The agent-stack consolidation — how four Fortune 100 buyers picked in Q2.

UnitedHealth's Optum Insight: the Bedrock-Anthropic consolidation

Optum Insight, the data and analytics division inside UnitedHealth Group, ran an agent-stack consolidation procurement between January and April 2026. The starting state was a fragmented architecture: 14 distinct agent deployments running across LangChain orchestration, a mix of OpenAI and Anthropic models on Azure OpenAI Service and Bedrock respectively, three separate vector database deployments (Pinecone, Weaviate, and a Snowflake-native vector workload), and four different observability and governance tools cobbled together by the engineering teams over 18 months. The fragmentation had become an operational concern by Q4 2025, when an internal audit identified 81 distinct configuration parameters across the agent stack that were not centrally documented and could not be modified without coordinated change-management across multiple engineering teams. The consolidation procurement was the response.

The winning configuration is AWS Bedrock Agents as the primary orchestration runtime, Anthropic Claude Opus 4.7 and Claude Sonnet 4.6 as the primary model providers (with Opus reserved for high-context reasoning workloads and Sonnet for high-volume retrieval and routing), Bedrock Knowledge Bases as the primary vector retrieval surface, and AWS-native CloudWatch and CloudTrail integration as the primary observability and audit surface. The procurement value across the three-year horizon is approximately $31 million in committed AWS spend, allocated across Bedrock Agents license, Bedrock inference, and the associated AWS infrastructure consumption. The decision criteria documented in the procurement memo are: data residency posture (Optum's HIPAA-regulated data must remain in AWS US-East-1 and US-East-2 regions with a documented HIPAA Business Associate Agreement, which the Bedrock posture satisfied without exception); native integration with UnitedHealth's existing AWS-resident infrastructure (Optum's primary data warehouse runs on AWS, and Bedrock's integration with that warehouse was procurement-decisive); and a single-vendor accountability model for the orchestration, inference, and governance layer (AWS as the prime contractor, with Anthropic as a named subcontractor under the AWS master agreement).

The runners-up in the Optum evaluation were instructive. Microsoft's Azure AI Foundry plus Azure OpenAI Service was the second-place evaluation, narrowly losing on the data warehouse integration criterion. The Optum procurement team rated Foundry as architecturally comparable to Bedrock Agents and rated GPT-4o and the o3 family as comparable to Claude on the reasoning workloads, but the warehouse integration friction was the decisive cost difference over the three-year horizon. Google's Vertex Agent Builder plus Gemini 3 Pro was the third-place evaluation, losing primarily on the HIPAA contracting posture, where Google Cloud's healthcare-vertical contracting was rated less mature than AWS or Azure for a buyer of Optum's regulatory profile. Salesforce Agentforce and ServiceNow's Now Assist were both evaluated as adjacent rather than primary stacks — they would compose on top of the primary orchestration runtime rather than replace it — and the Optum architecture team has retained Agentforce for specific customer-success workflow surfaces that do not interact with HIPAA-regulated data.

The eliminations are the consolidation outcome. Optum eliminated LangChain orchestration from its primary architecture, Pinecone and Weaviate from its vector retrieval architecture, and three of the four observability tools that had accumulated. The LangChain elimination is the most consequential single signal in the Optum consolidation, because Optum was one of LangChain's largest healthcare deployments by token volume and the elimination represents a procurement vote against orchestration-framework-as-primary in favour of runtime-as-primary. The Optum architecture team's published rationale, circulated internally and forwarded to procurement counterparts at three other Fortune 100 buyers, is that LangChain's value proposition as a model-agnostic orchestration layer is genuine but is not procurement-decisive in a single-cloud, single-model-provider deployment. The model-agnostic value is purchased at the cost of an additional governance and observability layer that the primary runtime now provides natively. Optum decided that the additional layer was not worth the operational overhead.

The renewal-risk math for the Optum deployment is instructive. The three-year AWS commitment carries an early-termination clause that imposes a 40 per cent penalty on the remaining contract value, capped at $4 million for the first 12 months and stepping down to $1.5 million for the final 12 months. The clause is the procurement-side leverage against Bedrock Agents quality regressions. The Optum procurement team's published internal scorecard requires Bedrock to maintain a defined set of operational performance thresholds (P95 latency, availability, model output quality on a quarterly internal evaluation), with breach triggering a renegotiation right and, in the case of cumulative breaches, a contract exit right that bypasses the early-termination penalty. The clause is non-standard in cloud contracts of this size and reflects the procurement leverage that a $31 million committed-spend buyer can exercise. AWS accepted the clause. The procurement signal is that competitive pressure inside Bedrock is sufficient that AWS conceded contract-protection terms it would not have conceded 18 months ago.

