← Findings 🕐 18 min read
Findings

The AI Vendor Pitch Decoder: Red Flags, Green Flags, and the Jargon They Hope You Won't Question

Brandon Sneider · March 2026

Every enterprise software vendor now claims AI capabilities. Most pitches follow the same pattern: impressive demo, industry jargon, urgency to sign.

How to Use This Card

Every enterprise software vendor now claims AI capabilities. Most pitches follow the same pattern: impressive demo, industry jargon, urgency to sign. This two-part reference card helps executives evaluate those pitches in real time. The front side identifies what to watch for. The back side translates the terminology vendors assume you already understand.

Executive Summary

  • 45% of marketing technology leaders say vendor-offered AI agents fail to meet promised business performance. The gap between the demo and the deployment is not a bug — it is the dominant pattern (Gartner, n=413 martech leaders, June-August 2025).
  • Only ~130 of the thousands of vendors claiming “agentic AI” capabilities have genuine agentic functionality. Gartner calls this “agent washing” — rebranding chatbots, RPA scripts, and basic automation as AI agents (Gartner, January 2025 poll of 3,412 attendees).
  • The SEC has levied enforcement actions against companies misrepresenting AI capabilities since March 2024 — including Delphia ($225K penalty), Global Predictions ($175K penalty), Presto Automation (cease-and-desist), and Nate Inc. (founder charged with fraud after raising $42M on fabricated AI claims). AI washing is now a regulatory category, not just a marketing problem (SEC, 2024-2025).
  • 78% of IT leaders reported unexpected charges tied to consumption-based or AI pricing models in the last 12 months. The price on the slide deck and the price on the invoice are reliably different (Zylo 2026 SaaS Management Index, 40M+ licenses analyzed).

Part 1: Five Red Flags in an AI Vendor Pitch

Red Flag #1: “Our AI is fully autonomous / fully automated”

What you hear: “Our solution automates the entire workflow end-to-end with no human intervention.”

What it means: The SEC charged Presto Automation in January 2025 for claiming its drive-thru voice assistant was AI-automated. It was operated by a third party with significant human intervention. Nate Inc.'s founder was indicted after raising $42M on claims of AI-powered shopping — transactions were processed manually by contract workers overseas. Securities class actions targeting AI misrepresentations doubled between 2023 and 2024 (SEC enforcement data, 2024-2025).

What to ask: “What percentage of transactions go through without human intervention? Can I see the audit log?”

Red Flag #2: Vague security language without specifics

What you hear: “We use industry-leading security practices” or “advanced encryption” or “trust us, we work with Fortune 500 companies.”

What it means: A vendor that cannot name their encryption standard (AES-256), their compliance certifications (SOC 2 Type II, ISO 27001), or explain where your data physically resides is either hiding something or does not understand their own architecture. 88% of AI vendors cap liability at one month’s subscription fee — if their AI causes a breach, your maximum recovery is $3,000-$5,000 per seat while you face uncapped regulatory penalties and client claims (CIO, Deshpande et al., October 2025).

What to ask: “Show me your SOC 2 Type II report, your sub-processor list, and your data residency map. If you can’t produce those today, schedule a follow-up before the next meeting.”

Red Flag #3: The demo runs on the vendor’s data, not yours

What you hear: “Let me show you what this looks like” — followed by a polished demonstration on a curated dataset.

What it means: This is the single most reliable red flag across every major AI evaluation framework (DUNNIXER, CloudEagle.ai, Dan Cumberland Labs, 2025). Your data is messy, incomplete, and idiosyncratic. Curated demos hide that reality. Organizations without well-defined use cases experience failed AI projects 70% of the time (CloudEagle.ai, 2025). 60% of AI project failures stem from inadequate data — the vendor knows this and shows you clean data for a reason (Gartner, 2025).

What to ask: “Run this on our data. If you need a week to prepare, fine. But I will not evaluate a tool I have only seen work on your data.”

Red Flag #4: No mid-market references or “our customers include [Fortune 500 logos]”

What you hear: The vendor’s case study page features Walmart, JPMorgan, and Deloitte. Your company has 400 employees.

What it means: A Fortune 500 deployment with a dedicated IT team, a $2M implementation budget, and 18 months of integration work tells you nothing about what this tool does at a 400-person company with a CIO wearing two hats. A vendor that cannot produce three references at your company size does not have a mid-market product — it has a mid-market price list. All references under six months old suggest no long-term track record (DUNNIXER enterprise evaluation framework, 2025).

What to ask: “Connect me with three customers between 200 and 1,000 employees who have been live for at least six months. I’ll ask them what broke that you didn’t warn them about.”

