In January 2026, SaaStr, one of the largest B2B SaaS media and events businesses in the world, was running on three humans and twenty-plus AI agents.

Eight-figure revenue. Single-digit headcount. The last salespeople had quit, the CEO decided to stop replacing them with humans, and over ten months, a Chief AI Officer deployed agent after agent until the operation ran almost entirely on AI.

That is a media company story. And I want to be honest about that. SaaStr runs content, events, and sponsorship sales: workflows that are largely digital, largely structured, and highly repeatable. Not every organisation becomes a three-person operation. Procurement teams inside complex enterprises have regulatory constraints, relationship requirements, and judgement calls that do not compress that way.

But here is the part that does translate: the ratio is shifting. The amount of work a small team can cover, and cover well, is about to change significantly. And the teams that understand this early are going to operate at a throughput that looks completely different from the ones that do not. I covered what is actually happening right now in AI procurement in a recent video if you want the broader picture before reading on.

I run the AI Procurement Blueprint and a range of content projects alongside a full-time role. I do that with agents. Not because I have unlimited time, but because the right agents handle the volume so I can focus on the parts that actually require me. That is the model. Not three humans replacing a hundred. The same people covering far more ground.

Procurement is about to get a version of this, and the question is not whether you believe it is coming. It is whether you are preparing for it.

The instinct when people start using AI in procurement is to reach for a general assistant. Something that can do a bit of everything: answer questions, draft emails, summarise documents, help with research.

That is not what I mean by agents. And it is not what the organisations getting ahead are deploying.

What works is specialist agents, purpose-built for a specific job, with specific criteria, specific escalation rules, and a defined scope. A generalist AI assistant is useful in the same way a smart intern is useful. A specialist agent is more like a trained employee who does one job very well, handles volume without fatigue, and routes the exceptions to you.

If you want the fuller picture of what agents actually are and how they work, I wrote a plain-language explainer here, or watch the real-life procurement agent showcase on YouTube.

In procurement specifically, the specialist agent model is already taking shape. Contract review agents that extract clause-by-clause analysis against your playbook and route deviations to the right people. Supplier risk and due diligence agents that monitor your vendor base continuously rather than annually, flagging changes in financial health, compliance status, or regulatory exposure. Spend analysis agents that surface anomalies, consolidation opportunities, and off-contract spend without someone pulling a report. Renewal monitoring agents that track every contract end date in your portfolio and trigger the right workflow at the right time. Vendor onboarding agents that chase document submissions, validate insurance certificates and tax forms, and update your systems without manual data entry.

Gatekeeper, where I work, has built a range of these agents across contract management, vendor management, and third-party risk, including a DDQ approval agent, W-9 and W-8 validators, DORA compliance reviewers, and more. The pattern is consistent across all of them: the agent handles the volume, the human handles the judgment.

The question that is starting to emerge is what comes next. Category intelligence agents. Negotiation support agents. Savings identification and tracking agents. Regulatory change monitoring agents. These are not science fiction. They are a logical extension of what is already working at the operational layer. I covered what some of these look like in practice in AI Agents in Action.

The SaaStr story has a line that gets less attention than the headcount numbers: the biggest variable in their transformation was not which platform they used. It was how deeply they invested in training the agents.

Their observation was blunt. The companies that figure out agent training in 2025 will have massive advantages by 2026.

The difference between mediocre results and genuinely useful agents is not the tool. It is the investment in understanding how to build instructions, how to test them, and how to refine them when something goes wrong.

I have built over fifty agents at this point. I have had agents that rejected contracts they should have approved, flagged non-issues as critical problems, and misread form structures completely. Every one of those failures was an instructions problem, not a technology problem. Twenty minutes of refinement fixed most of them. I wrote up the exact process I use here.

But here is the thing I do not see discussed much: the risk on the other side of that learning curve.

Some of my agents are creating more work than they save. Not because they are broken, but because they are too productive in the wrong direction. An agent with no output constraints will generate near-infinite output. My attention and time are not infinite. So I have a review agent that produces a thorough, detailed analysis of every single thing it processes, when what I actually need is a clear signal: action required or not.

The bottleneck shifts. In the beginning, the challenge is getting agents to do useful things. Once you are past that, the challenge becomes managing the volume of what they produce and keeping your own focus on what actually requires a human decision. Constraining output is not failure; it is maturity. The best-configured agents I have do not tell me everything they found. They tell me what I need to do about it.

This is the part of agent management that only reveals itself once you have been doing it for a while. And it is the part that makes the difference between agents that genuinely free up capacity and agents that just move the administrative burden from one place to another.

The procurement professionals who will lead in the next five years will not necessarily be the ones with the deepest category expertise or the longest track record of negotiation. They will be the ones who understand how to deploy specialist agents, how to train them properly, how to manage their outputs, and how to build the human oversight structures that keep the judgement calls where they belong.

That is not a prediction about AI replacing procurement. It is a prediction about what procurement capability looks like when the operational layer is largely automated and the strategic layer is where all the human effort goes. I have written before about what that team structure looks like and the skills it requires, and I also covered the five ways AI will redefine procurement’s future in detail on YouTube alongside the specific skills you will need by 2030.

The SaaStr story is instructive not because every organisation should become a three-person team, but because it demonstrates that the operating model is possible, and that the people inside the organisations that figure it out are the ones shaping what comes next.

You can debate whether this is coming. Or you can start building the capability to manage it.

I know which one I would recommend.

A specialist AI agent is a purpose-built tool that performs one specific procurement task, such as contract clause review, due diligence assessment, or renewal monitoring, with defined criteria, escalation rules, and human oversight built into the workflow. Unlike a general AI assistant, it does not do a bit of everything; it does one job very well at scale.

Not necessarily. The more likely near-term outcome is the same people managing significantly more work. SaaStr’s three-person model reflects their specific media business context, not a universal template for all organisations. What translates is the ratio shift: more throughput per person, with agents handling the operational volume.

Contract review and clause analysis, vendor due diligence, supplier onboarding, renewal monitoring, spend analysis, and compliance monitoring are all live in production today. You can see specific agent examples in the LuminIQ agent library.

For someone already using AI tools regularly, a basic agent can be up and working in ten to twenty minutes. More complex agents with multiple decision criteria typically take a few hours to build and test properly. The time investment is in writing clear instructions and building test data, not technical configuration.

Poorly configured agents that create more work than they save. An agent with no output constraints will generate near-infinite output. The skill is not just building agents; it is constraining and managing them so your attention stays on what actually requires human judgement. Many teams discover this only after they have already deployed.

Start with What Actually Is an AI Agent? for the foundations, then AI Agents in Action for real workflow examples, and Building Agents That Actually Work for the practical build process. If you prefer video, the real-life procurement agent showcase is a good starting point on YouTube.

The AI Procurement Blueprint publishes every week. If someone on your team would find this useful, forward it on.

Thanks for reading The AI Procurement Blueprint! This post is public so feel free to share it.