From Handlooms to Agents: Where Value Moves When Intelligence Gets Cheap
When intelligence becomes cheap, where does value move? That was the question Sandeep Srivastava put to the room at this month's Beyond the Deck, our monthly webinar with domain experts on emerging technologies and their commercial potential.
By SilverX Fund
Sandeep has spent 25 years on both sides of that question. He studied computer science at IIT Bombay, started his first campus company in 1998, and went on to build Mintoos (an early HFT venture) and Zycus. He then spent nearly two decades at Microsoft, shipping more than 15 products across OneCare, Office Live, Office 365, SharePoint and Visio, incubating Kaizala, and later leading product strategy for M365, Project and Visio. In 2025 he founded krsna.ai.
His answer, in short: as models commoditise, the money moves to the system around the model. Here is how he got there.
The Zycus lesson: sell outcomes, not intelligence
Sandeep's thesis started with a failure. When telcos pulled out of his first startup, Myzus, he joined Zycus in 2003 with one question: what do successful companies still sell when budgets tighten?
The answer was verified savings. Zycus ran a loop that looks a lot like today's agentic stack:
1. **Data** \- analyse Fortune 500 procurement spend and cost inputs with AI.
2. **Evaluation** \- check accuracy and preserve provenance, with human review teams online and offline.
3. **Domain validation** \- consultants confirm the analysis before any savings are recommended.
4. **Outcome fees** \- cut procurement or production costs and charge a fraction of the savings.
Every correction and every realised outcome fed the next analysis. Intelligence was never the product. The verified result was.
When a capability gets cheap, the system around it gets valuable
Sandeep framed today's shift with two older ones. Each time a core skill was automated, value did not vanish. It moved.
The investor takeaway follows directly: don't invest in the capability. Invest in the system around it.
What enterprises actually want from agents
Sandeep's customers are moving past off-the-shelf AI to build their own learning loops. Compute and models are the base. Workflows, controls and feedback are where value gets created. Five demands keep coming up:
* **Adoption is not transformation.** Measure workflows adopted and outcomes delivered, not seats.
* **Organisational productivity.** Complete whole processes across people, agents and applications.
* **Interoperability.** Work inside the enterprise systems customers already run.
* **Agent lifecycle governance.** Discover, test, own, monitor, recertify and retire agents.
* **Reliability and measurable ROI.** Track quality, completion, cost and business impact.
He defines an agent as digital labour with three properties: **agency** (it can propose or take an action), **autonomy** (it can pick the next step within a defined scope) and **authority** (it can change data, communicate, spend or approve). Authority is the one that matters commercially, and the one that needs the most care.
Inside a workflow, agents play one of three roles:
Across all three, value is measured the same way: quality, completion and business impact.
Where the startups are, and why India
Sandeep mapped the opportunity into four zones, all sitting on platforms:
* **Verticals** - claims, compliance, procurement and revenue operations.
* **Trust** - identity, policy, evaluation, provenance and escalation.
* **Data** - businesses where each completed action improves the next decision.
* **Physical AI** - logistics, manufacturing, energy and health operations.
Then he turned the question to India: which Indian workflows need that whole system, not just a model? His case is that India's population-scale needs make it a product origin, not just a market. He grouped the openings in three bands:
The common thread: companies that own the workflow end to end compound data, integration depth, trust and distribution at the same time.
What he would back, and what he would avoid
Sandeep's checklist is short:
1. **A narrow workflow with material economics** - a named buyer, high frequency, a measurable baseline and a clear loss rate.
2. **Authority with guardrails** - bounded permissions, human escalation and traceable decisions.
3. **A compounding asset** - every outcome adds data, workflow position or trusted distribution.
He gave examples of workflow businesses that pass:
Legal, support and salons fit the same pattern. What he would avoid: generic wrappers, hidden human costs, agent marketplaces without distribution, and time-saved claims with no outcome economics behind them.
The free-model test
Sandeep closed with one question every founder and investor should ask: **if the best model became free tomorrow, why would this company still win?**
If the answer is the model, it won't. If the answer is that the company owns the workflow, the authority to act, the learning loop and the accountability for the outcome, it will.
That test maps closely to how we think at SilverX. Our Non-Disruptible Technologies thesis looks for companies whose advantage survives the next wave of commoditisation. In agentic AI, that means trusted systems of action, built for workflows where India's scale is the advantage, not the obstacle.
*Beyond the Deck runs monthly. If you're building a system of action in one of these workflows, we'd like to hear from you.*