Red Flag #5: Pressure to sign before you can evaluate

What you hear: “This pricing expires at end of quarter” or “We only have three onboarding slots left this month.”

What it means: Artificial urgency is a sales tactic, not a supply constraint. The median abandoned AI project consumed 11 months and $4.2M before termination (Pertama Partners, n=2,400+ initiatives, 2025-2026). A 30-day evaluation before signing prevents a 12-month recovery after signing. Every vendor who is confident in their product will give you time to run a structured pilot.

What to ask: “I will not sign without a 30-day paid pilot with 5-10 users on a real workflow. If your calendar doesn’t accommodate that, we’re not a fit.”

Part 2: Five Green Flags That Signal a Serious Vendor

Green Flag What It Looks Like
Volunteers limitations “Here is where the AI gets it wrong, and here is the user experience when that happens.” Every AI tool has failure modes. The honest vendor names them before you discover them.
Demos on your data Offers a proof-of-concept on your actual data (anonymized if needed) within 1-2 weeks. Brings deployment architecture and data flow diagrams to the second meeting.
References at your scale Provides direct contact with 3+ customers between 200 and 2,000 employees who have been live for 6+ months. Encourages you to ask hard questions.
Transparent total cost Presents the full year-1 cost: license + implementation + training + support + consumption overages + renewal terms. Does not hide the 60-75% of cost that sits above the license fee.
Recommends starting small Suggests a 30-day pilot on one specific workflow with pre-defined success metrics and kill criteria. Asks substantive questions about your business before proposing a solution.

Part 3: Three Questions Before Activating AI Features in Tools You Already Pay For

Your existing vendors — Microsoft, Google, Salesforce, Zoom, ServiceNow — have embedded AI features into products you already own. Most are enabled by default. These questions apply before turning on (or discovering you cannot turn off) AI in your current stack.

Question 1: “What data does this AI feature process, and has the data processing agreement been updated?”

Microsoft added Anthropic as a sub-processor for Microsoft 365 Copilot by default for most commercial tenants as of January 7, 2026 — without requiring explicit consent. Google deployed Gemini to Workspace subscribers between January and March 2026, enabled by default. Business-tier subscribers had to request access to admin controls — meaning AI was activated before the governance mechanism was available (2toLead, December 2025; Google Workspace Updates, 2025-2026).

Action: Check your admin console for AI toggles you did not configure. Compare your current DPA against the vendor’s posted version. If AI model providers (OpenAI, Anthropic, Google) have been added as sub-processors since your last review, your data flow has changed without your approval.

Question 2: “Does the cost justify activating it for all seats — or should it start with 10-20?”

M365 Copilot adds $30/user/month on top of existing licenses. For a 300-person company, that is $108,000/year. Google Workspace Business Standard includes Gemini at $14/user/month. Salesforce Agentforce starts at $2/conversation. 62% of AI deployments achieve less than 40% user adoption in the first six months (Pertama Partners, 2025-2026). Activating enterprise-wide is paying for 300 seats when 40 people will use it.

Action: Start with 10-20 power users on a specific workflow. Measure adoption at day 30. Expand only if adoption exceeds 60% and the workflow metric improves.

Question 3: “Can individual attendees opt out — and does it matter for our client obligations?”

Zoom’s AI Companion processes every participant’s speech when the host enables it. Individual attendees cannot opt out. The notification during meetings is a declaration, not a consent request. For companies with client confidentiality obligations — legal, financial services, healthcare, consulting — this creates risk that no technology configuration resolves (TechCrunch, 2025; Bloomberg analysis, 2025).

Action: Review client engagement letters and NDAs for data processing restrictions. If any client agreement prohibits AI processing of communications, disable AI meeting features for those client interactions or obtain written consent.