ING Wholesale Banking: the multi-vendor architecture

ING Group's wholesale banking division ran a parallel consolidation between February and May 2026, with a structurally different outcome. The starting state for ING was less fragmented than Optum's — the bank had been operating under a more centralised technology architecture inherited from its 2023-2024 platform modernisation — but it included three significant strategic constraints: EU data residency for European wholesale banking data, AI Act compliance under the systemic-risk model provisions for high-impact use cases, and a board-level commitment to vendor diversification that the bank's chief risk officer Ljiljana Cortan articulated in the 2025 annual report. The diversification commitment was the procurement-shaping constraint that produced the multi-vendor outcome.

ING's winning configuration is a deliberately multi-vendor stack: Azure AI Foundry as the primary orchestration runtime, with Anthropic Claude Opus 4.7 and OpenAI GPT-5 Pro both running inside Foundry as primary model providers (with workload-specific routing logic), Microsoft Fabric as the primary data and analytics surface, and a governance layer built on top of Foundry's native compliance tooling extended with a custom rule library developed in partnership with ING's internal risk-engineering team. The procurement value across the three-year horizon is approximately $22 million in committed Azure spend, allocated across Foundry, Azure OpenAI, the Anthropic-on-Azure inference, and Fabric. The Anthropic-on-Azure availability — through Microsoft's strategic partnership with Anthropic and the corresponding inference availability inside Foundry — was procurement-decisive for ING, because it allowed the bank to maintain its single-cloud strategic posture (Azure as the primary cloud) while preserving model-provider diversification.

The decision criteria documented in the ING procurement memo are: EU data residency posture (Azure Foundry's commitment to EU-domiciled inference and Azure's EU Sovereign Cloud offering, which the procurement team rated as the most mature of the EU sovereignty-grade hyperscaler offerings); AI Act compliance documentation (the alignment between Foundry's compliance tooling and the EU AI Act systemic-risk provisions, where Microsoft has published more detailed compliance documentation than the AWS or Google equivalents); model-provider diversification (the ability to operate both Anthropic and OpenAI inference inside a single runtime, which neither AWS Bedrock nor Google Vertex offers at the same parity in May 2026); and the strategic relationship with Microsoft (ING has been a strategic Microsoft customer since 2018 and the wholesale banking division's existing Fabric deployment was a procurement entry point). The criteria together produced the Foundry-primary, dual-model outcome.

The runners-up in the ING evaluation were AWS Bedrock Agents and Google Vertex Agent Builder. Bedrock lost on the model-provider diversification criterion — Bedrock's primary positioning is Anthropic Claude with secondary positioning for Meta Llama, Cohere Command, and AI21 Jamba, and the OpenAI parity that Foundry offers is not currently available inside Bedrock. Vertex lost on the EU sovereignty posture; Google Cloud's EU sovereignty offering is structurally less mature than Azure's Sovereign Cloud and Microsoft's broader EU compliance posture. Mistral, evaluated as an EU-native alternative, was retained as a secondary model provider for specific French-language workloads but was not selected as a primary model provider. Aleph Alpha was evaluated and not retained; the ING procurement team rated Aleph Alpha as commercially less mature for the wholesale banking deployment scale.

The ING consolidation eliminated five legacy agent deployments and consolidated three vector retrieval implementations into a single Microsoft Fabric-resident vector workload. The structural significance is the explicit multi-vendor commitment at the model layer. ING is the first Fortune 100 buyer to publicly commit to a dual-model primary architecture (Anthropic and OpenAI) inside a single runtime, and the commitment is the procurement-side response to the model-provider concentration risk that the bank's chief risk officer has been articulating since the 2024 annual report. The risk position is structurally informative: ING has decided that single-model-provider lock-in is a more material risk than the additional operational complexity of dual-model orchestration. The position will be tested as the dual-model architecture goes into production through Q3 2026, and the operational data from that production cycle will be the most consequential field test of whether the dual-model pattern can hold inside a regulated enterprise at scale.

