The Enterprise AI Stack: How Companies Will Deploy AI

Executive Intelligence · Special Report · November 14, 2026

As artificial intelligence moves from experimentation into core business operations, enterprises are building increasingly sophisticated technology stacks to develop, deploy, govern, and scale AI across their organizations. From foundation models and data infrastructure to AI agents, cloud platforms, security, and orchestration, the emerging enterprise AI stack will shape how companies integrate intelligence into everyday workflows and compete in an AI-driven economy.

The Enterprise AI Stack: How Companies Will Deploy AI

Central thesis. Enterprise AI will not be deployed through foundation models alone. It will depend on an integrated stack that connects computing infrastructure, cloud platforms, models, enterprise data, application development, AI agents, security, and business workflows. The most important distinction in this market is between companies that sell AI capabilities and companies that supply the infrastructure and operating systems that make AI useful inside a business.

1. Introduction: from experimentation to implementation

By late 2026, almost every large company is using AI somewhere. Far fewer have it running reliably inside the processes that generate revenue or control costs. The published numbers disagree, mostly because each survey defines “production,” “agent,” and “scale” differently:

A Recon Analytics survey of more than 120,000 enterprise respondents (March 2025 to January 2026) found only 8.6% of companies with AI agents deployed in production, with a further 14% still piloting them.
S&P Global Market Intelligence puts agents in production at scale near 11%, and Forrester has reported that about 88% of agent pilots never reach production.
In contrast, a KPMG pulse survey of large US organizations (Q2 2026) found 53% with agents deployed, and McKinsey’s 2026 survey reports 44% of respondents saying AI is scaling across the enterprise.

Read together, the data says adoption is broad, and depth is thin. Gartner has predicted that 40% of enterprise applications will include task-specific agents by the end of 2026, and has also warned that over 40% of agentic projects may be cancelled by the end of 2027 because of cost, unclear value, or weak risk controls. Both can be true, and the gap between them is the subject of this report: the surrounding stack decides which projects survive.

This report does not rank models. It maps the eight layers an enterprise has to assemble, the companies supplying each, four deployment models, an eight-step lifecycle, six worked scenarios, and the economics that determine whether any of it pays back.

2. The enterprise AI stack explained

The framework below is conceptual. It is not a mandatory build sequence or any single vendor’s architecture. Many platforms span several layers (Microsoft, AWS, Google, Databricks and Salesforce each do), and three concerns cut across every layer: security, governance and data management.

8	Business applications and industry solutions — where employees and customers actually meet AI
7	Security, identity and governance — who and what may access which data and actions
6	AI agents and orchestration — from answering questions to executing tasks
5	AI development, evaluation and operations — build, test, monitor, maintain
4	Data and knowledge infrastructure — proprietary information, retrieval, pipelines
3	Foundation models and inference — the reasoning and generation engines
2	Cloud and managed AI platforms — the procurement and control plane most enterprises start from
1	AI infrastructure — chips, networking, GPU capacity, data centers
Cutting across all layers: security · governance · data management · cost control

Figure: the eight-layer enterprise AI stack, from business applications (top) to computing infrastructure (bottom).

Layer	Why enterprises need it	Companies to evaluate
1. Infrastructure	Compute, networking, and power determine capacity, latency, and unit cost	NVIDIA, AMD, Broadcom, CoreWeave, Lambda, Equinix, Digital Realty
2. Cloud & managed platforms	One contract, one identity system and managed access to many models	AWS, Microsoft, Google Cloud, Oracle, IBM, Salesforce
3. Models & inference	Task fit, cost, latency, context, multimodality, data handling	OpenAI, Anthropic, Google DeepMind, Meta, Mistral, Cohere, Amazon
4. Data & knowledge	Models know nothing about the company until data is connected and governed	Databricks, Snowflake, MongoDB, Elastic, Pinecone, Redis, Confluent
5. Development, evaluation & ops	Probabilistic systems need testing and monitoring after launch	Hugging Face, Weights & Biases, LangChain, LlamaIndex, Arize, MLflow, Datadog
6. Agents & orchestration	Connects AI to systems of record so work is completed, not just discussed	Sierra, ServiceNow, Salesforce, UiPath, Glean, CrewAI, Microsoft
7. Security, identity & governance	Agents hold credentials and act; permissions and audit must keep up	CrowdStrike, Palo Alto Networks, Wiz, Okta, Microsoft, Protect AI
8. Applications & industry solutions	The layer where productivity, revenue, or cost impact is measured	Microsoft, Salesforce, ServiceNow, SAP, Workday, Adobe, Intuit, vertical specialists

