AI
Agentic AI in procurement: where the most value is
Procurement has the processes and the numbers that AI agents need, yet it adopts them less than other functions
Agentic AI in procurement is moving more slowly than in the rest of the organization. In a Harvard Business Review Analytic Services survey of 385 respondents, conducted in March and April 2026, 35% reported AI agents in software development and 31% in IT operations, against 9% in procurement. The gap stands out because procurement combines structured processes, results measured in money, and manual friction, the conditions in which an agent pays off most.
In an article published in HBR in August 2026, Heiner Himmelreich, Ilan Oshri, Paolo Scala, and Anas Zaidani, who have followed the function for more than a decade, attribute the delay to the way organizations decide, control, and measure procurement. The cases they document already run in production. The HBR Analytic Services survey points the same way: only 18% of respondents say that, in most processes that use AI, it is integrated into the workflow, and 69% agree that legacy systems limit their ability to scale AI. An AI agent, here, is a system that receives a goal, queries data and other systems, carries out steps, and adjusts its course based on the result. In Brazilian public procurement, the word "agent" already has a legal meaning: Law 14,133/2021, the country's public procurement law, calls the civil servant appointed to run a tender and make its decisions the "contracting agent." When an AI agent joins a tender, the decision stays with the contracting agent, and governance has to spell out what belongs to each.
Why should procurement lead the adoption of AI agents?
Because procurement combines three features that rarely appear together elsewhere: results measured in the same currency the CFO uses, multistep processes repeated at high volume, and manual friction that traditional automation never removed. Any one of them would justify an agent; together they put procurement ahead of almost any other function.
The first is economic visibility. A better supplier choice, a more disciplined price, or a well-drafted clause shows up directly in the results, and the gain can be measured contract by contract.
The second is process structure. From supplier registration to quotes and contract monitoring, procurement runs on multistep workflows with defined documents, deadlines, and owners. In the Brazilian public sector, Law 14,133/2021 gives that workflow legal form, from planning the purchase to selecting the supplier and managing the contract.
The third is persistent friction. Checking qualification documents, classifying items, researching prices, and coordinating across departments still depend on people at critical points. This work resisted rules-based automation because it requires judgment across several steps, which AI agents can partly take on.
What the first cases show
Himmelreich and his coauthors describe a large European energy company that receives 30,000 to 40,000 supplier registration requests a year, each with up to 25 documents and more than 20 qualification steps, and that monitors 18,000 active suppliers with a team of about 25 people. In 2025, it put four agents into production. One guides suppliers through the form and flags the wrong choice of service category, which happens in about 30% of requests. Another orders the queue by what the business will actually need in the coming weeks. The third looks for risk signals, such as unpaid taxes and labor lawsuits, in internal and external sources. The fourth prepares a briefing on any supplier for the chief procurement officer before a meeting.
Six months after the first agent went live, the backlog had dropped 20%, and accuracy, which started at 60%, stood between 70% and 80%. The system has run under human oversight from the start.
The second case, an automaker, began in 2023 with text assistants for preparing quotes and found that the time saved did not turn into results within the old processes. In 2024, it redesigned its entire procurement cycle, with agents tied to economic results such as supplier selection and negotiation. About 70% of procurement work moved into a single chat interface, and a three-month A/B test measured the gain with the same metrics used to evaluate the function. In Brazil's federal government, the first steps are taking place in Compras.gov.br, the federal procurement platform. Since December 2025, Alice, a tool from the Office of the Comptroller General (CGU), has shown alerts inside the platform on electronic reverse-auction notices after they are published; the alert is guidance, and the manager reviews it and records a justification when disagreeing. On March 2, 2026, the Ministry of Management and Innovation in Public Services (MGI) and Serpro, the federal data processing company, launched MentorIA, an assistant that answers questions about purchase planning based on selected laws, decrees, and manuals. Both tools advise and leave the decision with the civil servant, the first rung of the autonomy discussed in the sections below.
Three organizational barriers
In the authors' interviews with procurement leaders and technology providers, three barriers recur, and all three are organizational design choices.
Accountability for the decision
Procurement runs on formal approval levels and makes decisions with contractual and financial consequences. A system that shapes supplier selection raises liability and audit questions that a marketing assistant never raises. In the Brazilian public sector, a tender is run by an appointed contracting agent, and the law requires segregation of duties, barring the same civil servant from holding the functions most exposed to risk. Many organizations respond by restricting agent autonomy so tightly that the agent adds nothing. In the HBR Analytic Services survey, 92% of respondents at organizations using or exploring agentic AI agree that agents need defined rules to operate safely, and 48% say their organization has defined them.