Part 4: The Jargon Guide — 25 Terms in Plain English

Term What the Vendor Says What It Actually Means
Agentic AI “Our agents work autonomously to complete tasks” AI that takes multi-step actions — plans, uses tools, checks results, retries. Only ~130 of thousands of vendors claiming this capability have it. Most are chatbots with a new label (Gartner, January 2025).
AI Agent “Deploy agents across your workflows” A program that receives a goal and decides its own steps to achieve it, including calling other tools. The word “agent” is the most inflated term in enterprise AI — verify what it actually does by asking for a workflow log.
Foundation Model “Built on a state-of-the-art foundation model” A large AI model (GPT-4, Claude, Gemini, Llama) trained on broad data that other products build on top of. The vendor likely did not build it. They built the interface. Ask: “Which foundation model? Whose API?”
Large Language Model (LLM) “Powered by our proprietary LLM” An AI trained on text that generates text. Very few vendors have a proprietary LLM. Most use OpenAI, Anthropic, or Google APIs. “Proprietary” sometimes means “fine-tuned someone else’s model.” Ask which base model.
RAG (Retrieval-Augmented Generation) “Our RAG architecture ensures accurate answers” Instead of relying on the AI’s training data alone, the system searches your documents first, then generates an answer based on what it finds. Reduces hallucination. Does not eliminate it.
Fine-Tuning “We fine-tuned the model on your industry” Training an existing AI model on a specific dataset to improve performance for a narrow use case. Sounds impressive. Effective fine-tuning requires large, clean datasets. Ask: how much data, from where, and what measurable improvement over the base model?
Hallucination Vendors rarely say this word voluntarily AI generating false information presented as fact. Every LLM does this. The rate varies from 3% to 27% depending on the task. Ask: “What is your hallucination rate on our type of data, and how does the system flag uncertain outputs?”
Tokens “Usage-based pricing per token” The unit AI models process — roughly ¾ of a word. A page of text is ~500 tokens. This is how consumption-based pricing works. A 100-page document costs 50,000 tokens to process. Ask for a monthly token cost estimate based on your actual volume.
Embedding “We create embeddings of your data” Converting text into numerical representations that capture meaning, so the AI can search by concept rather than keyword. This is what makes “smart search” work. Your data is converted but not visible to other customers (in properly designed systems).
Vector Database “Powered by a vector database” A database that stores embeddings and enables the AI to find relevant information quickly. The plumbing behind RAG. Not proprietary — most vendors use the same 3-4 open-source or commercial options.
Prompt Engineering “Our prompt engineering ensures quality” Writing the instructions that tell the AI what to do. This is important but not proprietary technology. A good prompt is like a good brief — it improves the output. The vendor’s prompts can be replicated.
Multi-Modal “Our multi-modal AI processes text, images, and video” The AI handles multiple input types, not just text. Useful for document processing (invoices with logos, forms with handwriting). Ask: which modes are production-ready and which are “coming soon”?
Guardrails “Enterprise-grade guardrails” Rules that prevent the AI from generating harmful, inaccurate, or off-topic responses. Critical for production use. Ask: “Can I define custom guardrails for our industry’s compliance requirements?”
Context Window “200K token context window” How much text the AI can consider at once. A 200K-token window processes ~150 pages. Relevant for document-heavy workflows. Larger is not always better — performance degrades near the limits.
Inference “Low-latency inference” The AI generating an output from an input. Every time the AI answers a question, that is inference. Inference costs money. This is the usage-based cost that surprises buyers at month three.
SLM (Small Language Model) “Our SLM runs on-premise for data privacy” A smaller, cheaper AI model that can run on your own servers. Less capable than GPT-4 or Claude, but keeps data inside your network. Good for narrow, well-defined tasks. Not a replacement for general-purpose AI.
Copilot “Your AI copilot for [function]” An AI assistant that suggests actions while a human makes the final decision. Microsoft branded this term, but every vendor uses it. The key question: does the human actually review the output, or does the “copilot” quietly become the pilot?
AI-Native “Built AI-native from the ground up” The application was designed around AI, not retrofitted. Theoretically means tighter integration. In practice, “AI-native” is often marketing for “we added AI to the login page.” Ask what the product did before AI and what changed.
Model Drift Vendors rarely mention this AI model performance degrades over time as the data it was trained on becomes stale. A model trained on 2024 data gives worse answers about 2026 conditions. Ask: “How often is the model updated, and how do you detect performance degradation?”
Data Residency “Data stays in your region” Where your data is physically stored and processed. Critical for GDPR and industry regulations. Microsoft’s Anthropic integration routes data outside the EU Data Boundary. “Data residency” claims need sub-processor verification.
Sub-Processor Buried in the DPA appendix A third party that processes your data on behalf of your vendor. When Microsoft uses Anthropic, Anthropic is a sub-processor. GDPR requires 30-day notice before adding sub-processors. Most vendors add them with a blog post, not a formal notification.
Zero-Shot / Few-Shot “Works out of the box with zero-shot capability” Zero-shot: the AI performs a task without examples. Few-shot: it needs a handful of examples. Vendors claim zero-shot to imply no setup. Reality: few-shot (giving the AI examples of your data) almost always performs better.
Latency “Sub-second response times” How long it takes the AI to respond. Matters for customer-facing applications. Vendor claims are measured under ideal conditions. Ask: “What is the p95 latency under production load with our data volume?”
Orchestration “Our orchestration layer coordinates multiple AI models” Software that routes tasks to different AI models based on the request. Like a traffic controller for AI. Sounds sophisticated. Ask: what models, and can you switch providers if one raises prices 300%?
Total Cost of Ownership (TCO) Vendors avoid this phrase The full cost: license + implementation + integration + training + productivity dip + ongoing support + overages. The license is 25-40% of actual year-1 cost for most AI tools. The vendor will not calculate this for you. You must (RSM, n=966 mid-market firms, 2025; Xenoss TCO analysis, 2025).

