The AI Decision Your IT Team Already Missed

Between May 11 and May 19, three of the largest distribution channels in enterprise software made the same vendor call. SAP embedded Claude as the primary reasoning and agentic capability inside Joule on the SAP Business AI Platform, a surface that reaches 300 million end users. KPMG put Claude in front of 276,000 employees across 138 countries and inside its Digital Gateway client platform. PwC expanded its alliance to put Claude Code in front of 30,000 US professionals, with the full 364,000-person rollout already underway. Nine days. Three deals. One model.

Most enterprise AI steering committees are still in workshop two of their vendor evaluation. The decision they were going to make in Q3 already got made for them in May, by the ERP partner and the audit firm and the consulting bench they outsourced “AI strategy” to last year.

That’s the governance story. The procurement process companies trained themselves on after 2010 was built to police vendor decisions that happen inside the IT perimeter. Claude got into the org through three doors none of those processes are watching.

Quick Verdict

Signal	What It Means
SAP-Anthropic alliance (May 11-13)	Claude is the reasoning layer inside Joule for finance, HR, procurement, supply chain — 300M+ end users
KPMG Digital Gateway Powered by Claude (May 19)	276,000 KPMG employees in 138 countries, plus the client audit and advisory platform
PwC alliance expansion (May 14)	30,000 US professionals on Claude Code, 364,000 global staff rollout underway, up to 70% delivery improvement reported
Time between first and last announcement	9 days
Where the vendor decision actually got made	At the SAP platform, the KPMG partnership, the PwC delivery method — not in your steering committee
Where your AI governance program is looking	At the model your team chose to pilot, not the model your ERP and audit partners just installed
Your real move this week	Map the Claude footprint your platform and consulting vendors are already shipping into your environment

The Nine-Day Window

The sequencing is the first thing worth slowing down on. SAP went first at Sapphire Orlando on May 11-13, with Anthropic President Daniela Amodei appearing by video and a partnership announcement that puts Claude inside the Joule Assistants for finance, HR, procurement, and supply chain. The Treasury Manager demo prepared a CFO briefing with live ERP data in minutes. The Joule Studio agent build flow runs against Claude by default.

PwC followed on May 14 with the alliance expansion. Claude Code firmwide. 30,000 professionals trained in the first wave. A joint Center of Excellence targeting financial services, pharma, life sciences, healthcare, and consumer markets. Insurance underwriting compressed from ten weeks to ten days. Cybersecurity incident response cut from hours to minutes. The 364,000-person global rollout is the destination, not the headline.

KPMG closed the trio on May 19 with the most aggressive of the three deployments. The full 276,000-person workforce gets Claude. The Digital Gateway platform, KPMG’s client audit and advisory environment built on Azure, now runs Claude inside the agent layer. Regulatory compliance agents that previously took weeks are completing in minutes. The initial focus is tax clients and private equity portfolio companies.

Three different deal structures. Three different parts of the enterprise stack. Same vendor selection inside a window most procurement cycles cannot react inside.

What Just Happened to the Enterprise Procurement Model

The post-cloud procurement playbook for enterprise AI looked something like this. Spin up an AI steering committee. Inventory the use cases. Shortlist three to five model vendors. Run a bake-off. Pick one or two. Negotiate enterprise terms. Sign. The cycle takes six to nine months in a healthy org, eighteen in a regulated one. That’s the model every CIO who lived through the Salesforce and ServiceNow rollouts is running.

The cycle assumes the vendor decision is yours to make. May 11 through May 19 quietly retired that assumption inside any org that runs on SAP, uses KPMG or PwC for audit and advisory, or has a meaningful consulting spend with the Big Four.

If your ERP is SAP, your finance team is going to be talking to Joule, and Joule is now talking to Claude. If your audit firm is KPMG, your engagement teams are running Claude inside Digital Gateway against your data. If PwC is delivering a major program for you, the consultants in your conference room are working in Claude Code. None of those decisions required a sign-off from your steering committee. None of them is waiting for one.

This is the same pattern I described in the enterprise AI vote analysis at the buyer level. The vendor consolidation is happening, the marginal new purchase is going to Anthropic, and the formal procurement process is the last thing to acknowledge it.

What is the Enterprise AI Governance Blind Spot?