McKinsey's Lilli: the agent-native rebuild

McKinsey & Company's internal Lilli platform — the firm's proprietary generative AI surface for its 45,000 consultants — went through a substantial architectural rebuild between November 2025 and April 2026. The rebuild is not strictly a consolidation in the sense that the other three buyers' decisions are, because Lilli was already a relatively consolidated stack at the start of the rebuild. The rebuild is more accurately a re-platforming on agent-native primitives, and the result is a configuration that displaces multiple components that had been adequate in 2024 but had reached their architectural limits as the agent surface evolved. The decision team inside McKinsey was led by Helen Mayhew, the firm's chief technology officer at QuantumBlack, and the operational lead inside the Lilli team was the platform engineering director Erik Roth.

The winning configuration for the rebuilt Lilli is Microsoft Azure AI Foundry as the primary runtime, Anthropic Claude as the primary model provider with OpenAI as a secondary model provider for specific code-generation workloads, a custom McKinsey-developed retrieval layer built on top of Foundry that integrates the firm's proprietary research and client engagement archives, and a governance layer that extends Foundry's native compliance tooling with McKinsey-specific rules for client confidentiality, conflict-of-interest checks, and partner-level review workflows. The procurement value, by intelligence triangulated from McKinsey's published technology spend commentary and the procurement-side conversations, is between $14 million and $19 million in committed Azure spend over three years. The figure understates the total Lilli investment because McKinsey's internal engineering spend on the rebuild is substantial and is not included in the committed vendor spend.

The decision criteria documented in the McKinsey procurement memo are: confidentiality posture (Lilli operates against McKinsey's most sensitive client engagement data, and the procurement team required vendor commitments on data isolation, contractor access, and personnel security that Microsoft was uniquely positioned to provide given the existing M365 Enterprise relationship); model performance on long-context reasoning workloads (Claude Opus 4.7's 1-million-context handling was rated as substantially better than the alternatives on the long-document tasks that the consulting workload generates); developer-tooling integration (the McKinsey engineering team's existing development environment is Microsoft-aligned, and the Foundry integration with that environment was procurement-decisive); and the firm's strategic Microsoft relationship (McKinsey is one of Microsoft's largest enterprise customers by seat count and the strategic relationship was a procurement-shaping factor).

The runners-up in the McKinsey evaluation were AWS Bedrock Agents and the Anthropic-direct integration. Bedrock lost on the strategic Microsoft relationship criterion and on the developer-tooling integration criterion; McKinsey's engineering environment is more Microsoft-aligned than AWS-aligned, and the friction of operating against AWS for the Lilli rebuild was rated as materially greater than the alternative. The Anthropic-direct integration — using Anthropic's own infrastructure rather than the Foundry-hosted Anthropic inference — was evaluated as architecturally simpler but procurement-friction-heavier; McKinsey's procurement office preferred to operate under the existing Microsoft master agreement rather than negotiate a separate Anthropic enterprise agreement at the scale of the rebuild. The procurement-friction calculus has been a consistent theme across the four field-study buyers and is one of the structural reasons that Bedrock and Foundry are winning the orchestration runtime competition against the direct-vendor alternatives.

The McKinsey rebuild eliminated a previous LangChain-based orchestration layer that had been running since the original Lilli launch in 2023, eliminated a separate Pinecone vector deployment in favour of the Foundry-resident retrieval surface, and consolidated three observability tools into the Foundry-native monitoring surface plus a McKinsey-specific telemetry overlay. The LangChain elimination at McKinsey is structurally significant in the same way that the LangChain elimination at Optum is: it is a procurement vote against orchestration-framework-as-primary at a Fortune 100 reference customer. McKinsey's published rationale, communicated by Roth at the firm's internal technology summit in March 2026, is that the Lilli rebuild needed a primary runtime that could provide the governance, observability, and compliance integration natively, and that the LangChain-on-top-of-Foundry alternative produced operational complexity that the rebuild was specifically designed to eliminate.

FedEx Enterprise Logistics Intelligence: the Salesforce-Agentforce anchor

FedEx's enterprise logistics intelligence group ran a consolidation procurement between January and April 2026 that produced the most structurally divergent outcome of the four field-study buyers. FedEx selected Salesforce Agentforce as the primary orchestration surface for its customer-facing logistics intelligence deployments, with Anthropic Claude as the primary model provider through Salesforce's strategic partnership and Snowflake as the primary data surface. The procurement value across the three-year horizon is approximately $17 million in committed Salesforce spend, with the Anthropic and Snowflake components flowing through the Salesforce contracting envelope. The Agentforce-primary choice is the structurally divergent decision because the other three field-study buyers selected hyperscaler-primary architectures (Bedrock at Optum, Foundry at ING and McKinsey), and Agentforce is positioned as an application-platform-primary alternative rather than a hyperscaler-primary one.