Part 5: Reading Provider Case Studies — The Methodology Problem

OpenAI, Anthropic, Microsoft, and Google each publish case study libraries. The cases are real. The results are often real. The problem is structural: every published case is a selected winner, with no control group and no independent verification. They represent what is possible under ideal conditions — not what the average deployment produces.

The standard caveat to apply to every vendor case study:

These are vendor-published, represent selected wins, and have no control group and no independent verification. Cross-reference against independent baselines: METR RCT (n=16 experienced developers, 246 tasks, July 2025 — 19% slower with AI); CMU analysis (807 AI-assisted repos — 39% cognitive complexity increase); Atlan 200-deployment B2B analysis (median +159.8% ROI over 24 months, but 27% failure rate, success concentrated in projects where training consumed 25%+ of budget).

The Three Questions That Cut Through Case Study Marketing

1. What was the scope? The strongest vendor cases involve one specific, narrow workflow with a clear input and output. “Morgan Stanley expanded answerable queries from 7,000 to 350,000” is specific and verifiable. “Moderna deployed AI to 80% of its workforce” is a deployment fact, not an outcome.

2. Was there a baseline? Cases without a before/after comparison are announcements, not evidence. “50% faster” is only meaningful if you know what “baseline speed” means and when it was measured.

3. Who measured it? Vendor-reported metrics rely on the vendor’s customer telling the vendor what happened. BCG’s independent sampling (n=10,635, June 2025) and McKinsey’s independent survey (n=1,993, November 2025) both find only 5–6% of organizations generating substantial financial gains from AI — a very different picture from vendor case study portfolios.

What Provider Case Studies Are Useful For

  • Existence proof: Yes, this technology can work for this type of task. The question is whether your organization has the conditions that made it work there.
  • Scoping reference: The cases define what “narrow, well-designed application” looks like for your industry. Use them to scope your pilot, not to predict your outcome.
  • Structural conditions audit: The strongest cases share three conditions — narrow task scope with clear success criteria; workflow redesigned before deployment; human oversight maintained. Use each case to ask: “Which of these conditions does my organization have in place?”

Key Data Points

Metric Finding Source
Vendor AI agents failing expectations 45% of martech leaders report underperformance Gartner, n=413, June-August 2025
Real vs. claimed “agentic AI” vendors ~130 genuine out of thousands claiming it Gartner, January 2025 poll (n=3,412)
IT leaders with surprise AI charges 78% in last 12 months Zylo 2026 (40M+ licenses analyzed)
Organizations with substantial financial gains from AI 5–6% BCG AI at Work (n=10,635, June 2025) + McKinsey State of AI (n=1,993, Nov 2025)
Median cost of abandoned AI project $4.2M over 11 months Pertama Partners, n=2,400+ initiatives
AI vendor liability cap 88% cap at one month’s subscription CIO, October 2025
SEC AI washing enforcement (2024-2025) 4 actions including fraud charges SEC (Delphia, Global Predictions, Presto, Nate Inc.)
Agentic AI project cancellation prediction 40%+ by end of 2027 Gartner, January 2025 poll (n=3,412)
Experienced developer speed with AI (METR RCT) 19% slower — independent baseline for vendor coding claims METR, n=16, 246 tasks, July 2025
AI-assisted repos cognitive complexity change +39% increase CMU analysis, 807 repos, through August 2025

What This Means for Your Organization

The vendor pitch is not the problem. The pitch is doing what pitches do. The problem is that most mid-market companies lack an internal framework for evaluating AI claims at the speed vendors deliver them. A CIO at a 300-person company sees 5-10 AI vendor pitches per quarter, each with its own terminology, its own metrics, and its own urgency. Without a shared reference for what the words mean and what the evidence shows, every pitch starts from zero.

This card is a filter, not a verdict. Some vendors with red flags have good products. Some vendors with green flags oversell. The card gives the evaluation team a common vocabulary and a shared set of questions — so the conversation after the demo is about fit, not about decoding jargon.

