← Knowledge Base 🕐 14 min read
Knowledge Base

AI Platform Selection

· March 2026

The decision is harder than standard software procurement because the underlying models, the competitive landscape, a...

AI platform selection is the organizational decision about which AI tools, vendors, and ecosystems to deploy — and at what commitment depth. It encompasses the platform vs. point-solution trade-off, the build vs. buy question, and the specific procurement discipline required when evaluating vendors without in-house analyst resources or a large IT team.

The decision is harder than standard software procurement because the underlying models, the competitive landscape, and the pricing structures are all changing faster than annual contracts can accommodate. A platform bet that looks rational in Q1 can look premature by Q3, as Anthropic’s growth from 12% to 40% of enterprise API share — and OpenAI’s fall from 50% to 27% — in two years illustrates (Menlo Ventures, ~500 U.S. enterprise decision-makers, December 2025).

Why this matters for mid-market buyers

  • Platform lock-in is deeper than license lock-in. Once AI workflows are woven into Microsoft Graph or Google’s APIs — reading email, calendars, documents, retained context — migration costs exceed anything companies have faced with traditional SaaS. One documented case: migrating 40 AI workflows cost $315,000 and three months of service degradation (Swfte AI, 2026).
  • User preference and enterprise procurement are misaligned. When employees have access to both Microsoft Copilot and ChatGPT, 76% choose ChatGPT. Copilot’s conversion rate among employees with paid access is 35.8%, versus ChatGPT’s 83.1% (Recon Analytics, n=150,000+, January 2026). Buying the platform tool does not guarantee the platform tool gets used.
  • The cheapest first move is activating what is already installed. Platform-native AI (M365 Copilot, Google Gemini in Workspace) has near-zero integration cost. The right question is not whether it is the best tool — it is whether its quality is good enough that employees will actually use it, at the price being charged.

Platform vs. Point Solution: Consolidate or Diversify? (March 2026)

  • 28% of enterprises run 10+ AI apps; 76% have experienced negative outcomes from disconnected tools (Zapier/Centiment, n=550 C-suite, October 2025).
  • The recommended mid-market architecture: one primary ecosystem for productivity AI (activate the platform already in place), one or two best-of-breed tools for specific high-value workflows on monthly billing, and standalone subscriptions (ChatGPT, Claude) for general-purpose AI where switching cost is near zero.
  • Google Workspace with Gemini included: approximately $14/user/month. Microsoft 365 Business Standard + Copilot: $42.50/user/month. The $171,000 annual gap at 500 employees compounds over a commitment term.
  • The hidden cost multiplier: integration, training, governance, and maintenance run 2–3x the license fee regardless of approach; 78% of IT leaders report unexpected charges from AI pricing models (Zylo, 40M+ licenses analyzed, 2026).
  • 90-day review cadence: cut tools with under 25% adoption; evaluate platform feature releases against point-solution performance before renewing.

Source: research/07-adoption-challenges/ai-platform-vs-point-solution-vendor-stack.md

The AI Platform Commitment Decision: Microsoft, Google, or Independent (March 2026)

  • 70% of Fortune 500 companies have Copilot licenses; only 3.3% of the 450 million M365 commercial seats are paid Copilot seats. The largest installed base is not the most-used platform.
  • The independent, model-agnostic path is gaining share: 37% of enterprises now deploy five or more models in production, up from 29% the prior year (a16z, n=100 CIOs, May 2025). Abstraction layers (API gateways routing to the best model per task) are becoming a production standard.
  • Total cost of ownership runs 3–5x the advertised subscription price when accounting for integration, training, workflow redesign, and option cost of being unable to switch.
  • Google Gemini is included in all Workspace Business and Enterprise plans since January 2025; Microsoft’s E7 bundle at $99/user/month includes Copilot, identity management, and agent tools.
  • Lock-in test: keep AI workflows that touch the company’s core data on the platform (where integration value is highest); keep general-purpose AI access on standalone tools (where switching costs are lowest).

