AI agents in procurement, explained: what they actually do in source-to-pay today

What counts as an agent

The word gets stamped on everything, so start with a definition you can test. Generative AI automates routine processes, analyzes large volumes of information and supports decisions; agentic AI takes on tasks and interacts with other business systems, as the CPOstrategy article on procurement talent frames the difference. In practice an agent does three things a macro or a rules engine cannot: it reads live state from a system (open contracts, shipment status, an email queue), decides which step comes next, and performs an action there — under a person’s permissions, with the option to hand off to a human.

Two plumbing standards make this possible in procurement stacks. Model Context Protocol (MCP) is an open standard for connecting AI applications to external systems — data sources, tools and workflows — supported across major assistants and developer tools. Agent2Agent (A2A) covers agents calling other agents. SAP, for one, says its Joule Work now uses both protocols to plan and execute multistep workflows across SAP and non-SAP systems.

Below is what agents are doing in source-to-pay as of late October 2026, with each example drawn from a launch announcement or report we verified this month.

Sourcing events and post-award price reviews

Arkestro took two agent-facing capabilities to general availability on Oct. 14 at its Optimal ’26 conference in Houston:

• Pricing Updates schedules recurring price reviews on live awards. Suppliers are prompted to confirm prices or submit changes; the system evaluates each proposed change in the context of the original negotiation and can suggest counteroffers or tailored feedback before new pricing takes effect, accounting for local currencies, freight, tooling and commodity-index pricing. It was previously limited to select customers.
• MCP connectivity lets enterprise AI assistants pull Arkestro’s intelligence into tools staff already use. The company says agents can monitor sourcing activity, summarize what needs attention, draft supplier communications and prepare new competitive bids, with the enterprise controlling what information an agent can access and what actions it can take.

Crown Holdings’ director of indirect sourcing for the Americas, James Cuff, said in the release that Arkestro has cut the time the company spends reviewing and coordinating supplier price updates by roughly 50% — a vendor-reported figure, but with a named customer attached. This is the least glamorous and possibly most valuable agent category: negotiated savings decay after award, and most teams have no mechanism to re-check prices until a supplier raises them. Our Arkestro launch coverage breaks down the mechanics.

Requisition intake and routing

The front door of procurement is where orchestration agents work. On Oct. 5, ORO Labs announced that SAP will sell its platform as a solution extension called SAP Ariba Procurement Orchestration by ORO Labs, with general availability planned for November 2026. Employees start and follow procurement work in one experience; pre-built connectors and APIs reach the SAP and third-party systems behind it, SAP Integration Suite supports the plumbing, and low-code workflows let teams change processes as policies shift. ORO says agentic steps sit inside governed processes that keep approvals and human oversight in place, and that the extension is designed to work alongside Joule Work deployments. Named users include Coca-Cola, Pfizer, Siemens Energy and Novartis.

The pattern to notice: the agent does not replace the suite. It routes a request across an ERP, a sourcing tool, a contract system and the like, keeping each transaction inside existing approval structures. We covered the deal in SAP’s ORO Labs resale story.

Spend classification and supplier data

Classification was the first procurement task AI got right, because the failure mode — a miscoded invoice line — is cheap and the volumes are enormous. SAP Ariba Spend Analysis and Insights, generally available since Sept. 16, continuously ingests spend from SAP Ariba, Fieldglass, Concur and SAP Cloud ERP plus customer-managed and third-party sources, classifies it to UNSPSC v25 or a custom taxonomy at a claimed 95% accuracy, and enriches supplier records with Dun & Bradstreet corporate hierarchy weekly, moving to soon-daily. SAP says SAP-managed sources are available within minutes, and that AI and machine-learning models are trained to learn each customer’s business across multiple languages.

The agents connect on the other side: SAP built the solution on its Business AI Platform so clean, classified spend data feeds Joule agents and recommendations in next-generation Ariba workflows. SAP’s own launch post states the constraint plainly — one of the greatest barriers to harnessing AI is the lack of high-quality data — which is why classification agents are getting budget before negotiation agents do. See our spend data coverage for the master-data work this depends on.

Freight quoting — the email-queue agents

Freight is where agent adoption is furthest along, because quoting is high-volume, deadline-driven and still runs on email. Nexcade’s CEO Dan Bailey told FreightWaves that roughly 90% of quote requests still arrive by email. Its agents read customer inboxes, classify the freight type, extract shipment details and run lookups against contracts, spot-market rates and overseas agent portals via API or browser, staging 10 to 15 quotes per operator each morning instead of four or five email chains per quote. Bailey says teams on the platform have doubled files handled per person, one customer now answers overseas agent requests arriving at 3 a.m. before competitors are awake, and about 75% of volume is air and ocean with 25% road. His caveat applies to every vendor in this category: “Context is the biggest challenge in our space” — turning years of lane knowledge into something an agent can use is the hard part. Our logistics AI ROI analysis covers the measurement side.

