If you’re a Procurement leader, the most important AI decision you make this year (or next) will not really be about AI. It will, and should, be about your people.
Source-to-Pay platforms used to be where your team did the work. They are now quietly becoming systems that do procurement alongside them. That makes the choice you are about to make bigger than picking a vendor. You are choosing what your function gets reinvented into, and whether that reinvention scales the people you already have or buries them under software they have to babysit.
The future of procurement is not about autonomous AI. It is about your people, and how they deliver outsized results with governed AI.
Imagine running your department the way most of the market is telling you to run your AI. Every time a new problem lands on your desk, you hire a new person to handle it. One for tail-spend analysis, another for supplier onboarding, a third for contract redlines. Each one is good at their narrow job, but each one has their own picture of how the company works, their own desk full of context, their own memory of what happened last week. They all report through a coordinator you also had to hire. By the end of the year you would not be running procurement. You would be running a staffing agency, and a badly organized one.
That is what most of the market is actually selling you, once you strip away the language. A framework where every new problem is a new agent, every agent needs its own orchestration, and every one of them needs its own governance. And guest what, you are not going to run out of new problems. If you follow that philosophy, you do not end up with intelligent procurement.
You end up with a fleet that demos beautifully and ages terribly, because you accumulate these agents the way you accumulate any other unmanaged headcount, one well-meaning success at a time, until you are spending more energy coordinating the team than the team saves you.
You Already Know How to Run This
You already know how to do this, and you have known for your entire career.
You hire someone capable. You do not hand them the keys on day one; you teach them how your organization actually works, the written policies and the unwritten rules and the way category strategies really get approved around here. You keep them close at first and give them more room as they earn your trust. And when they turn out to be good, you do everything you can to make sure their expertise does not leave the building when they do.
Capable hire, taught your way, trust earned over time, knowledge retained. That is not an AI strategy. That is just management. And it is the right lens for every decision you are about to make.
A 2026 Approach to Agentic AI, Not a 2025 One
Before getting to what we built, an important bit on the architecture, because it is the part most procurement leaders are not getting a clear answer on.
The frontier AI labs spent the last year converging on something. Instead of training narrow agents that each handle one task and hand work off to the next one, the systems that actually perform are a small number of broadly capable agents that get taught new behaviors through skill files, with tools and memory attached.
The shorthand for it is “skills-based architecture,” and you can see it across what the leading labs (Anthropic, Open AI, Google) have published over the last twelve months.
Here is what that actually means in procurement and finance, with an example boring enough to be honest. An invoice arrives and it needs to be validated against the contract, the purchase order, and the receipt. Four artifacts, sitting in three different parts of the system, governed by different rules.
In the swarm version of that problem, your AP agent picks it up and calls your contract agent, which calls your sourcing agent for the original RFx context, which then has to ask your supplier agent for the right legal entity.
Each one has its own context window, its own memory, its own slightly different picture of how the company works. By the time the answer comes back, it has been through a game of telephone, and nobody is sure which version of reality it is operating on.
In the skills-based version of the same problem, one agent loads the invoice-validation skill, pulls the contract and SLAs directly, checks the PO, reconciles the receipt, and flags the discrepancy. Same data, one context, one accountable action.
Nothing was lost in translation, because nothing was translated. The skill itself is assembled on the fly, from the tools and knowledge IVA already has access to, which means the next invoice with a slightly different shape gets the same intelligence applied dynamically.
We know this approach works, because we have already left the swarm behind. It was a 2025 way to think about agentic AI. Some vendors are still trying to sell it to you in 2026.
Procurement AI is Evolving Towards Agentic Systems
What We Actually Built
A quick orient before going further. IVA is our agentic core. IVA Studio is the AI control tower around it, the place where Skills are written and improved, where tools and external connections are managed, where the governance lives. We launched both together because neither one means very much on its own.