3. The infrastructure foundation (Layer 1)

Demand for AI compute is the best-documented part of the stack, and the most concentrated. Reported figures from the latest quarters:

Company	Latest reported period	AI-relevant figure
NVIDIA	Quarter ended July 26, 2026	Revenue $96.2B; data center about $89B. Hyperscale about $49B; the “ACIE” group (industrial and regular enterprises, among others) about $40B, up 25% sequentially. Next-quarter guidance about $108B.
Broadcom	Quarter ended Aug. 2, 2026	Revenue $29.6B; AI semiconductor revenue $16.7B (custom accelerators and AI networking), up 221% year over year.
AMD	Q2 2026	Revenue $11.5B; data center $6.7B, up 107% year over year, with Helios rack-scale systems beginning to ramp.
CoreWeave	Q3 2026 (reported Nov. 10)	Revenue $3.66B, up 168%; backlog about $132.5B; more than 1.7 GW active power; full-year revenue guidance raised to $13.0–13.3B; debt reported near $47B.
Equinix	Q1 2026 / FY2026 outlook	About 60% of its largest Q1 deals were AI-related; full-year revenue guided to roughly $10.1–10.2B.

These figures are not directly comparable. NVIDIA reports a data center segment that includes networking; AMD reports CPUs and accelerators together; Broadcom reports a narrower AI semiconductor line; fiscal calendars differ.

Research angle: own, rent, or combine? The evidence points to a combination, with ownership as the exception. Three things drive that:

Utilization decides the economics. Uptime Institute analysis reported by Network World found that dedicated infrastructure beats an established cloud provider only above roughly 22% sustained utilization, and beats a neocloud only above roughly 66%. Few enterprises keep GPUs that busy.
Neoclouds sell capacity in committed blocks. CoreWeave and Lambda offer dense GPU clusters with high-speed networking, often at lower hourly rates than hyperscaler list prices, but with a narrower service catalog. They suit large reserved training and high-volume inference. A company with modest or unpredictable demand is usually better served by its cloud provider’s managed services.
Colocation is the bridge for private AI. Equinix and Digital Realty position themselves as places to keep sensitive data under company control while sitting close to the clouds that host the models. Their enterprise colocation business has been slower to benefit than their hyperscale leasing, but inference and private AI are the stated growth case.

Counterparty risk is part of the decision. CoreWeave’s backlog is large, but its interest expense has been reported as exceeding operating income, and its growth depends on continued financing. An enterprise signing a multi-year reservation with any neocloud should assess financial resilience, delivery timelines for new capacity, and portability of workloads.

4. Cloud platforms and the model layer (Layers 2 and 3)

Why enterprises start with the cloud they already use

The managed platforms from AWS (Amazon Bedrock and its AgentCore agent runtime), Microsoft (Microsoft Foundry, formerly Azure AI Foundry) and Google (Vertex AI, reportedly rebranded as the Gemini Enterprise Agent Platform in 2026) are converging on the same architecture: an agent runtime, memory, a tool gateway, identity, observability and governance. The decision therefore tends to follow existing commitments:

Identity and data gravity. Microsoft is strongest where Entra identity and Microsoft 365 content already form the substrate. AWS is the default for AWS-centric estates and broad model choice. Google is strongest for Gemini access and data-platform adjacency.
Committed spend. A pre-signed cloud commitment often decides the platform, because the same Claude models are offered on all three at similar list prices, while GPT models are first-party on Microsoft and Gemini is first-party on Google. Reports in 2026 indicate OpenAI models are beginning to appear on Bedrock in limited preview, which weakens the old assumption that OpenAI workloads mean Azure.
Real lock-in is not the API. It is the embeddings, vector stores, fine-tuned models, evaluation sets, and pipelines built on top of it.