Ownership of the project
In many companies, procurement benefits from the AI initiative while IT or a central digital team designs and runs it, with little input from operations. The pilot works under controlled conditions and stalls when it meets the exceptions of the real process. In the same survey, 46% say their organization needs to clarify process ownership and accountability to improve AI results, and 47% point to cross-functional coordination.
Data
Supplier master data is inconsistent, spend classification is incomplete, and enterprise resource planning (ERP) systems keep information scattered across units and systems. In the HBR Analytic Services survey, among respondents using or exploring agents, 34% report that agents lack access to the right data. When an agent fails for lack of data, the usual conclusion is that the technology failed.
The three barriers lead to the same impasse, which the authors' interviewees call the "belief stage": the pilot launches, adoption is celebrated, and no one is accountable for the value. Only 16% of HBR Analytic Services respondents say their organization has realized a high degree of measurable value from AI.
Seven decisions to move past the pilot
The organizations already capturing value from agentic AI in procurement have made decisions about ownership, sequence, data, autonomy, capability, governance, and measurement. None is primarily technical, and each one strengthens the others.
Procurement owns the project
In the projects that scale, procurement defines the use case, owns the metrics, and answers for the operational outcome. IT enables the solution, and the trade-offs of each deployment stay with the people who will live with them.
What to do: name a procurement owner for each agent, with targets in the function's plan, and record IT as responsible for integration and security.
The biggest bottleneck first
Mature cases choose the process with the clearest economic or operational bottleneck: supplier registration, preparing quotes and tender notices, compliance monitoring. The most mature cases in the authors' research focus agents on direct materials and complex services, which weigh more in spend than technology purchases. A concentrated result speeds up learning and is hard to ignore when the scaling discussion comes.
What to do: rank procurement processes by spend volume and by time spent waiting in the queue, and start with the first on the list.
Data as infrastructure
Data quality is a condition for trusting an agent with a decision. Supplier master data, spend classification, and integration across systems are investments made before deployment, even if they delay the first agent.
What to do: before the first agent, measure the completeness of supplier master data and the share of unclassified spend, and set a target for each.
Autonomy in steps
Autonomy moves from advice to decisions with human approval and then to execution under oversight. In structured, low-risk tasks, such as document validation, the agent moves up quickly. In negotiation, it may stay in a support role indefinitely. In the Brazilian public sector, the final decision stays with the contracting agent, and the AI agent's step adjusts to that.
What to do: for each task, write down the agent's current step, the condition for moving up, and who approves the change.
Capability that carries from one agent to the next
Instead of isolated projects, the organizations that advance build a repeatable method for identifying, deploying, and scaling agents. The team behind the first agent carries its components, integrations, and governance rules into the second. Procurement professionals move toward supervising agents and deciding where AI falls short.
What to do: treat the first agent as the start of a portfolio, with the same team, a repository of reusable components, and a review routine for each new deployment.
Governance from the first agent
Approval levels, audit trails, escalation paths, and mechanisms to override the agent are defined before it goes into production. In the survey, 54% of respondents cite human review of AI outputs among the mechanisms already in place or in progress, while only 27% cite audit trails and 17% escalation paths for AI decisions. In mature cases, each agent has an investment case tied to a process step and a financial metric.
What to do: do not put an agent into production without an approved audit trail, escalation rule, and investment case.
Results measured in money
Mature cases measure agents with the function's own metrics: captured savings, negotiation gains, cost avoidance, and time to decision in critical situations. Chatbot usage and cycle-time reduction remain secondary metrics. In the survey, the most tracked measures are operational efficiency (57%) and productivity (56%), and only 18% track new revenue streams. What to do: include agents in the procurement results report, with savings and cost avoidance attributed to each one.
Conclusion
Procurement adopts less agentic AI than other functions because accountability for decisions, project ownership, and data stayed as they were before the technology arrived. The organizations that advance give procurement responsibility for the agents, prepare data and governance before production, and measure results in money. In the Brazilian public sector, Alice and MentorIA already show the first step, with the decision in the hands of the civil servant.
Agents work on contract and supplier data. Takter Converge CLM brings together in one place the contract registry, electronic service attestations, financial execution, and the traceable history of each contract, and it integrates with ERPs such as SAP and Oracle through REST APIs. To assess that foundation across your organization's contract portfolio, talk to Takter.
Sources
- Harvard Business Review Analytic Services, What Drives AI Value, April 2026 survey, sponsored by Appian.
- Heiner Himmelreich, Ilan Oshri, Paolo Scala, and Anas Zaidani, Why Agentic AI Could Transform Procurement, Harvard Business Review, August 2026.
- Law 14,133/2021, Brazil's public procurement law (in Portuguese).
- CGU, Alice integrated into Compras.gov.br, December 2025 (in Portuguese).
- MGI, MentorIA launch, February 2026 (in Portuguese).