Start with the hire. IVA is not a fleet of narrow agents. It is one agentic core with access to the entire Source-to-Pay action surface, meaning every action a person can take in the platform, and then some. Think of it as the seasoned operator who already knows the building and where everything is kept, not a temp who needs the org chart explained.
There is nothing to assemble before it starts working, because the agent lives inside the platform and has access to all of it and knows how to use it, rather than a layer of AI bolted on top.
Then you teach it. The way you do that is a Skill, which is simply a plain-text file that captures how your organization does something. IVA already has a lot of Skills. How to run a sourcing event, how to qualify a supplier, how to create an amendment, invoice, PO, etc. What you teach it is how your business runs those things, the policies, the thresholds, the exceptions.
IVA Studio is where those Skills are written, reviewed, governed, and improved over time. If you have ever written an onboarding doc or an SOP, you already understand the mechanic. You are not building a new worker for every task. You are teaching the expert you already have, and with AI, that expertise sits alongside every user.
The dynamic assembly part is what makes this hold up over time. When IVA picks up a new problem, it does not wait for someone to build it as an agent. It pulls together the right Skills, tools, and knowledge in the moment, the same way a strong operator combines what they know on the fly to handle whatever shows up that day.
How the Work Gets Done
There is a working assumption out there that AI lives in a chat window. You ask it something, it gives you a smart answer, you go off and do the actual work. That is not what IVA is.
IVA runs a loop. It plans, it searches your data, it reasons about what to do, it acts inside the platform, and it monitors what happens next. It does that whether the task is drafting a sourcing strategy or validating a hundred invoices against contract terms buried inside PDF documents. A category manager preparing to re-source an expiring contract can have one conversation in which IVA pulls the current contract, benchmarks it against comparable agreements, identifies stronger suppliers, sets up the RFx, and launches the event. That is not five interactions across five tools. That is one agent, one conversation, the work done in minutes.
The conversation is only one of several ways IVA gets pulled in. It also runs inside a workflow as a step, triggered by a background event or a scheduled job, called from an external system through MCP, or invoked from a button somewhere in the platform that a user already trusts.
A supplier risk event hits the system, and IVA surfaces every affected contract, opens the right purchase orders, drafts a mitigation plan, and proposes qualified alternative suppliers, before anyone on your team has had time to ask. For work that genuinely benefits from parallel execution, IVA spins up temporary sub-agents to handle the orchestration and tears them down when the work is done. Nobody designs that, and nobody has to govern those sub-agents separately, because they inherit everything from the parent.
In plain English, IVA is not a feature on top of your procurement system. It is at the core of the system, available in any of the ways your team already operates: a conversation, a workflow step, a background process, a button.
Governed by Design, Dialed By You
The question every serious leader asks next is the right one. If this thing can act, who is holding the keys?
The honest answer is that autonomy is a dial, not a switch, and you are the one turning it. At the low end, your people direct IVA and review everything it produces. In the middle, IVA does the work, checking with your team at key decision points, and your team manages the exceptions. At the high end, it runs on its own and only surfaces what genuinely needs a human. You move along that dial at the pace you trust, exactly the way you would extend rope to a strong new hire who keeps proving themselves.
Autonomy is a Dial, Not a Switch
What makes that safe is structural, not a policy you hope everyone remembers. IVA inherits the permissions of whoever invokes it and cannot exceed them. Any sub-agent it spins up to handle a complex task inherits the same boundaries and cannot escalate past them.
When IVA is running autonomously inside a workflow or a background process, it still has an accountable human user behind it, with that person’s permissions and that person’s exception path. There is no anonymous AI activity happening on your data. Autonomous actions are absolutely possible, but they happen under control.
Every action is logged with a continuous audit trail. Governance lives at the platform level, inside the access model and business rules you already maintain, rather than in a separate rulebook you have to write and enforce. AI works only where it is allowed to work, and you can prove it line by line.
The same principle applies outside the platform, in both directions. When IVA reaches out to other systems your business runs, it does so through Model Context Protocol (MCP), using the accountable user’s identity and credentials, so the receiving system’s own governance still applies on the way in.