Microsoft’s fiscal fourth quarter and October 28 call show how quickly the managed layer is scaling: more than 30 million paid Microsoft 365 Copilot seats; Agent 365, its control plane for agent governance, with nearly 40 million agents registered two months after launch and over 50 million by late October; and Foundry revenue more than doubling year over year. Google Cloud reported a $514 billion backlog in Q2 2026. Oracle Cloud Infrastructure (GPU capacity and enterprise applications), IBM (watsonx and hybrid deployments) and Salesforce (Agentforce, covered in Layer 6) serve customers whose AI decisions are tied to existing database, mainframe or CRM estates.

When a specialized platform wins: when a workload needs GPU price-performance the hyperscaler cannot match, when data residency or sovereignty rules prevent public-cloud use, or when a vendor’s application (for example, a CRM or IT service platform) already holds the workflow and data.

The model layer: multi-model is becoming the default

Evidence of share is mixed, because surveys measure different things (usage share, wallet share, adoption). Menlo Ventures’ mid-2025 survey put Anthropic at 32% of enterprise LLM usage, OpenAI at 25% and Google at 20%. An a16z survey of 100 Global 2000 CIOs (January 2026) found OpenAI still ahead at roughly 56% of wallet share, with 78% of respondents using its models in production, while Anthropic and Google were gaining. Enterprise Technology Research data reported by SaaStr shows OpenAI’s share slipping from a September 2025 peak of 62% to 56% by March 2026. The consistent finding is not who leads but that no one holds a lock.

Reasons enterprises are using several models at once:

Task fit. Models differ on coding, long documents, multilingual work, tool use, and multimodal inputs. a16z noted that routing across models is how application vendors capture each one’s strengths.
Cost tiers. Frontier models for hard reasoning, smaller or open models for classification, extraction and summarization. Microsoft cited up to 77% lower cost for a GitHub Copilot configuration it says delivers frontier-level quality; treat that as a vendor claim, but it shows why routing matters.
Resilience and negotiating leverage. A single provider is an outage, pricing, and roadmap risk.
Data handling and deployment constraints. Some workloads must stay in a region, a private cloud or on company hardware.

Vendor-reported multi-model figures vary (one 2026 infrastructure report cites average usage rising from 2.1 to 4.7 models per account), and should be read as directional.

Model-selection criterion	What to test
Accuracy on your tasks	A private evaluation set built from real documents, tickets or transactions, not public benchmarks
Inference cost and retries	Cost per completed task, including failures and re-runs
Latency	Time to useful answer under production load, not demo load
Context window and multimodality	Quality at the lengths you actually use; claimed maximums often degrade in practice
Data handling	Retention, training use, regional processing, audit logging, certifications
Deployment flexibility	API, managed cloud, private cloud or self-hosted weights; portability if you switch

Open models. Llama (Meta), Mistral, Qwen, DeepSeek and others now handle classification, extraction and well-scoped summarization at a level that is hard to distinguish from closed models, while complex reasoning and long code still tend to favor closed frontier models. The share figures conflict sharply: Menlo put open-source at 11% of enterprise API use in 2025, an infrastructure vendor reports 38% of its token volume in Q1 2026, and academic research cited by one 2026 review finds closed models still account for about 80% of worldwide token usage. Open models may carry a third of usage while earning a few percent of revenue. The enterprise reasons to use them are cost at scale, privacy and control, with licensing terms and data residency, rather than capability, cited as the main blockers. Mistral and Cohere are the two vendors on this list that pair open or enterprise-tuned models with private-deployment options; Amazon’s Nova models are the in-house option for AWS customers.

5. The data layer (Layer 4)

A model is only as useful as the company information it can safely reach. Retrieval-augmented generation (RAG) is the standard pattern: documents are chunked, embedded as vectors, stored in a search engine or vector database, and retrieved at question time, with the user’s permissions enforced on the results.