The FedEx procurement memo documents the decision criteria explicitly. First, the customer relationship management surface: FedEx's customer-facing logistics intelligence deployment must integrate natively with the customer record, the engagement history, the service-level agreement tracking, and the case management surface that Salesforce CRM already manages for FedEx's enterprise customer base. The Agentforce integration with the underlying Salesforce CRM is architecturally tighter than any hyperscaler-primary alternative could provide. Second, the agent-action surface: Agentforce's design pattern, in which agents take parameterised actions against the Salesforce data model under explicit permission constraints, mapped directly to the FedEx use case requirements for customer-facing agents that can quote service options, schedule pickups, and resolve service exceptions. Third, the existing Salesforce relationship: FedEx has been a major Salesforce customer since 2014, and the strategic procurement relationship made the Agentforce contracting envelope structurally easier to negotiate than a new hyperscaler-primary procurement.

The runners-up in the FedEx evaluation were AWS Bedrock Agents with a Salesforce CRM integration overlay, and Microsoft Azure AI Foundry with a Dynamics 365 alternative to Salesforce. Bedrock lost on the CRM integration depth — the AWS-Salesforce integration was rated as architecturally more complex than the native Salesforce alternative — and the Bedrock procurement envelope would have required FedEx to negotiate separately with AWS and Salesforce, increasing procurement friction. Foundry plus Dynamics 365 was evaluated as a strategic alternative that would have required FedEx to migrate from Salesforce CRM to Dynamics 365 — a change that the FedEx procurement team rated as out of scope for the agent consolidation procurement and that would have generated migration costs in the $40 million to $60 million range that the consolidation business case could not absorb. The Foundry alternative was therefore not procurement-credible inside the FedEx context, even though it was rated as technologically comparable to the winning Agentforce configuration.

The FedEx consolidation eliminated four legacy agent deployments that had been running across a mix of Salesforce Service Cloud automations, custom internally-developed bots, and a small ServiceNow deployment that had been used for internal IT service management. The ServiceNow deployment was retained for internal IT service management — it was scoped out of the customer-facing agent consolidation — but the customer-facing automations were consolidated into the Agentforce surface. The procurement signal is that for FedEx, the application-platform-primary architecture is the right choice because the dominant constraint is CRM integration depth rather than hyperscaler-primary considerations like data residency or model-provider diversification. The four field-study buyers, taken together, illustrate that the right consolidation outcome is buyer-specific and depends on which constraint dominates the procurement-decisive criteria: Optum's HIPAA-data residency leads to hyperscaler-primary, ING's EU sovereignty and model-provider diversification leads to hyperscaler-primary with dual-model orchestration, McKinsey's strategic Microsoft relationship leads to Foundry, and FedEx's CRM integration depth leads to Agentforce.

The LangChain–Bedrock Agents–Agentforce three-way race

The four field-study consolidations together produce a structural signal about the three-way competition between LangChain, AWS Bedrock Agents, and Salesforce Agentforce for the procurement-primary agent runtime position at the Fortune 100 tier. The signal is not a clean victory for any of the three. It is a redistribution of the addressable procurement category along the dominant-constraint dimension. LangChain lost ground in all four field-study consolidations. Bedrock won at Optum and was the runner-up at McKinsey and FedEx. Agentforce won at FedEx and was an adjacent rather than primary stack at Optum. Foundry — which is not one of the three named runtimes but is structurally part of the same category — won at ING and McKinsey. The most consequential pattern is the LangChain ground-loss, which appeared in all four consolidations.

The LangChain ground-loss is not the same as a LangChain commercial failure. LangChain is still the dominant orchestration framework by deployment count across the Fortune 1000, with an installed base that exceeds Bedrock Agents and Agentforce combined. The ground-loss is specifically at the Fortune 100 procurement-primary position, where the four field-study buyers all decided that the orchestration-framework-as-primary architecture produced more operational complexity than the runtime-as-primary alternative. LangChain's response strategy, articulated by Harrison Chase at the company's recent NYC developer summit, is that the orchestration-framework category provides model-agnostic portability that the runtime-primary alternatives do not, and that the portability value will become more visible as the model-provider competition intensifies. The argument is internally coherent. It is also the argument that LangChain has been making for two years, and the four field-study buyers' decisions indicate that the procurement-grade buyers are not currently weighting the model-portability value as procurement-decisive.