The enterprise AI governance blind spot is the gap between the model your steering committee thinks it is evaluating and the model your platform partners, audit firms, and management consultants are already running against your data. Committees are scoped to vet vendor contracts the company signs directly. The Claude footprint inside SAP Joule, inside KPMG Digital Gateway, and inside PwC delivery shows up under your ERP agreement, your audit engagement letter, and your consulting SOW respectively — none of which routes through AI vendor governance. The model is in the environment before the committee meets to evaluate it.

That has three operational consequences worth naming.

1. Your data is already moving through Claude through at least one of these channels if you are a midsize-to-large enterprise running SAP, audited by KPMG, or consulting with PwC. The question is not whether Claude is touching your data. It is which workflows it is touching and under whose data processing agreement.

2. Your model risk inventory is incomplete if it lists only the vendors you contracted directly. The Claude that runs inside Joule sits under the SAP MSA. The Claude inside Digital Gateway sits under the KPMG engagement letter. The Claude that PwC consultants run sits under the PwC SOW. Three different contracts, three different liability postures, one model.

3. Your AI policy probably doesn’t cover this because most enterprise AI policies were written against the assumption that the company would be the contracting party for the model. The “approved vendor list” approach breaks down when the model arrives as a feature of a platform or service the company already approved years ago.

The Real Decision That Got Made

The interesting part of these three deals is not that they all picked Claude. It is what the deal structures say about who made the decision and what process they ran.

SAP didn’t put out an RFP for the Joule reasoning layer. SAP picked Anthropic based on a year of partnership work and integration testing against the Model Context Protocol, which Anthropic authored. The decision was made by SAP product leadership and ratified by the joint engineering teams. Your steering committee wasn’t in the room.

KPMG didn’t run a model bake-off across its 276,000-person workforce. KPMG’s leadership decided Claude was the model it wanted in front of every employee and embedded in its client platform. The decision was made at the firm level, signed at the alliance level, and announced as a fait accompli. Your KPMG engagement team will use what KPMG bought.

PwC didn’t survey its 30,000 US professionals on model preference. PwC committed firmwide to Claude Code as the standard for delivering production software to clients. The early data, up to 70% delivery time reduction across reference deployments, is the rationale. Your PwC SOW doesn’t give you input on which model the consultants build with.

The pattern matters. None of the three decisions was made by an IT or risk function looking at model cards and SOC reports. All three were made by business leadership looking at delivery margins, productivity per consultant, and platform competitive positioning. The IT and risk functions inside SAP, KPMG, and PwC are downstream of the decision, the same way the IT and risk functions inside the customer base now are.

The Anti-Hype Read

Two cautions before the headline writes itself.

The first is that “300 million end users” is a SAP install-base number, not a Claude usage number. Most of those end users will never directly invoke an agentic workflow. The realistic year-one usage footprint is the population of finance, HR, procurement, and supply chain power users who actively work in Joule. That is a real number in the millions, not the hundreds of millions, and it matters because it sets the realistic data flow into Claude for the next 12 months.

The second is that PwC’s “up to 70% delivery time reduction” is a cherry-picked figure from the most successful early deployments. The realistic median is probably 25 to 40% across less-favorable use cases. That’s still enough to reshape consulting unit economics, but it isn’t the headline. The enterprise AI ROI reckoning I wrote about earlier applies here too. Vendor case studies pick the best-case projects. Your case will land somewhere lower.

Those cautions don’t change the governance story. The governance story isn’t about how much Claude usage actually materializes in year one. It’s about who chose the model and whether your existing controls were in the room when the choice happened.

What the Steering Committees Have to Do Now

The committees that are still operating on the 2024 playbook need to re-scope this quarter. Three moves are doable inside 30 days.

Move 1: Inventory the Claude exposure you already have. Not the Claude exposure you contracted for. The Claude exposure that arrived through your platform agreements and your consulting SOWs. Pull the SAP roadmap your account team is briefing on for Q3. Joule with Claude is going to be in it. Pull your active KPMG and PwC engagement letters and ask the engagement partners what AI tooling is being used in delivery. The answer is going to involve Claude. Document the footprint.

Move 2: Re-route the data processing agreements through the AI committee, not the platform committee. The SAP MSA your legal team signed in 2019 didn’t anticipate a frontier model sitting in Joule. The KPMG engagement letter didn’t anticipate Claude running against your audit data. The PwC SOW didn’t name Claude Code. None of those agreements is illegal. All of them are now incomplete from a model-risk and data-processing perspective. Get the AI governance function reviewing the addenda the platform and consulting partners are about to send you anyway.