Data platforms: Databricks reported crossing a $7B revenue run-rate, growing more than 80% year over year (company-reported and unaudited, August 2026). Snowflake reported fiscal Q2 2027 product revenue of $1.49B, up 37%. Databricks has said its AI products reached a $1.4B run-rate earlier in 2026, against a much smaller disclosed AI line at Snowflake. The competition is now about who hosts the governed data that agents will use.
Operational and search databases: MongoDB (documents plus vector search), Elastic (search and retrieval) and Redis (caching and vector search) let teams add retrieval to systems they already run. Pinecone is a dedicated vector database for teams that want a separate retrieval service.
Streaming: Confluent supplies real-time event data, which matters when an agent must act on current state (inventory, orders, fraud signals) rather than yesterday’s export.
Microsoft Fabric reported more than 40,000 paid customers, and Microsoft says more than 17,000 customers use Foundry and Fabric together, an example of the cloud platforms bundling data and model layers.

Will RAG stay essential as models improve? Our assessment: yes. Longer context windows reduce the need to chunk small document sets, but they do not solve permissions (a model should only see what the requesting user may see), freshness, citations, auditability, or cost (sending a large corpus on every request is expensive). Open-model reviews also report weaker reliability at very long contexts. The likely evolution is retrieval that is more agentic and better integrated with enterprise permission systems, not retrieval that disappears. Several analysts also note that RAG quality depends less on the vector database brand than on data cleanliness, chunking, hybrid keyword-plus-semantic search, and access control.

6. The development and operations layer (Layer 5)

Traditional software either works or fails visibly. AI systems fail quietly: a plausible wrong answer, a slow drift in quality when a model version changes, a cost spike from a looping agent. This layer exists to catch that.

Build: LangChain (applications and agents), LlamaIndex (data integration and retrieval), and Hugging Face (models, datasets, libraries) are the common open tooling. Each hyperscaler also ships its own agent SDK.
Evaluate and track: Weights & Biases (experiment tracking and evaluation), Arize AI (observability and evaluation), and MLflow (open-source model lifecycle management) let teams compare model and prompt versions against a fixed test set before shipping.
Monitor: Datadog and similar tools bring AI traces, latency, token cost, and errors into the same dashboards as the rest of the application estate.

Practical standard for production: every change to a model, prompt, retrieval index or tool gets regression-tested against a company-owned evaluation set; every request is traced; cost, quality and escalation rate are reviewed on a schedule. Teams that skip this layer typically discover problems through customer complaints. (Several vendors in this layer have been consolidating; confirm current ownership and independence when evaluating.)

7. The agent layer (Layer 6)

An assistant answers questions. An agent reads context, decides on steps, calls tools in business systems and completes work, sometimes across many applications. That shift changes pricing, risk and architecture.

Vendor	Role	Latest disclosed signal
Salesforce	Agentforce on CRM, data and Slack	Agentforce ARR passed $1B in May 2026 and was reported above $1.5B by September, though Salesforce widened its ARR definition that quarter.
ServiceNow	Agents on IT, HR and customer workflows; AI Control Tower for governance	Q3 2026: AI net new ACV up 65%; year-end AI ACV target raised to $1.7B+; agentic consumption in production up 13x year to date; goal of AI at 30% of ACV by 2030.
Microsoft	Copilot Studio, Foundry Agent Service, Agent 365 control plane	More than 50 million agents running in Agent 365 as of its October 28 call.
UiPath, Sierra, Glean, CrewAI	Agentic automation / RPA, customer experience agents, enterprise search and assistants, multi-agent frameworks	Not independently researched for this edition; see the quarterly tracker.

Salesforce reports ARR; ServiceNow reports ACV. The metrics are defined by each company and are not interchangeable.

Interoperability is settling around open protocols. Anthropic’s Model Context Protocol (MCP) standardizes how agents connect to tools and data; Google’s Agent2Agent (A2A) protocol standardizes how agents talk to each other. Both are now under Linux Foundation governance, and Microsoft has adopted A2A in Copilot Studio and Foundry. For buyers, this reduces integration lock-in and makes it easier to mix a vendor agent, a custom agent, and a cloud runtime.