Bedrock Agents has accumulated the strongest position in the four-way pattern. The Bedrock win at Optum and the Bedrock runner-up positions at McKinsey and FedEx indicate that for buyers whose dominant constraint is data residency, regulated industry posture, or AWS-strategic-relationship, Bedrock is the procurement-preferred choice. The AWS strategic position with Anthropic — which provides Bedrock with primary access to the Claude family and which Anthropic prices through Bedrock at parity with the direct Anthropic API — is the structural advantage that Bedrock has been building on since the AWS-Anthropic partnership deepened in 2023-2024. The Bedrock Agents commercial trajectory through 2026 will depend on whether AWS can convert its Optum-class wins into a broader Fortune 100 procurement pattern, and the early field-study evidence indicates that the conversion rate is high in the segment where AWS has the strategic relationship advantage.

Agentforce's position is structurally different from Bedrock's. Agentforce is winning where the dominant procurement constraint is CRM integration depth or application-platform-primary considerations, and Agentforce is not currently winning in the hyperscaler-primary segment where Bedrock and Foundry are dominant. The Salesforce go-to-market is positioned to address that asymmetry through the Agentforce-on-Hyperscaler integration narrative, in which Agentforce orchestrates agent workflows that span Salesforce's data model and hyperscaler-resident data surfaces. The integration narrative is plausible but has not yet been demonstrated at Fortune 100 procurement scale. The FedEx win is the first major reference for Agentforce-primary at the Fortune 100 tier, and the FedEx deployment will be the first production test of whether the Agentforce architecture can scale to the customer-facing logistics intelligence volumes that FedEx is committing to it.

The three-way race is therefore not a race for a single winner. It is a category-redistribution in which three different runtime architectures will occupy three different segments of the procurement-grade enterprise agent market: Bedrock Agents and Azure AI Foundry split the hyperscaler-primary segment along AWS-versus-Microsoft strategic relationships, Agentforce occupies the application-platform-primary segment where CRM-integration is the dominant constraint, and LangChain occupies the model-agnostic orchestration segment where the buyer prioritises model portability and is willing to accept the operational complexity of running the orchestration layer separately from the runtime. The four field-study consolidations illustrate the pattern. The next 12 months of Fortune 100 procurement will test whether the pattern is stable.

What to watch

The four field-study consolidations are the procurement-grade evidence base. The Q3-Q4 2026 cycle will determine whether the patterns generalise across the broader Fortune 100 cohort or fragment as the runtime category continues to mature.

• Whether the ING dual-model architecture holds operationally through Q3 2026 production; the dual-model pattern is the most architecturally novel of the four consolidation outcomes and the production-cycle operational data will determine whether the pattern can spread to other regulated buyers who have been articulating model-provider concentration risk.
• Whether AWS Bedrock Agents converts its Optum-class win into a broader healthcare-vertical procurement pattern; UnitedHealth, Anthem (Elevance Health), CVS Health, and Humana are all running parallel agent-stack consolidations in 2026 and the Bedrock pattern from Optum will be the procurement reference that the other three buyers' procurement teams evaluate against.
• Whether the LangChain ground-loss at the Fortune 100 tier triggers a strategic adjustment in LangChain's commercial positioning; Harrison Chase has signalled that LangGraph will continue to be positioned as the model-agnostic primary runtime, but the four-of-four ground-loss in the field-study consolidations indicates a procurement-grade structural challenge that LangChain's response will need to address materially.
• Whether Microsoft Azure AI Foundry continues to win the dual-model orchestration competition against AWS Bedrock; the Anthropic-on-Azure availability is the structural feature that gave ING the dual-model option inside Foundry, and AWS's response will determine whether Bedrock can add OpenAI-on-Bedrock parity in a credible timeline.
• Whether Agentforce's FedEx win produces a broader application-platform-primary consolidation pattern across customer-facing enterprise agent deployments; Walmart, Target, Walgreens, and Home Depot are all running customer-facing agent procurements in 2026 and the Agentforce pattern from FedEx will be the precedent that those procurement teams evaluate against.

Frequently asked

What is the most consistent pattern across the four consolidation outcomes?