Source: research/07-adoption-challenges/ai-platform-commitment-decision.md

Model Provider Market Share: The 2026 Reality (February 2026)

  • OpenAI holds 78% production penetration across Global 2000 enterprises but only ~56% of wallet share — usage is broad but spend is concentrating in alternatives (a16z 3rd Annual CIO Survey, n=100 Global 2000 executives, Feb 2026)
  • Anthropic added 25 percentage points of enterprise production usage in nine months (44% now vs. ~19% in May 2025) — fastest share gain of any frontier lab; cross-validated by Ramp AI Index (Anthropic 34.4% vs. OpenAI 32.3% business adoption, n=50,000+ U.S. businesses, April 2026)
  • 81% of enterprises use three or more model families in production — multi-model is the new enterprise default, not a specialized one
  • CIO projections for end-2026: OpenAI 53%, Anthropic 18%, Google 18% — the model is commoditizing faster than vendor contracts anticipated
  • Average enterprise LLM spend: $4.5M (2024) → $7M (early 2026) → $11.6M projected (2026 year-end); app spend consistently over-runs budget by ~54% ($3.9M budgeted → $6M actual)
  • 65% prefer incumbent solutions for integration and procurement simplicity — the AI ceiling is set by the platform chosen, not the underlying model family
  • Caveat: a16z is an Anthropic investor; Anthropic share data should be calibrated against independent transaction-verified sources

Source: research/01-ai-native-landscape/a16z-enterprise-cio-third-annual-2026.md

AI Tool Evaluation and Selection Framework (March 2026)

  • Buy beats build for mid-market AI: purchased solutions convert pilot-to-production at 47% — nearly double typical SaaS; custom builds fail at mid-market scale primarily due to data preparation burden (61% of build timelines) and 34% annual ML engineer turnover.
  • Companies that define success metrics before vendor selection achieve 54% project success versus 12% without (Pertama Partners, 2,400+ AI initiatives, 2025–2026).
  • A Gartner subscription runs $30,000–$100,000/year. Mid-market companies can replicate 80% of the evaluation rigor using G2, Gartner Peer Insights, and TrustRadius with structured pilot protocols.
  • The average failed AI project costs $4.2M and takes 11 months to abandon.
  • Platform-native AI (M365 Copilot, Google Gemini, Salesforce Agentforce) is the path of least resistance — but lock-in risk and per-seat cost premiums make the decision non-obvious.

Source: research/07-adoption-challenges/ai-tool-evaluation-selection-framework.md

AI Vendor Evaluation for Non-Technical Buyers (March 2026)

  • Business units control 81% of SaaS spend; IT directly manages 15%. The AI buyer is no longer the CIO — it is the Marketing Director, VP of Sales, General Counsel, and Controller (Zylo 2026 SaaS Management Index, 40M+ licenses analyzed).
  • 27% of enterprise AI application spend enters through product-led growth channels; shadow adoption pushes the figure toward 40%. Tools “land” through individual users before any formal procurement process begins.
  • Five department-led failure modes: demo syndrome purchase (tool works on vendor’s clean data, not the company’s messy CRM), free-tier-to-enterprise drift, duplicate purchase of tools already available in existing licenses, vendor lock-in surprise, and compliance blind spot (AI screening tools triggering state employment AI laws without audit).
  • The 10-question framework for non-technical buyers: define the specific workflow change and measurement before the demo; demand a pilot with the company’s actual data; verify that the vendor’s security posture matches the sensitivity of data being processed.

Source: research/07-adoption-challenges/ai-vendor-evaluation-non-technical-buyers.md

How to Pick Your First AI Tool (March 2026)

  • Only 6% of firms capture >5% EBIT impact from AI; the single largest predictor of failure is selecting a tool before defining the problem it solves.
  • Three pre-procurement questions: (1) What specific workflow costs time, money, or errors today? (2) What does success look like in 60 days, quantified? (3) What happens if this fails — is failure cheap and reversible?
  • Proof of concept protocol: demand a 2–4 week trial with the company’s actual data, not vendor showcase data. Test against 10–20 real cases before signing an annual contract.
  • 42% of companies abandoned most AI initiatives before production in 2025, up from 17% in 2024 (S&P Global Voice of the Enterprise, n=1,006, 2025).

Source: research/07-adoption-challenges/how-to-pick-your-first-ai-tool.md

Practitioner voices (pillar 13)

“Start with the simplest, most painful problem you can find. And the simplest, most obvious solution to that. Don’t start complicated. Start with the simplest. Don’t overcomplicate the stack. You’re not going to have to pre-train models at the very first LLM launch you’re about to do.” — Pranav Pathak, Product AI Development Lead, Booking.com (Dec 2025) Source: research/13-multimodal-sources/beyond-the-pilot/2025-12-03-how-bookingcom-boosted-agent-accuracy-2x-with-mini-llms-with.md

“For Zoom, our customers are very selective. So Zoom AI Companion offers two configurations. One is a federated AI that offers the best quality. We combine our own small language model with the best models in the industry, either OpenAI or Anthropic. […] Another configuration is, a lot of companies, Zoom customers, are worried about this federation […] they demanded a Zoom-only model.” — Xuedong Huang, Chief Technology Officer, Zoom (Apr 2026) Source: research/13-multimodal-sources/beyond-the-pilot/2026-04-13-open-closed-or-hybrid-choosing-the-right-model-strategy-for-.md