Pricing is moving from seats to usage. Salesforce meters “agentic work units,” ServiceNow reports consumption in production, and Microsoft has described customer service as the first category with hybrid pricing, billing agent usage on top of per-seat licences. Budget for variable cost, not just licences.

Which processes are ready for autonomy? Our assessment, based on the deployments vendors and surveys describe:

Ready now with guardrails: tier-1 customer and employee support, IT service desk triage, document classification and extraction with sampling review, code generation with automated tests, internal research and summarization.
Autonomous only within narrow limits: routine procurement and invoice matching, order status changes, scheduling, standard HR queries — with spend, scope, and data limits enforced outside the model.
Keep human approval: payments and credit decisions, hiring and termination decisions, regulatory filings, medical or legal determinations, production system changes, anything irreversible or customer-visible at scale.

8. Security, identity and governance (Layer 7)

Agents hold credentials, read sensitive data and take actions. That makes them a new kind of identity, and a new attack surface: an agent can be steered by what it reads (prompt injection), can misuse a legitimate tool, or can inherit more access than any one task requires.

Standards are arriving. NIST’s Center for AI Standards and Innovation launched an AI Agent Standards Initiative on February 17, 2026, and NIST’s NCCoE published a concept paper on agent identity and authorization on February 5. The IETF has a draft framework for agent identity management. The OWASP Top 10 for Agentic Applications (December 2025) lists goal hijacking, tool misuse, and identity and privilege abuse among the leading risks. The NIST AI Risk Management Framework and the OWASP LLM Top 10 remain the baseline references.
Controls must sit outside the model. NIST guidance on agent identity argues that prompt-level guardrails alone are insufficient; identity, authorization, and runtime enforcement should be deterministic and independent of the model. Each agent should have its own identity, an owner, a purpose, and permissions granted per task.
Vendors are assembling the control layer. Okta offers Okta for AI Agents and its Cross App Access protocol for agent authentication, with endpoint discovery integrations (including CrowdStrike) announced for 2026. Microsoft ties Entra identity and Purview compliance to Agent 365. Palo Alto Networks markets Prisma AIRS for agent security; CrowdStrike covers endpoint and cloud telemetry; Wiz provides cloud risk visibility; Protect AI addresses model and supply-chain risk. ServiceNow has launched an AI Passport and Control Tower to verify agent scope, and has formed a Control Tower Alliance with Anthropic, AWS, Cisco, NVIDIA and Okta.

How readiness looks today. Vendor-sponsored surveys are alarming and should be read as indicative: one Gravitee survey reports 88% of organizations with confirmed or suspected agent security incidents in the past year and only 22% treating agents as identity-bearing entities in their own right. The practical gap is visibility: most companies cannot yet say how many agents they run, who owns them, what credentials they hold, or how to shut one down.

Minimum governance for production agents: an inventory of every agent and owner; per-agent identity with least-privilege, short-lived access; data permissions inherited from the requesting user; approved-action lists; human approval thresholds; full audit logs of prompts, retrievals and tool calls; a kill switch; and periodic access reviews.

9. Business applications (Layer 8) and where value is measurable

Most employees will meet AI through applications they already use. The Layer 8 question is not which assistant is smartest but which applications produce measurable results.

Productivity and collaboration: Microsoft 365 Copilot passed 30 million paid seats, up from about 20 million a quarter earlier, with customers buying more than 50,000 seats growing sharply. That is real scale, though paid seats measure purchase, not benefit.
Customer service and CRM: Salesforce, ServiceNow and specialists such as Sierra. Customer service is consistently the first function in production in surveys, and the first with usage-based pricing.
Software engineering: coding is the clearest early “killer app”; Menlo found Anthropic led enterprise code generation usage in 2025, and GitHub Copilot is widely deployed.
Finance, HR, supply chain, marketing and sales: SAP, Workday, Adobe, Intuit and others are embedding AI in systems of record. Adobe reported AI-influenced ARR of about $5B in September 2025, a vendor-defined metric.
Regulated industries (healthcare, banking, manufacturing, insurance): adoption is led by administrative and back-office use; clinical, credit, and safety-critical uses add validation, explainability, and audit requirements.