Freight booking — agents that transact

On Sept. 29, Flexport launched a Model Context Protocol server that moves freight from quoting to committing spend. A shipper’s own assistant — Claude, ChatGPT or Codex, Microsoft Copilot, Meta’s Muse or an internal tool — can call six operations against the shipper’s Flexport account: track a shipment with milestones, customs status and holds; browse active shipments flagged for exceptions or demurrage and detention risk; search active contract and available spot rates on ocean FCL/LCL and air lanes, requesting pricing from Flexport’s team where no rate exists; browse quotes with line-item charges; create a booking against a rate, with price, routing and cargo details validated first; and look up account details such as HS codes, port identifiers and verified addresses. FreightWaves’ report of the launch notes an agent can submit cargo details, book against negotiated rates and return a booking ID and transit time without logging into the platform.

The numbers behind it, from Flexport and FreightWaves: more than 13,000 companies use the platform, and requests landing through MCP are processed by Flexport’s own internal AI agents, which the company says handle 21 million tasks a year, including a quoting agent. Every agent action is scoped to one person’s login — “always within the permissions you already have,” in Flexport’s words — an administrator enables the connection org-wide and can revoke any user or switch the whole organization off, ending active sessions immediately, and suppliers booking on a shipper’s behalf see cargo details only, never rates or pricing. Full detail in our Flexport MCP story.

Suite-embedded agents and the governance layer

At SAP TechEd in Berlin on Oct. 27, SAP described out-of-the-box assistants and agents across five horizontal domains, business users and developers building and extending agents in Joule Studio with no-code and pro-code tools drawing on more than 70 AI models, and new agentic thinking capabilities that plan and execute multistep workflows across SAP landscapes via A2A and MCP. Third-party agents can invoke Joule agents through a governed agent gateway with enforced access control.

The governance half arrived with it: Joule Studio gained observability and administrative controls over agent behavior, model usage and spend; a new SAP Enterprise AI Gateway will centralize authentication, policy enforcement, routing, monitoring, usage tracking and auditing for AI traffic including MCPs, APIs, A2As and LLMs (planned for Q1 2027); and a new AI Governance Assistant targets EU AI Act compliance obligations. SAP also says customer data remains the customer’s: it does not share customer data with anyone to train or improve large language models.

That last cluster matters because the gap is real. Pathlock’s survey of 286 IT, compliance and security decision-makers, covered by SAPinsider, found 52% of organizations cannot fully verify every action their agents execute, even though 38% allow agents to create or modify business records and 28% allow them to approve transactions. Agents are arriving in procurement faster than the ability to audit them.

What agents do not do yet

Set expectations against what is shipping:

• Backward-looking analysis. Flexport’s own FAQ says historical questions like average transit time by lane are not available through its agent tools today, arriving later through a planned Insights Builder.
• Unsupervised commitments. Booking flows validate details and stage no-rate requests for approval; orchestration vendors keep agentic steps inside governed approval processes by design.
• Judgment with nobody accountable. Permissions inherit from a named person’s login, and misclassification, wrong-price or wrong-book consequences still land on the shipper or buyer. Treat agent output like a direct report’s: review the ones that matter, log everything.

What agents do in source-to-pay today: a summary

Process	What the agent does	Human role left standing	Example verified this month
Requisition intake	Routes requests across ERP, sourcing, contract and risk systems	Sets routing rules; approves exceptions	SAP Ariba Procurement Orchestration by ORO Labs
Sourcing events	Monitors activity, summarizes, drafts supplier communications, prepares bids	Runs negotiation strategy; signs awards	Arkestro MCP connectivity
Post-award price review	Schedules reviews, evaluates changes against the negotiation, suggests counteroffers	Approves price moves	Arkestro Pricing Updates
Spend classification	Ingests, codes to taxonomy, enriches supplier hierarchy	Owns the category map; samples accuracy	SAP Ariba Spend Analysis and Insights
Freight quoting	Reads email requests, pulls contract and spot rates, stages quotes	Picks the pricing decision	Nexcade agents
Freight booking	Books against negotiated rates, returns booking ID	Approves writes; owns exceptions	Flexport MCP server

This mapping is ours; the capability claims come from the cited vendor and press sources.

What to watch

• ORO Labs general availability, planned for November 2026, will be the first wide test of SAP-branded intake orchestration agents.
• SAP Enterprise AI Gateway availability in Q1 2027 will show whether cross-vendor agent auditing becomes a product category or stays per-suite.
• Independent accuracy data. Classification claims like SAP’s 95% and customer results like Crown Holdings’ ~50% need third-party verification as these tools reach production scale.
• Copycat agent interfaces at other forwarders, which would make MCP support standard freight RFP language within a year, as we argued in our Flexport coverage.