Shorter source-to-award cycle
Less time to move from sourcing to a supplier decision.
Journey & Experience / ZycusMerlin Agentic
Zycus · 0→1 enterprise AI · Product leadership
Merlin Agentic
Sourcing Platform
Making more spend worth sourcing. Moving the coordination from the buyer into the platform.
My role: Product Manager, primary PM ownerUsers: enterprise buyers, sourcing teams and suppliersScope: requirement to award, then purchase-request sync
A faster workflow. A better commercial outcome.
I led the 0→1 product strategy and roadmap for a configurable multi-agent sourcing journey. The goal was to reduce manual coordination and make lower-value, high-volume purchases economically worth sourcing and negotiating.
Supplier participation
More participation in sourcing events.
Tail-spend savings
Commercial value supported by evidence-grounded price benchmarking.
Negotiation success
Supplier negotiations that met the success criteria.
Award recommendation acceptance
Buyers accepted the recommended award decisions.
The workflow was digital.
The buyer was still the orchestrator.
Buyers already had sourcing software. They still had to move between products, select suppliers, compare bids, negotiate outside the system and update the original purchase request. The coordination and decision effort remained with them.
Before: buyer-coordinated work
- Navigate multiple procurement modules.
- Manually sequence sourcing steps.
- Spend time comparing and negotiating every event.
- Carry the final decision back to the source request.
After: one coordinated journey
- Start with a requirement or sourcing opportunity.
- Specialized agents progress through the workflow.
- Buyers set priorities and retain decision control.
- Approved supplier and pricing return to the purchase request.
“Tail spend” means lower-value, high-volume purchases that often receive less procurement attention. Manually negotiating each one can consume more buyer time than the saving justifies. Reducing that effort creates room to optimize spend that was previously underserved.
I owned the product direction and the trade-offs.
Strategy and roadmap
Defined the customer problem, end-to-end scope, priorities and success metrics. Reframed the product around the sourcing outcome rather than individual modules.
AI and autonomy decisions
Defined agent responsibilities, supplier recommendation requirements, benchmarking, negotiation and award experiences, including where buyers needed control.
Cross-functional delivery
Worked with engineering, AI/ML, UX, QA and customer-facing teams to translate product choices into delivery, demos, feedback and iteration.
Team contribution: Engineering and AI teams owned production services, models, retrieval, APIs, orchestration implementation and integrations. My ownership was the problem, product behavior, priorities, acceptable trade-offs and measures of success.
Three decisions shaped the platform.
Make underserved spend part of the opportunity.
I framed automation as a way to expand sourcing coverage, with a proactive opportunity capability alongside buyer-initiated requirements.
- Alternative
- Focus only on speeding up events buyers already chose to run.
- Why I chose this
- The larger business opportunity was to reduce the effort needed to pursue smaller savings opportunities.
- Trade-off
- Surfacing opportunities adds prioritization complexity. Buyer attention must go toward opportunities with meaningful commercial value.
Make source-to-award the product boundary.
I chose one coordinated sourcing journey using specialized agents and existing Zycus capabilities underneath.
- Alternative
- Add separate AI features to supplier management, sourcing and award screens.
- Why I chose this
- Smarter screens would still leave buyers switching products and deciding what happened next.
- Trade-off
- We accepted more internal orchestration and integration complexity to reduce the customer’s coordination burden.
Earn autonomy stage by stage.
I phased autonomy according to consequence, reversibility and buyer confidence.
- Alternative
- Make every stage autonomous from the first release.
- Why I chose this
- An editable supplier shortlist and a consequential commercial decision require different levels of control.
- Trade-off
- Human review limits automation, but lets buyers delegate within boundaries they understand and trust.
Seven agents. One business outcome.
Each agent had a defined responsibility, the intelligence suited to its decision and an appropriate level of buyer control.
Explore the sourcing journey
Workflow illustrationFrom a purchase need to a completed request.
Inputs
Buyer request
Intake / purchase request
1
Requirement Gathering
Turn a purchase need into a sourcing-ready requirement.
Outputs
Structured sourcing need
Originating request context
01
Define the need
Requirement Gathering
Turn a purchase need into a sourcing-ready requirement.
Capabilities
Receive a buyer-created requirement, an intake request or an existing purchase request. Capture what is needed and route it into sourcing.
AI & logic
Structured requirement capture and routing establish the context the downstream agents need. The opportunity capability could also surface spend worth sourcing.
My product decision
I designed multiple entry paths so buyers could start from the work already in their procurement systems, rather than recreate it in a separate tool.
Measures of quality
requirement completeness, conversion into sourcing events and eligible spend brought into the workflow.
Buyer control
Buyer or requester defines the purchase need.
02
Prepare the event
Event Creation
Create a sourcing event without starting from a blank page.
Capabilities
Create the RFx, prepare supplier questions and configure event information for the requirement. RFx is a request for supplier information, quotations or proposals.
AI & logic
Generative AI prepares relevant questions on price, delivery, qualifications, certifications and commercial terms. Event rules provide the structure.
My product decision
I prioritized early autonomy here because the output was reviewable and editable, with no final commercial commitment.
Measures of quality
question relevance, completeness, buyer edit rate and event setup time.
Buyer control
Buyer reviews or edits the event according to configuration.
03
Build competition
Supplier Recommendation
Find suppliers that fit the actual item and business context.
Capabilities
Retrieve supplier candidates, apply business filters, rank them and explain why they are relevant. Buyers can add or remove suppliers.