Test for any application: can the vendor show cost per completed task, accuracy against your data, escalation rate, and a baseline? If the benefit cannot be measured before and after, it is probably another assistant added to the software stack.

10. How companies will actually deploy AI

Four deployment models translate the layers into practical architecture. They are not mutually exclusive; a large company will run several at once.

Model	How it works	Best suited for	Main risk
A. Cloud-first	Existing cloud provider; managed models via API; internal data connected; identity and governance from the provider.	Companies starting enterprise AI without building infrastructure.	Platform lock-in; variable consumption cost.
B. Hybrid	Sensitive data in controlled environments; public-cloud models where appropriate; workloads routed by security, performance and cost.	Large enterprises, financial institutions, regulated organizations.	Integration complexity; inconsistent controls across environments.
C. Private / customized	Dedicated or reserved GPUs; open or customized models; fine-tuning where justified; control over serving.	AI-intensive firms with specific performance, privacy or cost needs.	Low utilization; scarce engineering talent; model upkeep.
D. Agent-driven operations	Agents connected to CRM, ERP and HR systems; defined permissions and approved actions; human approval for sensitive steps; monitored cost and accuracy.	Organizations moving from assistance to workflow automation.	Over-permissioned agents; agent sprawl; unmeasured cost.

The eight-step deployment lifecycle

Identify business problems. Start with measurable cases: customer support, document processing, software development, research, finance, internal knowledge retrieval. Record a baseline.
Select the model and platform. Compare proprietary and open models, cloud providers, inference cost, security needs, and integration options against your own test set.
Connect enterprise data. Integrate approved information through APIs, pipelines, retrieval systems, and knowledge bases, preserving source-system permissions.
Build the application or agent. Interface, business logic, integrations, and workflow execution.
Implement security and governance. Identity controls, permissions, audit trails, data protection, human oversight, usage policies.
Test and evaluate. Accuracy, hallucination rate, latency, reliability, security (including red-teaming and prompt-injection tests), and business performance.
Deploy and monitor. Usage, infrastructure cost, model performance, failures, and adoption.
Scale and optimize. Expand what works, move tasks to cheaper models where quality holds, automate adjacent workflows.

11. Real-world enterprise scenarios

These are reference patterns assembled from common architectures, not accounts of specific named deployments. Measurement targets should be set from each company’s own baseline.

Use case	Workflow	Integrations	Security needs	Success measures
Customer service	Agent identifies the customer, retrieves policy and account data, drafts or sends the answer, takes approved actions (refund under a limit, address change), escalates the rest with full context.	CRM, ticketing, knowledge base, order and billing systems, telephony.	Customer-data isolation, action limits, PII redaction, call and chat logging.	Resolution rate, human escalation rate, cost per interaction, satisfaction, repeat contacts.
Financial document processing	OCR and an LLM extract fields from invoices, statements, or contracts; rules validate totals and vendors; exceptions routed to staff; approved items posted.	ERP or accounting system, document storage, workflow or RPA tooling.	Segregation of duties, audit trail, retention rules, approval before posting.	Straight-through rate, extraction accuracy, processing time, cost per document, error rate.
Enterprise knowledge assistant	Employee asks a question; retrieval fetches only documents that the employee may access; model answers with citations.	SharePoint or Drive, wikis, ticketing, chat, vector or search index.	Permission-aware retrieval, identity federation, stale-content handling.	Answer accuracy on a test set, time to find information, active usage, deflected support requests.
Software development	Assistants and coding agents propose changes, write tests and open pull requests; CI runs checks; humans review and merge.	Repositories, CI/CD, issue tracker, observability.	Secrets protection, code-leak controls, license scanning, restricted production access.	Cycle time, accepted-change rate, defect and rollback rates, review load, cost per merged change.
Supply chain	Predictive models forecast demand and disruption; an agent drafts reorder or rerouting options; planners approve; ERP updated.	ERP, warehouse and logistics systems, supplier feeds, streaming data.	Spend and commitment limits, supplier-data protection, human approval for commitments.	Forecast error, stockouts, inventory carrying cost, expedite cost, planner hours.
HR operations	Employee asks about leave, benefits or payroll; assistant answers from policy and the person’s own records; routine requests processed; sensitive cases handed to staff.	HRIS, payroll, service-management platform, policy library.	Strict access to personal data, regional privacy rules, no automated employment decisions without human review.	Tickets resolved without staff, time to resolution, employee satisfaction, HR hours saved.