“The moats for developer tools, some of the OEMs, may not be that weak, because what I see is if there’s a better OEM, we switch API calls on a dime, so the API moats is weaker.” — Swami Sivasubramanian, Vice President of Agentic AI, Amazon Web Services (Apr 2026) Source: research/13-multimodal-sources/snowflake-summit/2026-04-14-the-ai-blueprint-for-the-next-decade-build-2025-luminary-con.md

“I think right now, most of this is initially vibes, right? Like you go in there and you play around with it and you’re like, is this close enough to the frontier that we want to put some effort in here? And usually the answer is no. The ones that we’ve tested — Gemini, DeepSeek, and Google, plus obviously OpenAI — are the ones where, okay, actually this is pretty close to the frontier, so it’s worth doing.” — Andrew Lee, CEO, Tasklet (May 2026) Source: research/13-multimodal-sources/cognitive-revolution/2026-05-18-three-kinds-of-software-survive-tasklets-andrew-lee-on-compe.md

Lee’s framing captures how AI-native product companies actually run model evaluations: rapid qualitative assessment at the frontier boundary (“close enough to frontier?”), not benchmark scores. This is the practitioner heuristic behind the a16z finding that 81% of enterprises use three or more model families — the evaluation overhead is low enough that multi-model is now the default starting position.

Menlo Ventures State of GenAI in the Enterprise 2025

n=495 U.S. enterprise decision-makers, November 7–25, 2025 (Tier 1 — MEDIUM-HIGH credibility; Menlo is an investor in Anthropic).

  • 76% of AI use cases are now purchased rather than built (up from 53% in 2024). The build-vs.-buy question has a clearer answer than at any prior point.
  • Startups captured 63% of AI application revenue in 2025, reversing incumbents’ 64% advantage from 2024. In finance and operations specifically, startups hold 91% of application revenue — meaning the incumbent ERP/CRM vendor’s AI roadmap is not where enterprise spend is going.
  • 47% of AI deals reach production — nearly double traditional SaaS. Once a purchase decision is made, deployment rates are high. The risk has shifted from “will we ever deploy this” to “will we capture value after we deploy.”
  • Open-source LLM enterprise adoption fell from 19% to 11% in one year. The total-cost-of-ownership argument for open-source has not been winning against the support and compliance simplicity of commercial APIs.
  • Shadow AI is underestimated: ~40% of AI application spend may be entering through product-led growth channels (formal PLG at 27%, personal credit card/shadow at an estimated additional 13%).

Source: research/01-ai-native-landscape/menlo-ventures-state-genai-enterprise-2025.md

Supporting research

What this means for mid-market buyers

  • Define the specific workflow problem and 60-day success metric before any vendor demos. Organizations that skip this step fail AI initiatives at 88% rates.
  • Run a proof of concept with the company’s actual data before signing an annual contract. Every vendor demo uses clean data. The test that matters is whether the tool works with the company’s messy CRM, not the vendor’s showcase account.
  • Maintain flexibility by keeping general-purpose AI (research, drafting, analysis) on monthly standalone subscriptions where switching costs are near zero. Lock into the platform layer only where integration with core company data creates genuine integration value.

See also

  • AI Vendor Contracts — lock-in clauses, portability terms, and exit provisions
  • Inference Economics — cost structure of AI model consumption
  • Data Readiness — data quality prerequisites that determine whether a vendor can deliver on demo promises
  • Shadow AI — what happens when department-led AI purchasing bypasses evaluation discipline
  • AI Model Evaluation and Benchmarks — why vendor benchmark scores are unreliable as procurement inputs and how to run a domain-specific internal evaluation

Consulting Firm AI Positioning Landscape (Brandon Sneider Synthesis, March 2026)

Source: research/04-consulting-firms/consulting-firm-landscape.md

  • The global AI consulting market reached $11.07B in 2025, projected to $90.99B by 2035 (26.2% CAGR). Every Big 6 firm has positioned AI transformation as its primary revenue line — meaning buyers face sellers whose income depends on AI urgency.
  • McKinsey (QuantumBlack), BCG (BCG X), Accenture, Deloitte, PwC, EY, and IBM IBV all publish primary AI research as a demand-generation activity. Commercial interest should be held in mind when evaluating their adoption statistics and ROI claims.
  • The implication for mid-market platform selection: consulting firm recommendations reflect their implementation capability, not necessarily what delivers the best ROI for a 200–2,000 person company. Large firms route clients to platforms they have certified practices for; independent evaluation against your specific workflow stack is required.
  • In 2026, foundation model providers entered the services market directly: Anthropic/consulting JV structures and OpenAI’s Deployco venture represent a new conflict of interest — the same organization selling both the model and the implementation advisory. Mid-market buyers should require explicit disclosure of model-vendor relationships from any consulting firm recommending a specific foundation model.

