum session 2 discussion. Starting with highly regulated, data-sensitive sectors, leaders brought interesting perspectives from some of the most innovative hospitals in the US, to digital insurance technology and toolstacks. AI applications were discussed in legacy manufacturing behemoths to deep research institutions.
Challenges with data and knowledge quality, quantity, usability, and availability, plus privacy and security issues, were consistently highlighted. These challenges exist in pre-curated data
availability for RAG models and other fine-tuning techniques.
Quality data and knowledge items are also critical for test-time and inference compute stages, which are imperative for agentic AI usecases i.e. for machine reasoning and goal-oriented
systems, state-space machines.
Along with the data challenges, the critical hurdles of legacy enterprise systems and databases were discussed, e.g. when these systems had to be integrated into genAI and agentic AI
usecases.
In context of data governance and on AI regulations, discussions involved open exchanges of ideas regarding how AI governance is evolving, and who should have the controls and
accountability, in an enterprise scenario.
There were also discussions around measuring the tangible ROI of AI initiatives, and the critical question of adoption at scale, which is dependent on trust building.
Summarily, many service providers, in their over-eagerness in rapidly mass-producing agenticAI solution stacks, are “gen-washing” and “agent-washing” the nuanced requirements of different enterprises.
Providers often paint all AI applications in different sectors and sizes, constraints and realities, with the same brush. In this muddle, creating a set of simple effective agentic AI governance principles has become more critical than ever.
Agentic AI governance: Emerging challenges without proven answers
Governance is a critical lever to build trust in AI. From employees to customers, users, developers, to regulators and auditors, AI governance best practices need to be built, proliferated and practiced on a scale. This is because every stakeholder persona needs a clear understanding of the outcomes targeted and achieved by their AI leverage, in a responsible and accountable manner.
AI governance can be flexible and based on simple principles e.g. Asimov’s Three Laws of Robotics, to a general guideline for example in case of DevOps the principles of agile manifesto.
Enterprise AI leaders can start with 3 simple principles
- Honesty (for building trust) of AI models: AI models prioritized on explainability, reference ability or factuality of responses, transparency, feedback loops and control of
reasoning. Most of the maturing LLMs are working aggressively in this space, e.g. Gemini, GPT latest models are showing their reasoning and ‘thinking’ steps and interacting and allowing user guidance and feedback. Google’s latest Alpha Evolve shows new capabilities e.g. it suggests a fresh way to speed up complex matrix multiplication-type mathematical problems. But this also shows the steps and the code it is developing, to achieve this faster technique.
- Harmlessness: Enterprise AI models and solutions should follow a “Do no harm” strategy. A simple way to execute it is to expose the early-stage pilots and POCs to a group of users like beta-testers, to rigorously test the effects of low-factuality responses and hallucinations. Their feedback should be back-propagated to the model and usecase design stages.
- Human augmentation: Outcomes of AI models and solutions should augment human quality of work and productivity and impact. This is essential to reduce the general ‘fear’ in employees’ minds that AI agents will replace them. The AI story-telling technique needs to mature to a “digital twin” equivalent in an enterprise, that these agents augment human employee productivity and quality, which ultimately helps them getting rewarded.
Asimov’s Three Laws of Robotics
Regarding the “do no harm” principle, there was an interesting discussion in the session, on whether AI-powered productivity for automation leading to job losses, are considered harmful or not.
Multiple solution ideas also emerged including the third principle, e.g. up-skilling workforce to augment human productivity and repurposing human agents for more value-adding/ creative/relationship-driven, more “humane” work.
The STEP (Social-Technological-Environmental-Political) impact of enterprise AI in different sectors varies differently. This variance is seen even within sectors. For instance, small healthcare/ healthtech research firms use AI for drug discovery and new molecules design usecases that are different. They require varied regulatory guardrails, from the production-line
pharma AI-automation-agentic applications.
Similarly, a small regional community bank’s AI leverage use cases e.g. agentic AI for CRM, are very different from say Agentforce-Gemini implementations in large banks.
Also, even in current state with AI policy documents evolving across US, EU, GCC, Australia, Singapore, India et al, sovereign cloud, data infrastructure and AI are not mature to the extent
that there are benchmarks that can be fully applied to most application scenarios.
For example, synthetic data generation with genAI for sparse datasets, is technically a tried and tested application. But, using synthetic data for in-prod AI usecases in an enterprise, still doesn’t have explicit and mature guardrails.
Using open databases for pilots and POCs makes practical sense when enterprise data in production is hard to get, for build and test. But again, productionizing them requires measured
steps, in absence of clear guidelines.
Action items Monday Morning: Strike a Golden Balance between Shadow AI and Responsible Citizen AI
Navigating “shadow AI” e.g. employees using GPTs from different models in their regular functions, without adequate transparency, explainability and boundaries, can create havoc for
data security and IT controls. But completely stopping the usage reduces the AI leverage of the company overall. Hence, creating a balanced pathway of agentic AI adoption is necessary:
- Enterprise AI leaders can ensure that a few priority, high-impact, moderate-risk applications are tested, and control guidelines are documented and communicated across the organization, from board to the first-line workers.
- These applications can be tested by power-user citizen developers e.g. specific business process and function teams within the enterprise.
- AI CoE’s can encourage responsible citizen developers on LCNC development and experimentation platforms e.g. agentic AI development on ADK, CrewAI etc., using enterprise-ready LLMs like Gemini, anthropic, perplexity for deep research on their regular work applications.
- Systems can be established to train the AI-eager employees on basic fine-tuning of models, entropy management for hallucinations vs creativity, prompt engineering- single prompting and multi-turn feedback loops for reasoning, etc.
- If enterprise datasets are sparse, sensitive and high-security, then the first-user teams can build, test and share experimental sandboxes with precurated data, synthetic data or public databases like image net to UCI datasets and Kaggle/ GitHub datasets.
Challenges are enormous but so are the opportunities, when it comes to generative, agentic, robotic AI leverage in any enterprise context. Taking clear-headed, logical, practical steps, with simple but tested guardrails, is the only way forward.