12. The economics of enterprise AI

Model price is one line in a much larger budget. Total cost of ownership (TCO) includes everything required to make a model produce a reliable business result:

Cost category	What drives it
GPU and cloud infrastructure	Reserved vs on-demand capacity; utilization; networking and egress
Model inference and API consumption	Tokens in and out; reasoning-heavy calls; retries; agent loops
Data storage and processing	Pipelines, cleaning, embedding, refresh frequency
Vector databases and retrieval	Index size, query volume, hybrid search
Software licensing and subscriptions	Seats plus usage-based agent fees
AI engineering and integration	Connecting to ERP, CRM and legacy systems; often the largest line
Security and compliance	Identity, logging, red-teaming, audits, regulated-industry validation
Monitoring and maintenance	Evaluation reruns, model migrations, prompt and index upkeep
Training and change management	Workflow redesign, user training, adoption support

The cheapest model is not necessarily the cheapest outcome. Consider an illustrative example (hypothetical numbers, for explanation only). Model X costs half as much per call as Model Y. But X completes only 70% of tasks without help, needs an average of 1.4 attempts, and sends 30% of cases to a staff member costing far more than the model call. Y completes 92%, rarely retries, and escalates 8%. Once retries and human time are included, Y can cost less per completed task. The only way to know is to measure.

Cost per completed task = (inference + retrieval + orchestration + monitoring + human review time + rework from errors) ÷ tasks completed correctly.

KPI	Why it matters
Cost per completed AI task	The true unit cost, including failures and rework
Cost per customer interaction	Comparable to the pre-AI baseline in service operations
Accuracy and task completion rate	Quality gate for expanding autonomy
Time saved per employee	Productivity; must be converted to cost or revenue to count
Human escalation rate	Falling rates signal maturing automation; rising rates signal drift
Infrastructure utilization	Decides whether owned or reserved GPUs pay off
Revenue or operating-cost impact	The business outcome
Return on AI investment	Fully loaded benefit divided by fully loaded cost

Treat published ROI claims with care. Many “median ROI” and productivity statistics circulating in 2026 come from vendor-sponsored or small-sample surveys with differing definitions. One independent-style data point worth noting: Forrester has reported that 44% of AI projects that reach production achieve positive ROI within 12 months, as relayed by secondary compilations. Attribute vendor claims and verify them against your own baseline.

13. The future enterprise AI architecture: ten strategic questions

1. Hyperscalers or specialized AI infrastructure? Both. Hyperscalers will remain the default control plane because they own identity, contracts and data gravity. Neoclouds win on large reserved GPU blocks and price; hyperscalers themselves buy from them. Expect the boundary to blur through partnerships and acquisitions.

2. One model provider or many? Many. Share data shows rapid movement among providers and no stable monopoly, and routing across models is becoming a core platform capability. Standardize on the interface, evaluation set, and governance, not on one model.

3. How important will open models be? Important for cost-sensitive, high-volume, or private workloads, and as negotiating leverage; less dominant in revenue terms. Count usage and spend separately, because the two diverge sharply.

4. Will RAG remain essential? Yes, for permissions, freshness, citations, and cost, though the mechanics will become more agentic.

5. Which processes are ready for autonomous agents? Bounded, reversible, well-instrumented work: support triage, IT requests, document extraction, coding with tests. Keep human approval for money movement, employment, and legal or regulatory decisions (see Section 7).

6. How will agents be secured? As a new identity class: inventory, per-agent identity, least privilege, tool-call policy enforcement outside the model, full audit logs and kill switches. Standards work at NIST, IETF, and OWASP is converging on this.