The most consistent pattern is the consolidation toward a primary orchestration runtime — whether AWS Bedrock Agents, Microsoft Azure AI Foundry, or Salesforce Agentforce — paired with a primary model provider and a primary governance layer. All four buyers eliminated multiple legacy agent vendors and converged on three-component primary architectures. The model provider varied (Anthropic in three of four primary positions, dual-Anthropic-OpenAI at ING), the orchestration runtime varied across the four buyers, but the structural pattern of runtime-primary, model-primary, governance-primary is consistent. The pattern indicates that the procurement category is settling at the Fortune 100 tier into three-component reference architectures rather than the multi-vendor sprawl that characterised the 2024-2025 deployment period.

Why did LangChain lose ground in all four consolidations?

LangChain lost ground because all four buyers decided that the orchestration-framework-as-primary architecture produced more operational complexity than the runtime-as-primary alternative, particularly when the primary runtime (Bedrock, Foundry, or Agentforce) now provides governance, observability, and compliance integration natively. LangChain's model-agnostic value proposition is genuine but is not procurement-decisive in single-cloud, single-model-provider deployments, which is what three of the four field-study buyers selected. ING's dual-model architecture is the exception, but even ING chose Foundry as the primary runtime rather than LangChain-on-Foundry. The procurement-grade buyers are not currently weighting model-portability as the dominant criterion, and that weighting is what determines LangChain's commercial position at the Fortune 100 tier.

What makes the ING dual-model architecture structurally novel?

ING is the first Fortune 100 buyer to publicly commit to a dual-model primary architecture (Anthropic Claude and OpenAI GPT) inside a single runtime, driven by an explicit board-level commitment to vendor diversification articulated by the chief risk officer in the 2025 annual report. The architecture is structurally novel because it accepts the additional operational complexity of dual-model orchestration in exchange for reducing single-model-provider concentration risk. The bank's calculation is that single-model lock-in is a more material risk than the dual-model operational overhead. The architecture is enabled by Azure AI Foundry's ability to host both Anthropic and OpenAI inference at parity, which is a feature that neither AWS Bedrock nor Google Vertex currently offers at the same parity in May 2026.

How does the McKinsey Lilli rebuild differ from the other three consolidations?

The McKinsey Lilli rebuild is technically not a consolidation in the same sense as the other three; Lilli was already a relatively consolidated stack at the start of the rebuild. The rebuild is more accurately a re-platforming on agent-native primitives, replacing a 2023-era LangChain-on-OpenAI architecture with a Foundry-on-Anthropic architecture optimised for the long-context reasoning workloads that the consulting business generates. The decision criteria — confidentiality posture, long-context reasoning performance, developer-tooling integration, strategic Microsoft relationship — produced an outcome architecturally similar to the ING decision, with the structural difference that McKinsey selected Anthropic-primary with OpenAI-secondary rather than the ING dual-model primary configuration.

What is the procurement signal from the contract-protection clause in the Optum-AWS agreement?

The clause — which imposes a 40 per cent penalty on remaining contract value for early termination, capped at $4 million for the first 12 months and stepping down to $1.5 million for the final 12 months, with breach-trigger renegotiation and exit rights — is non-standard in cloud contracts of this size. Its inclusion indicates that competitive pressure inside the Bedrock procurement category is sufficient that AWS conceded contract-protection terms it would not have conceded 18 months ago. The procurement signal is that Fortune 100 buyers with $30 million-class committed-spend leverage can now negotiate operational performance protections that were not procurement-available in the previous procurement cycle. The signal is the procurement-leverage redistribution from vendor to buyer that the maturation of the agent-runtime category is producing.

What is the broader implication for the agent-stack category beyond the four field-study buyers?

The broader implication is that the agent-stack category is redistributing along the dominant-constraint dimension at the Fortune 100 tier, with three different runtime architectures occupying three different procurement segments: Bedrock Agents and Azure AI Foundry split the hyperscaler-primary segment along AWS-versus-Microsoft strategic relationships, Agentforce occupies the application-platform-primary segment where CRM-integration is the dominant constraint, and LangChain occupies the model-agnostic orchestration segment for buyers who prioritise model portability. The pattern is not vendor convergence; it is category-redistribution. The implication for procurement teams not yet through their consolidation cycle is that the right consolidation outcome depends on which constraint dominates their own decision criteria, and that the four-way field-study pattern provides four different procurement-grade reference outcomes.