Move 3: Stop running a vendor bake-off your platform partners already concluded. This part is going to be uncomfortable in some committees. If your enterprise architecture roadmap is built on SAP and you audit with KPMG and your delivery program runs through PwC, your model selection is functionally already Claude. Running a parallel bake-off between Claude, GPT-5, and Gemini at the steering-committee layer in that environment is a process exercise, not a decision exercise. Spend the cycles on integration governance and data-flow mapping instead. That’s the enterprise negotiation lever Anthropic is actually paying attention to right now.

These aren’t the moves a 2024 procurement consultant would tell you to make. They’re the moves the 2026 reality is telling you to make.

My Read

Three positions I am taking after the nine-day window.

The platform-and-consulting channel is now the most powerful vendor selection mechanism in enterprise AI. Not the buyer’s RFP. Not the analyst report. Not the security review. The model that gets embedded inside the ERP, the audit platform, and the consulting delivery method is the model that wins the next two years of enterprise usage. Anthropic just ran that play three times in nine days. OpenAI’s response will be the next thing worth watching.

Most AI steering committees are running an outdated process against the wrong question. The 2024 question was “which model do we approve.” The 2026 question is “which models are already running against our data, under whose contract, and where are the control gaps.” Committees that don’t pivot inside the next quarter will be doing model risk assessments on decisions that are six months old by the time the report lands.

The IT function is being routed around, and the explanation is structural, not political. Selling AI as a feature of an existing platform or service contract is faster than selling AI as a new tool that has to go through a separate procurement cycle. The SAP-n8n deal earlier this month was the first signal. The Claude embedding across SAP, KPMG, and PwC is the larger one. The platform and consulting layers have figured out that selling AI capabilities is faster than selling AI tools, and they aren’t asking the customer’s IT department to ratify the model underneath.

Your Three Moves Before the Next Steering Committee

Sized for a CIO, CISO, or AI program owner inside a mid-market or enterprise org running on SAP, audited by KPMG, or consulting with PwC. Doable inside 30 days. Will position your governance program against the May reality instead of inside the 2024 playbook.

1. Email your SAP, KPMG, and PwC account leads this week and ask for the Claude footprint document. The phrasing matters. Don’t ask “are you using AI in your delivery.” Ask “what is the current and planned Claude footprint inside the services you provide us, under which contract instrument, and what data processing addendum covers it.” All three vendors have a version of this document ready. Most customers haven’t asked for it.

2. Re-scope your next AI committee meeting around the embedded-model question. Pull the model-bake-off agenda. Replace it with a Claude-exposure mapping exercise across your platform and consulting vendors. Bring in legal and procurement, not just IT and risk. The first hour of that meeting will be uncomfortable. The output will be the first realistic AI risk inventory your org has produced.

3. Update your AI policy to cover embedded models, not just contracted models. The policy your team wrote in 2024 was probably built around a “list of approved vendors” model. That model breaks the moment the AI shows up as a feature of an ERP module or as the working medium of a consulting engagement. The fix is a clause that covers “AI capabilities introduced via third-party platforms and service providers” with named data flow and audit requirements. Three paragraphs. Most enterprise legal teams can ship it in two weeks.

Bottom Line

The vendor decision your steering committee was going to make in Q3 already got made by your ERP partner, your audit firm, and your top consulting vendor between May 11 and May 19. The model is Claude. The contract surface is the SAP MSA, the KPMG engagement letter, and the PwC SOW. The governance footprint is bigger than your AI policy assumed.

The good news is that the model selection is one most steering committees would have arrived at anyway. The Ramp data already showed 73% of new enterprise AI spend going to Anthropic this year. The bad news is that the procurement process you spent the last twelve months building was scoped against a vendor decision that no longer routes through it.

Stop running the bake-off. Start mapping the embedded footprint. Update the policy to cover the channel the model actually arrived through. The committees that pivot inside the next 30 days will be governing the AI deployment that is already happening. The committees that keep running the 2024 playbook will be writing model risk reports about a decision their platform partners closed two quarters ago.

The IT team didn’t lose the AI vendor decision. It wasn’t invited to it.

---

Related Reading:

• Big 4 Just Standardized on Claude Code
• SAP Just Made n8n the Enterprise Standard
• The Enterprise AI Vote Is In: Anthropic Won
• Anthropic Just Turned Profitable. Now Negotiate.
• The Enterprise AI ROI Reckoning