What Procurement Agents Must Earn: Rujul Zaparde on Zip, Execution and Enterprise Control

1. Guest Background

Craig S. Smith's guest on this episode of Eye on AI is Rujul Zaparde, co-founder of Zip. Their subject is the changing role of procurement software: how agents might take responsibility for ongoing work, and what organizations must retain control over when software can communicate, evaluate documents and affect business records. Zaparde speaks as a company founder describing his own platform, which makes the distinction between product claims and demonstrated outcomes essential throughout the interview.

His route to enterprise software began with a markedly different business. He recalls leaving college around late 2012 to start FlightCar, which offered travelers free airport parking by renting their vehicles to other people. After roughly four years operating the company, it completed an asset sale to Mercedes-Benz in 2016. Zaparde describes an operationally demanding business and connects that experience to his subsequent determination to build a B2B SaaS company.

His more direct exposure to procurement came at Airbnb, where he met his co-founder Lou and encountered the process's stakeholder complexity. He subsequently served as a visiting partner at YC. After further exploration, the founders began the business in its present form around August 2020. That background explains the interview's combination of requester experience and platform-building ambition; it does not independently validate the performance claims that follow.

2. What the Episode Covers

Zaparde presents Zip's development as a sequence: orchestrate procurement work, understand it, then help perform it. He says approximately $1.2 trillion in cumulative spending has passed through Zip and been approved across its customers. That is a statement about platform throughput. It is neither the amount autonomously spent by AI nor a measure of savings, and treating it as either would exaggerate what the episode establishes.

The starting point for automation is his rough estimate that around 80 percent of procurement can be broken into relatively discrete transactional tasks. Examples include checking W-9 or W-8 documentation and reviewing a supplier's SOC 2 report against company policy. The percentage describes his assessment of the work's structure, rather than an achieved automation rate. Zip's earlier agents addressed individual tasks: produce an output, then obtain human approval or proceed under an automatic configuration.

The limitation was continuity. Procurement often requires asking for something, waiting for another party and deciding what to do with the response. Zaparde says Superagents organize capabilities around roles that can pursue these exchanges. His legal example connects to email through MCP, uses a customer's legal playbook, processes contract redlines and sends revisions to the supplier or its counsel. It can wait for the reply and continue. If important issues remain unresolved after several exchanges, it raises the need for a human lawyer to intervene.

Sourcing supplies a second example of the intended scope. The agent can identify existing suppliers in the ERP vendor master, search for additional candidates, create and send an RFP, wait for responses, evaluate cost and value, and return a supplier selection. The operative claim is that it can advance the process between decisions, rather than merely draft the opening document. These remain Zaparde's descriptions of capabilities; the interview does not walk through a complete execution record.

Customer-specific guidance matters within those workflows. Zaparde describes category playbooks and gives the example of a consumer-products business with separate purchasing expertise for sugar and cocoa. An agent can use those materials to reflect the category's rules. A generic sourcing prompt would not, by itself, establish that the relevant business judgment had been supplied.

The episode also distinguishes the breadth of the platform from agent autonomy. Zaparde says Zip covers request intake and cross-functional orchestration as well as procure-to-pay: requisitions, approvals, purchase orders, invoice matching and payment. Covering that sequence creates opportunities for automation, but does not establish that every step operates without a person. Indeed, the closer the discussion gets to authoritative records, the more prominent human review becomes.

3. Core Views: Reasoning, Examples, and Limits

The central argument is that procurement agents must be judged together with the controls that make their actions acceptable. Zaparde emphasizes that procurement information flows into ERP systems, the general ledger and ultimately financial statements. His demand for correctness explains why a persuasive answer is insufficient when it becomes an operational input. It should not be mistaken for evidence that Zip has achieved error-free execution: the standard a system needs to meet and its measured performance are different questions.

Auditability is one part of his answer. He says Zip records what an agent did, the context and sources it used, when an event occurred, and who configured and authorized a capability. Downloadable audit packages make that history available for review. The reasoning is practical: a reviewer needs to reconstruct both the basis of a decision and the origin of its authority. Keeping the conclusion alone would leave a gap between an apparently valid review and permission to act on it. However, a complete record makes examination possible; it does not itself make the underlying decision correct.

His confidentiality example sharpens the access-control problem. Information about preparations for an IPO might sit in a consulting contract, invoice or statement of work. A user who cannot access that material must not learn its contents through an answer to a different question. Zaparde says Zip has 350 permission types that apply to people and agents, with an agent acting for a user constrained by that user's access. He also acknowledges that implementing and validating this behavior is difficult. The meaningful issue is therefore not the permission count alone, but whether restrictions survive the transformation of source material into an answer.