Credibility: MEDIUM — all named firms have commercial interest in AI adoption urgency. Financial disclosures HIGH; survey data MEDIUM. TIER 1/2 split. See also: research/04-consulting-firms/anthropic-enterprise-ai-services-jv-2026.md · research/04-consulting-firms/openai-deployco-consulting-venture-2026.md

Forrester AI Model Openness Framework (April 2026)

Source: research/16-procurement-contracting/forrester-ai-model-openness-framework-2026.md

Forrester’s MOF provides a structured procurement framework for evaluating open-weight AI model claims across three dimensions: Reproducibility, Usage Rights, and Community Momentum.

  • “Open” is a marketing claim, not a legal standard. Llama, pre-2026 Gemma, and DeepSeek all fail the OSI open source definition — they are source-available with commercial restrictions. Approved-vendor lists that only check Apache 2.0 / MIT / BSD will accept non-qualifying models.
  • Usage Rights is where enterprise deals break down. Field-of-use restrictions, MAU caps, and unilateral amendment rights disqualify models from regulated production use even when weights are publicly available. Verify current terms at point of procurement — Meta’s Llama license has been amended without notice.
  • Community Momentum flags long-term operational risk. A model with a single corporate sponsor, no active open governance, or amendable license terms is an operational risk beyond the initial legal review. Google (Gemma 4) and Alibaba (Qwen3+) moved to genuine Apache 2.0 in 2025–2026; Meta and DeepSeek did not.
  • Reproducibility maps directly to regulatory audit requirements: EU AI Act Article 13 (high-risk system documentation), NIST AI RMF transparency requirements, and SOC 2 vendor assessments all require documentation that only highly reproducible models provide.

Credibility: MEDIUM-HIGH for framework structure and regulatory mapping; MEDIUM for specific model scores — verify at point of procurement.

a16z — How 100 Enterprise CIOs Are Building and Buying Gen AI (2025)

Source: research/01-ai-native-landscape/a16z-enterprise-cio-build-buy-ai-2025.md

a16z’s 2025 CIO survey (n=100, 15 industries, 25+ buyer interviews) on how enterprise buyers are structuring build vs. buy decisions for gen AI.

  • Note: a16z has portfolio financial interest in AI-native vendors; apply VC bias caveat. Findings are directionally consistent with Menlo Ventures 2025 and independent build/buy data but should not be treated as independent research. TIER 2 (May 2025 fieldwork).
  • Use for: characterizing enterprise CIO decision frameworks and vendor evaluation criteria — with appropriate sourcing caveat.

Credibility: MEDIUM (VC commercial interest; n=100 adequate; qualitative CIO interviews supplement quantitative survey).

AI PCs and Desktop-First Architecture: Hardware Now Precedes Strategy (May 2026)

Source: research/20-local-tiny-models/ai-pc-desktop-first-enterprise-strategy-2026.md · Gartner (Sep 2024, Aug 2025), Forrester Predictions 2026, Six Colors Enterprise Report Card (n≈100 IT admins, May 2026) · TIER 1/TIER 2 · MEDIUM-HIGH

  • Platform selection is increasingly constrained by hardware defaults. Gartner projects AI laptops will be the only commercial laptop available to large businesses by end of 2026. Enterprises choosing a cloud-only AI platform in 2025–2026 are leaving on-device inference capacity unused across their entire endpoint fleet.
  • Only 15% of enterprises have an explicit private-AI strategy (Forrester Predictions 2026). The gap between 55% AI PC penetration (Gartner) and 15% deliberate strategy is the governance and platform selection vacuum.
  • Regulated industries have already made the platform decision by compliance logic. HIPAA-covered entities, FINRA-regulated firms, and defense contractors route sensitive data to local inference — eliminating BAA requirements and reducing regulatory surface area. Platform selection at these firms is driven by the compliance team as much as IT.
  • The practical implication for platform selection: enterprise AI platform evaluations in 2026 should include a local/hybrid inference tier alongside cloud APIs. Microsoft Copilot+ PC, Apple Intelligence, and developer-driven Ollama/LM Studio deployment are already happening at the endpoint; the question is whether IT governs the inference tier or inherits it.
  • Model quality gap has closed. DeepSeek V3, Qwen3, Llama 3.3, and Mistral-class models match GPT-4-level performance on most enterprise use cases. The technical argument for mandatory cloud API consumption has materially weakened since 2024, removing one constraint on private/local platform choices.