Build, Partner, Buy or Defer: McKinsey's agentic AI framework, read from the procurement seat

TL;DR : McKinsey's agentic framework is a make-or-buy applied to AI. For CPOs and digital/process teams: real AI infrastructure is about to get complex and expensive, procurement has to be in the room to keep the bill sane, and "wait" (Defer) is a legitimate, often optimal move. This is not an argument against buying AI. The upside is real when you apply it well, so experiment and own the topic.

"Don't worry, we'll just vibecode it, it'll be easy." If you're a CPO, or on a digital or process team, you've heard some version of that line about AI agents. It's rarely that easy, and it's rarely that cheap.

For 18 months, every IT and digital committee has received the same request: "we want to deploy AI agents on [finance, HR, support, sourcing, legal]". The business shows up with a PoC, the incumbent vendor pushes its agentic module, a startup pitches a dedicated solution, and IT wavers between insourcing everything and not. McKinsey published an architecture for enterprise agentic platforms, paired with a Build / Partner / Buy / Defer decision tree. It doesn't decide for you, but it asks the right questions in the right order, which is exactly what's missing to structure a make-or-buy when the object is an agentic platform rather than classic SaaS.

1. The agentic architecture (the short version)

Read this if you're interested. It's short on purpose, and genuinely useful, but if you only want the decision, skip to Part 2. For the full picture, the QuantumBlack piece is excellent: Creating a future-proof enterprise agentic platform architecture.

McKinsey breaks an enterprise agentic platform into four layers, from the most visible to the deepest, where a lot of things will need to be bought. Procurement should be involved, because every one of these layers represents real money and real risk. Start with the infrastructure, because that's the most obvious place the money goes: compute, inference, model routing, gateways. Then notice that governance isn't a box at the end, it sits at the onset, cuts across every layer, and is becoming the capability to build first.

• Marketplace & Workflows: the shop window: pre-built agents, packaged workflows, invocable tools. The layer vendors fight to own, and where lock-in starts.
• Agentic systems: the plumbing, in four archetypes (productivity agents, highly custom, purpose-fit, workflow automation). Each means a different vendor type and pricing model.
• Runtimes & Interfaces: runtimes plus API / MCP / LLM gateways. Invisible, but the single biggest driver of TCO. Treat it like cloud infrastructure, not SaaS.
• Shared services & Cloud / Infrastructure: registries, observability, foundation models, and the most underestimated line of all: human and agent identity. It decides whether you can audit what your agents did.

Once you see the whole stack, one thing becomes obvious: most of it is not yours to build, and the technology is still moving fast. Which is exactly why "wait" stops being a cop-out and becomes a normal, defensible move.

2. Build vs Partner vs Buy vs Defer: the decision

This is the half that should land on every leader's desk in 2026. McKinsey's decision tree reads top to bottom. I walk through it below, each time with the translation into procurement language.

The Build / Partner / Buy / Defer decision tree, after McKinsey QuantumBlack.

Question 1: is there a strategic reason to build the capability?

If YES, move to question 2 (Partner or Build). If NO, drop to the Buy / Defer branch.

Procurement reading: 80% of agentic use cases in procurement (spend categorization, invoice processing, RFQ generation, contract analysis) are not differentiating. The "let's code it in-house with Claude / GPT" reflex usually costs 3 to 5x more over 3 years than an off-the-shelf solution, once you add run, maintenance and model retrofit. Rule of thumb: if it isn't in the company's top 3 competitive levers, the answer is NO.

Question 2: can we partner to guarantee our requirements?

If YES, it's Partner. If NO, it's Build.

Procurement reading: this is where the negotiation happens. A real partnership implies evolution SLAs (not just uptime), an exclusivity or co-development clause, oversight on fine-tuning with your data, and reversibility on trained artifacts (weights, prompts, embeddings). If the vendor refuses, the "partnership" is disguised SaaS: treat it as Buy.

Question 3: a fit-for-purpose solution you can integrate while keeping control?

If YES, it's Buy (favoring open-source). If NO, move to question 4.

Procurement reading: "influence on feature roadmap" is rare in RFPs, and that's a mistake. On an agent running a critical process, no roadmap influence means total dependence on the vendor's calendar. Make it a weighted criterion.

Question 4: is the impact of deferring larger than the TCO of Build / Partner?

If YES, go back up to Partner or Build. If NO, it's Defer.

Procurement reading: the most counter-intuitive question, and the most powerful. It legitimizes doing nothing as an active decision. On a technology depreciating 30 to 50% a year, waiting 6 to 12 months is often the highest-ROI scenario. Document it so no one can hold it against you later.

In 100% of the agentic committees I've watched, no one ever seriously proposed "let's wait 12 months." Yet it's often the most profitable.

What this really means for procurement

Setting up real AI infrastructure is going to be complicated and costly. Procurement needs to be there to help build this ecosystem, or costs spiral out of control. Exploding AI bills already show it, and we are at the very beginning. Yet the upside is great when the technology is applied correctly, so businesses, and procurement teams, should experiment and take the topic on rather than sit it out. The job is not to block AI. It's to make sure someone in the room is asking what it costs, what it locks you into, and whether this quarter is the right one to commit.

3. How to operationalize this

3.1 Add the decision tree to the RFP kit

Before any consultation on an agentic use case, ask the business to answer the four questions in writing, with justification. You get a record of the make-or-buy decision, an argument to counter vendors going direct to the business, and a frame to challenge IT when it pushes Build.

3.2 Build a weighting grid across the five layers

Today's agentic RFPs over-weight the visible layer (UX, marketplace) and under-weight the deep layers (runtimes, observability, identity). A healthy grid splits the points across the five layers from Part 1:

3.3 Force the Defer question in every committee

Scenario planning must always include a "do nothing for 12 months" case with a quantified opportunity cost. It's not sabotage, it's financial hygiene. And in 30 to 50% of cases, it's the scenario that wins.

The bottom line

The McKinsey framework invents nothing: it's a classic make-or-buy applied to a new object. But it has two merits. It gives a reasonably simple way to decide in a very uncertain environment. And it legitimizes Defer as an active decision, probably the most profitable posture against a still-immature technology. For procurement, it's a chance to take back files that often run straight between IT and the vendor, by imposing a decision frame, a weighting grid, and the "what if we waited?" question in every committee.

None of this is an argument against buying AI. The upside is real when you apply it well. So experiment, take the topic on, and make sure procurement is in the room while you do, because that is how the bill stays sane and the ecosystem gets built on purpose rather than by accident.

Want to stress-test your next agentic RFP before you launch it? Book 30 minutes here.

I'm Alex Lio. 10+ years in indirect procurement, 7 of them at Amazon, now working alongside my clients on digitalization and AI.

Sources

1. McKinsey & Company, Rethinking enterprise architecture for the agentic era.
1. QuantumBlack, AI by McKinsey, Creating a future-proof enterprise agentic platform architecture (architecture diagram + Build / Partner / Buy / Defer tree).
1. McKinsey & Company, Indirect procurement: Insource, outsource, or both.