Model selection follows a similarly task-specific logic. According to Zaparde, Zip compares models on an effectively daily basis, including tests with and without its business context. Evaluation can distinguish invoice headers from line-item coding and examine contracts at clause level. He argues that some invoice tasks benefit more from the right context than from the newest model. His example of an invoice around 1,200 pages long, potentially in another language, also shows why a single task label can conceal very different workloads. Routing by complexity has an intelligible cost-and-performance rationale. But when asked about long-context computation, he acknowledges that he is not the internal technical expert and does not explain the architecture. The interview cannot establish a particular retrieval method or long-document accuracy level.

The commercial argument extends beyond making existing work faster. Zaparde says customers generally retain people for especially consequential contracts, using a $50 million AWS agreement as an illustration. Yet purchases of $250,000 or even $1 million may receive limited competitive sourcing in some organizations because the effort is too high. Lowering the cost of running the process could bring previously unattended spending into competition. That is a distinct source of potential value: increased coverage, rather than just reduced handling time.

UCI Health is offered as an example, with important qualifications. Zaparde attributes more than $20 million in cumulative cost avoidance to its use of a Zip negotiation-agent use case. He confirms that the amount spans multiple contracts and says he would need to check whether the period was around six months or a year. The episode does not disclose the calculation baseline or independent verification. Cost avoidance should consequently retain that label, rather than becoming a definite cash-saving result or a transferable forecast for another buyer.

The deployment survey warrants the same discipline. After Smith mentions a report surveying more than a thousand spend leaders, Zaparde recalls ROI of about 39 percent for deeper deployments versus about 6 percent for pilots. The interview does not define ROI, explain the groupings or establish causation. It supports asking how much work a deployment actually changes; it does not prove that deeper deployment necessarily causes that return. Likewise, Zaparde's view that the person granting authority should bear responsibility for an agent's mistake is a personal position in what he calls an evolving framework, not a disclosed contractual allocation or legal conclusion. His broader expectation is that expanding autonomy makes governance more important. The episode supplies a coherent rationale for that expectation, while leaving its effectiveness to be demonstrated in specific deployments.

4. Learning and Application

A practical deployment can begin with roles, rather than a blanket decision to enable agents. Zaparde describes recommended defaults for requesters, legal reviewers and other personas, with individual permissions available for adjustment. Smaller organizations may accept those defaults; larger ones may involve auditors in reviewing every permission. The useful application is to treat the default as a reviewable starting configuration. It reduces setup effort, but the organization still needs to establish who may view information, edit it and authorize action. A convenient preset is not evidence that a particular company's confidentiality requirements have been met.

The next decision is which operations may proceed continuously and which must stop for approval. Zaparde describes an agent initially asking permission to send an email, after which the user may choose automatic handling. Customers can also require human sign-off under specified risk conditions. He strongly recommends review when information is pushed into a source system such as an ERP, and says customers generally want it. A team can apply that distinction by testing lower-risk exchanges before widening authority and separately accepting any workflow that changes authoritative records. The tradeoff is explicit: review preserves a decision point but consumes some of the speed automation promises. The interview offers a configurable approach, not a universal list of safe autonomous actions.

Use-case selection should then start from work already occurring. Zaparde illustrates the method with an analysis of roughly 85,000 approvals over a year and a possible set of 25 configured use cases; he expressly says that the latter number is made up. A procurement organization can borrow the method without treating those figures as a completed customer project. Examine repeated approvals and follow-up work, identify where a changed process could matter, and specify which outcome would count. Cycle-time reduction, broader competitive bidding and reductions in some BPO work are different benefits with different conditions. Reporting them separately makes the case more interpretable than collapsing everything into an unexplained savings figure.

Implementation also requires an owner for configuration and testing. Zaparde says Zip's forward-deployed engineers adjust prompts, configure agents and test them without changing code, using capabilities available to customers themselves. That suggests a concrete implementation responsibility: translate business rules into settings, review exception behavior and establish whether the resulting workflow can operate inside the organization. External assistance may accelerate this work, but someone internally still needs to understand what was configured. Otherwise a change in policy or workflow creates another dependency instead of a maintainable operating process.

Finally, workforce planning should distinguish task relief from forecasts about jobs. Zaparde expects efficiency gains to support reinvestment in research and development and potentially other roles. He also says persuasion, turning an idea into an initiative and running a project remain difficult to automate. A grounded application is to retain clear ownership for those activities and check whether removing transactional work actually gives those people usable time. His optimism about reinvestment and employment is a future judgment; it cannot serve as a guaranteed outcome in an individual organization's staffing or budget plan.

Source

• Original episode: How Zip's AI Superagents Automate Procurement End-to-End | Rujul Zaparde