AI & logic
Embeddings and semantic search retrieve similar items and supplier transaction evidence. Location and category signals support filtering and re-ranking.
My product decision
I prioritized item relevance and location over broad category matching. A supplier in the right category may still be wrong for the specific requirement.
Business outcome+80% supplier participation
Measures of quality
shortlist relevance, buyer additions or removals and supplier response rate.
Buyer control
Buyer can accept, add or remove suppliers.
04
Compare the options
Bid Collection & Evaluation
Make submitted bids easier to compare before negotiation.
Capabilities
Collect supplier responses, close submissions at the configured deadline, normalize valid bids and summarize differences.
AI & logic
AI analyzes and compares responses. Deterministic rules enforce the deadline and determine which submissions can move forward.
My product decision
I treated evaluation as analytical work, suitable for earlier autonomy, while keeping final commercial commitment outside this stage.
Measures of quality
extraction accuracy, comparison correctness, missing-field detection and deadline adherence.
Buyer control
Buyer can inspect supplier responses and comparisons.
05
Improve the deal
Negotiation / ANA
Negotiate directly with suppliers within buyer-defined boundaries.
Capabilities
Run multi-round supplier conversations. Price negotiation evolved into multi-parameter negotiation covering delivery, payment terms and other commercial conditions, with later human takeover.
AI & logic
Conversational AI uses the current bid, evidence-grounded benchmark, previous rounds and negotiation state to choose counteroffers and next actions. RAG price benchmarking retrieves historical spend, comparable purchases and available market context.
My product decision
I grounded price targets in evidence. As negotiation expanded beyond price, buyers defined acceptable ranges and priorities rather than leaving the model to invent their business preferences.
Business outcome68% negotiation success · ~7% tail-spend savings supported by benchmarking
Measures of quality
target achievement, rounds to convergence, boundary adherence and human takeover.
Buyer control
Buyer sets boundaries and priorities, and can take over the conversation.
06
Choose best value
Award Recommendation
Recommend the supplier or allocation that fits the buyer’s objective.
Capabilities
Recommend suppliers per item, allocations across suppliers or one supplier for consolidation. Recalculate recommendations against configurable priorities.
AI & logic
Configurable decision logic combines buyer criteria and weights with AI synthesis of bids and negotiated terms. The recommendation explains the choice.
My product decision
I made “best value” configurable across price, quality and delivery, rather than equating the best supplier with the lowest price. Final decisions remained reviewable.
Business outcome95% buyer acceptance of award recommendations
Measures of quality
acceptance, overrides, adherence to constraints and alignment with buyer priorities.
Buyer control
Buyer reviews and accepts the recommendation before the final decision.
07
Close the loop
PR Sync
Carry the sourcing decision back into the originating purchase request.
Capabilities
Update the originating request with the selected supplier and finalized pricing once sourcing is complete.
AI & logic
A deterministic integration action writes the final outcome to the source record. This stage needs predictable execution rather than generative reasoning.
My product decision
I included PR Sync as the seventh stage because the workflow should complete operational work, not stop at a recommendation that a buyer must copy elsewhere.
Measures of quality
sync success, correct supplier and pricing fields, and reconciliation of failed updates.
Buyer control
The final sourcing outcome is written back to the originating request.
More intelligence did not mean less buyer control.
Start with reversible outputs.
Event creation, supplier recommendations and bid evaluation could be reviewed or changed before a commercial commitment.
Expand only with explicit priorities.
Price negotiation evolved into multi-parameter negotiation across price, delivery and payment terms. Buyers defined acceptable ranges and priorities.
Keep human judgment available.
Later human takeover let buyers handle exceptions in negotiations. Award recommendations remained reviewable before the final decision.
“Best supplier” depended on the buyer’s objective.
Award recommendations supported per-item choices, allocations across suppliers or consolidation with one supplier. Configurable criteria let buyers express the importance of price, quality and delivery. Lowest price was not automatically the best outcome.
Measure the value of completed work.
My recommended North Star
Verified savings realized through agentic sourcing
Validated savings from completed sourcing events, measured against an agreed baseline. To call savings “realized,” procurement and finance should confirm that the saving was achieved beyond the award recommendation.
Coverage
Eligible spend sourced and opportunities converted into events.
Competition
Relevant suppliers and meaningful bid participation.
Commercial quality
Benchmark quality, negotiation outcomes and accepted awards.
Efficiency
Cycle time and buyer effort per completed event.
Guardrails: Savings must sit alongside policy compliance, supplier suitability, commercial boundary adherence and correct purchase-request updates.
How I would evaluate and diagnose the workflow
My evaluation approach connects the quality of each agent’s decisions to buyer behavior and business outcomes.
Observability: Trace each stage’s inputs, outputs, evidence, tool results, time, failures and human interventions to understand what happened.
Evals: Judge whether each stage did its job well. Supplier ranking needs relevance checks; negotiation needs commercial success and boundary checks; awards need alignment with buyer priorities; PR Sync needs field-level correctness.
Link those checks to buyer behavior and business outcomes, so a model score can be investigated in the context of a completed sourcing event.
Agentic AI was a workflow redesign problem.
The most important product work was deciding what the system could own, where evidence was needed, and where business judgment belonged. Specialized agents mattered because their responsibilities matched the work, not because seven was the goal.
My takeaway: adoption depends on understandable decisions, clear boundaries and the ability to intervene. The final operational update matters as much as the recommendation.
Return to my Zycus experience