When IVA is called from the outside, the inbound request is treated like any other action inside the platform, with the same permission check, the same audit trail, and the same accountable user behind it.
On the model side, we have built IVA to be LLM-agnostic by design. Out of the box you can run it on models such as OpenAI, Anthropic, Google, Mistral, Meta, and xAI, picking the right model for the task at hand. You can also bring your own, including private models you may already be running inside your environment.
None of that changes how Skills work, how IVA understands your data, or how governance is enforced. That counts for more than it might appear at first, because the frontier model market is moving faster than any procurement org’s roadmap, and you should not have to re-architect your function every time a better model ships.
Most vendors will tell you the way to use AI is to identify a use case, build an agent for it, then build another one for the next use case, and another. With IVA, the day-one default is something very different. You already have one agent that handles thousands of use cases out of the box, set to a governed posture, with humans firmly in the loop.
Your job is not to keep building agents. Your job is to turn the dial as your team builds trust, decide where IVA should run in the background, decide which workflows are ready to be redefined, and decide what your function gets reinvented into. We have eliminated the burden of “build an agent” and replaced it with the environment for continuous, progressive reinvention of how procurement actually works.
The Knowledge Stops Walking out the door
Most procurement expertise lives in a handful of heads. Your best category manager knows things that are written down absolutely nowhere, and when that person retires or leaves for a competitor, most of it walks out the door with them.
When your best buyer’s way of working becomes a Skill, that knowledge stops being a personal asset and becomes an organizational one. The whole team runs it, not just the one person who happened to figure it out. New hires inherit it on their first day instead of slowly absorbing it over three years. IVA also keeps memory across interactions, so the second time it works on a category, a supplier, or a process, it is materially better than the first. The expertise compounds rather than decays. Your organization gets smarter on a curve, which is the exact opposite of what happens when knowledge stays trapped in people and people move on.
You Can’t Manage a Team on a Broken Foundation
None of this works on a broken foundation, and this is where I would push back hardest on the idea that all AI is created equal.
A brilliant new hire is useless if you give them no access to the files, no real system of record, and three contradictory versions of the supplier list. The same is true here, and it is the part most vendors would prefer you did not examine too closely.
Putting hundreds of agents on top of fragmented data and patchwork permissions does not give you intelligence. It gives you a fast, confident, well-spoken way to be wrong at scale.
This is the argument I have been making for years and it has not changed. AI quality is downstream of data quality, and data quality is downstream of architecture. A single data model gives you one source of truth instead of a slice of it. A semantic layer means IVA understands what your data means, not just where it happens to sit. Platform-enforced permissions mean it only ever sees what it is allowed to see. That is the foundation that turns your proprietary data into a moat instead of a liability. The model is a commodity. Your data, structured properly, is the asset.
There is a reason the headline description of IVA is that it knows your data and follows your rules. The Ivalua platform is not just where IVA reaches when it needs information. It is the entire toolset IVA uses to do anything at all. Knowledge and action live in the same model. That sounds like a technical detail. It is the whole game.
The Bet You’re Actually Making
The agents will come and go, and some of them will be genuinely impressive. The question that actually matters is not how many you can deploy. It is what kind of function you are reinventing while you do it.
Source-to-Pay platforms are no longer just the systems your people work inside. They are becoming the systems that do procurement work alongside them. The architectural choice you make today is also a choice about how that reinvention unfolds. The many-agents path reinvents your procurement people into agent managers, where running a fleet of narrow specialists becomes most of the job, and the judgment that used to define the function gets distributed across software you do not fully control.
The skills-based path does something different. It scales the impact your people already have, codifies the expertise that used to live only in their heads, and keeps the judgment, the relationships, and the strategy where they belong, with the people running the function.
The best procurement leaders of the next decade will not be the ones running the largest agent fleets. They will be the ones who reinvented the function while keeping it human at its core. That has always been the job. It still is.
IVA, powered by IVA Studio, is how we think you do it.