7. Which layers capture the most value? Today, by reported dollars, infrastructure (NVIDIA’s data center revenue alone is about $89B a quarter). Over time, we expect value to broaden toward the layers closest to workflows and data: platforms that hold governed data (Databricks, Snowflake, the hyperscalers) and applications that own the process (Microsoft, Salesforce, ServiceNow, SAP and others). Models risk commoditizing on routine tasks while premium models keep pricing power at the frontier. This is analysis, not a forecast of any company’s results.

8. New software stack or extension of existing platforms? Mostly an extension. The large incumbents are bundling agents, data and governance into platforms customers already own, and Microsoft and ServiceNow are both selling the control plane for agents. Investors and analysts are debating whether agents will weaken seat-based software; the vendors’ consumption disclosures suggest pricing, not necessarily demand, is what is changing.

9. How should CIOs and CTOs prioritize? (a) Fix data access and permissions first; (b) establish identity and governance for agents before scaling them; (c) build a company-owned evaluation set and cost-per-task measurement; (d) keep model and platform choices portable; (e) concentrate on a few high-value workflows rather than many pilots; (f) negotiate for usage-based pricing transparency; (g) invest in change management.

10. What separates production from expensive experiments? Production deployments have a named business owner, a measured baseline, governed data, permission-aware retrieval, an evaluation and monitoring loop, defined human-approval points, and a cost model that includes integration and operations. Experiments typically have a model demo and an enthusiastic sponsor.

Bottom line. The winning enterprise AI architecture will be assembled, not bought as a single product. Models are essential but interchangeable; the durable advantage sits in clean, governed data, secure identity for agents, disciplined evaluation, and integration into the workflows where money is made or saved.

Scope notes and limits of this edition

Company financial figures come from reported results and press coverage as of November 13, 2026. Run-rate, ARR and ACV metrics are company-defined and unaudited; Databricks is private.
Several statistics (adoption rates, model shares, open-model share, multi-model counts, security incidents) come from surveys with different samples and definitions. Where sources disagree, we show the range.
Oracle, IBM, Cohere, Mistral, Pinecone, Redis, Elastic, MongoDB, Confluent, LlamaIndex, CrewAI, Sierra, UiPath, Glean, CrowdStrike, Wiz, SAP, Workday and Intuit are described by role. Fresh financial or adoption data for them was not gathered for this edition. Ownership of some vendors in Layers 5 and 7 has changed through acquisitions; confirm current status before citing.
Anthropic, whose models are discussed in this report, is also the maker of the assistant that helped research and draft it.

Related Intelligence

AI Infrastructure Intelligence · Cloud & Infrastructure Intelligence · Developer Technology Intelligence · Cybersecurity Intelligence · Industry Intelligence

Related Coverage

Company Deep Dive · November 7, 2026 ASML: Why Advanced AI Chips Depend on One Critical Manufacturing Layer

Company Analysis · November 7, 2026 ASML: The Company Behind the World’s Most Advanced Chips

Company Deep Dive · November 8, 2026 TSMC: Can Its Foundry Dominance Survive the AI Semiconductor Boom?

The Term Sheet · November 7, 2026 The Venture Capital Race for Enterprise AI: Where Are Investors Placing Their Bets?

Startup Spotlight · November 7, 2026 Tenstorrent: The Open Challenge to AI Chip Dominance

Hardware Watch · November 5, 2026 SambaNova and Supermicro target 10,000 SN50 chips — heterogeneous inference and the distribution race

Sources

Primary documentation to consult directly when evaluating vendors: AWS Bedrock, Microsoft Foundry, and Google Vertex AI documentation; OpenAI, Anthropic, and Hugging Face model documentation; Databricks and Snowflake platform documentation; LangChain and CrewAI; NIST AI Risk Management Framework and the OWASP LLM and Agentic Top 10; NVIDIA data center and CoreWeave documentation. Vendor productivity and cost-saving claims in this report are attributed and should be independently validated.

ABOUT THE AUTHOR

Erwin Castro

Founder, Publisher & SEO Writer at The CODEW

Erwin Castro is the founder and publisher of The CODEW, an independently operated technology and business intelligence publication covering Tech M&A, AI, enterprise software, SaaS, cloud infrastructure, startups, business operations, and digital strategy.

About Author · LinkedIn · X

The CODEW