Find the right counterpart, not just a product.
FromFind a product→ToFind the right counterpart
For custom work, services and business purchases, finding the right counterpart still takes many messages. Matching and agents can take the coordination off people’s hands, while people keep every decision that matters.
A dining table for eight
A buyer tells her agent: “I need a dining table for eight. Solid wood, something distinctive, under ₹80,000, delivered within a month.” illustrative
Her agent sends a structured requirement into the network. A furniture maker’s agent checks capacity, materials, lead time and service area, and replies: “Available. ₹82,000. Delivery in 24 days.” Another maker’s agent replies: “We can’t meet the 30-day deadline.”
Within the ceiling she set, her agent counters at ₹78,000, and the first maker agrees. A few minutes later it says: “Seller accepted ₹78,000 with delivery on the 14th. Shall I confirm?” She taps yes.
Marketplaces made buying standard products remarkably easy: search, select, buy. For custom work, services and business purchases, finding the right counterpart still takes many messages.
- Human
- personal agent
- open network
- participant agent
- human
How matching works
Two models are compared: what the buyer needs, and what the participant can provide. The buyer’s needs carry four kinds of signal.
| Signal | Example |
|---|---|
| Hard constraint | It must fit through this doorway |
| Strong preference | Prefer reclaimed wood |
| Soft preference | Ideally a local maker |
| Exploratory | Something unusual |
Trade-offs are shown, never hidden. “This maker fits everything except budget. They’re ₹10,000 above it. Want to see them?” The buyer decides.
Whole needs, not keywords. “I’m planning a wedding” can become venue, photography, catering and décor, found individually or through a coordinator. Matching can include working style (“someone who’ll collaborate closely with me”), and it works between businesses too: a supplier who can produce 50,000 precision components in six weeks is matched on capability, capacity, certification, lead time, performance and terms.
Explainable, gaps included, and honest about uncertainty. Strong match because: capability verified, available in time, meets your quality bar, supports milestone payments. Potential gap: price is above target. And where the evidence is thin: “I don’t have enough evidence to know whether this provider can handle a project this size.”
Agents handle the routine. People hold the commitment.
Agents handle search, qualification, availability, routine negotiation and coordination.
People handle preferences, important trade-offs, relationships, exceptions, judgement and every final commitment.
Authority is set by humans. “Negotiate up to 5% on price.” “Don’t exceed ₹80,000 without asking me.”
Agents can act on instructions. They should never silently redefine them.
Go deeper
What already exists
The pieces for agents that act under a person’s authority are appearing. Google’s AP2 protocol uses signed “mandates” that record what a user has authorised an agent to do[1]. The Razorpay, NPCI and OpenAI pilot, announced in October 2025, lets an agent place an order after a single confirmation, with real-time tracking and instant revocation for the user[2].
Scope. These are payment announcements. They don’t show agents matching and negotiating with each other over custom work, which is what this shift would test. Razorpay describes fully autonomous agent payments as a future possibility, not part of the current pilot.
Getting it right
Every agent is identified: who it represents, what authority it has and what it can commit to. The Trust teardown covers the protections.
No spam. Permissions, rate limits and relevance thresholds, and buyers can refuse unsolicited offers.
Fair to both sides. Bounded negotiation, clear terms and human escalation: coordination, not pressure.
A record of every interaction, so any dispute starts from what actually happened.
Who it helps. Buyers skip the back-and-forth and arrive at the conversations that matter. Sellers stop answering enquiries they can’t fulfil. Businesses compress days of procurement into structured, comparable options.
What could go wrong
- Unequal agents.A well-resourced agent could press a small seller into a poor deal, so negotiation is bounded and terms are clear.
- Offer spam.Seller agents could flood buyers with offers, so buyers can refuse unsolicited ones.
- Authority creep.An agent that quietly widens its own mandate undermines trust, so any change comes back to the person.
- Thin evidence.Matching a large order to a small maker on too little evidence can fail both sides, so the agent says what it doesn’t know.
How we’d test it
Custom furniture requests matched between buyers and 30 makers in one region.
- The test
- Matching first
- What we would measure
- Start with matching only, then add agent negotiation within strict limits. Measure time to a good match, how often matches become orders, and satisfaction on both sides.
- What would end the idea
- Matches rarely turn into orders, or either side feels pressured by the agents.
Sources
- [1]Google Cloud, Announcing Agent Payments Protocol (AP2) (16 September 2025)Checked; a company announcement
- [2]Razorpay, Razorpay, NPCI and OpenAI come together to launch agentic payments (9 October 2025)Checked; a company announcement of a pilot
Spotted a source we missed or a rule we misread? Tell us what we got wrong.
Matching and agents work only on an interface people can trust. Who controls that interface is the next question.