The three questions for existing vendor features deserve the most immediate attention. Microsoft, Google, Salesforce, and Zoom have already changed what their tools do with your data. Most organizations have not updated their vendor assessments or client agreements to reflect those changes. A 30-minute admin console review this week surfaces more actionable risk than any new vendor evaluation.

If this card raised questions about a specific vendor conversation — or about the admin console review for tools already in your stack — I would welcome that conversation at brandon@brandonsneider.com.

Sources

  1. Gartner — “Survey Finds 45% of Martech Leaders Say Existing Vendor-Offered AI Agents Fail to Meet Their Expectations of Promised Business Performance,” October 29, 2025. n=413 marketing technology leaders, surveyed June-August 2025. Source credibility: Independent analyst firm, defined methodology. High credibility. https://www.gartner.com/en/newsroom/press-releases/2025-10-29-gartner-survey-finds-45-percent-of-martech-leaders-say-existing-vendor-offered-ai-agents-fail-to-meet-their-expectations-of-promised-business-performance

  2. Gartner — “Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 25, 2025. January 2025 poll, n=3,412 webinar attendees. ~130 genuine agentic vendors identified. Source credibility: Independent analyst firm. High credibility for vendor landscape assessment. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

  3. SEC — “Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence,” March 18, 2024. Delphia ($225K penalty), Global Predictions ($175K penalty). Source credibility: Primary regulatory enforcement. High credibility. https://www.sec.gov/newsroom/press-releases/2024-36

  4. SEC / StoneTurn — “Next-Generation Compliance: Preparing for Continued SEC AI Washing Enforcement,” 2025. Presto Automation (January 2025), Nate Inc. ($42M fraud, April 2025). CETU unit created February 2025. Source credibility: Primary enforcement actions via independent compliance advisory. High credibility. https://stoneturn.com/insight/next-generation-compliance-preparing-for-continued-sec-ai-washing-enforcement/

  5. Zylo — 2026 SaaS Management Index, March 2026. 40M+ SaaS licenses analyzed. 78% of IT leaders reported unexpected AI charges. AI-native app spend up 108% YoY. Source credibility: Independent SaaS management platform; actual license data, not survey self-reporting. High credibility. https://zylo.com/reports/2026-saas-management-index/

  6. CIO — Deshpande, Stines, Vaughan. “Your vendor’s AI is your risk: 4 clauses that could save you from hidden liability,” October 30, 2025. 88% of AI vendors cap liability at one month’s subscription. Source credibility: Independent IT publication; practitioner-authored. High credibility. https://www.cio.com/article/4081326/your-vendors-ai-is-your-risk-4-clauses-that-could-save-you-from-hidden-liability.html

  7. CloudEagle.ai — “How to Evaluate AI Tools Before You Buy,” 2025-2026. 70% failure rate without defined use cases. 40% fewer implementation issues with structured pilots. Source credibility: Vendor-published framework. Moderate credibility; methodology is sound. https://www.cloudeagle.ai/blogs/how-to-evaluate-ai-tools

  8. DUNNIXER — “Six Dimensions of AI Vendor Evaluation for Enterprise RFPs,” 2025. PoC on customer data as primary red flag. Six-dimension evaluation framework. Source credibility: Independent advisory. Moderate-high credibility. https://www.dunnixer.com/insights/articles/the-six-dimensions-of-ai-vendor-evaluation-that-matter-most

  9. Pertama Partners — “AI Project Failure Statistics 2026.” n=2,400+ enterprise AI initiatives. $4.2M average failure cost. 62% achieve <40% adoption in six months. Source credibility: Independent advisory; large dataset. High credibility. https://www.pertamapartners.com/insights/ai-project-failure-statistics-2026

  10. BCG — “AI at Work 2025.” n=10,600+ workers, 11 countries. Only 5% of organizations achieving substantial AI returns. Source credibility: Independent survey. High credibility. https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain

  11. 2toLead — “Anthropic Models On by Default in Copilot,” December 8, 2025. Microsoft sub-processor change documentation. Source credibility: IT advisory; detailed technical analysis. High credibility for factual timeline. https://www.2tolead.com/insights/anthropic-models-on-default-copilot-admin-action-plan-and-risks

  12. Gartner — “60% of AI Project Failures Stem from Inadequate Data Governance,” 2025. Source credibility: Independent analyst firm. High credibility. https://www.gartner.com/

  13. AI Xccelerate — “The AI Vendor Selection Guide: Red Flags, Green Flags, and Questions That Expose the Truth,” 2025. Comprehensive vendor evaluation framework. Source credibility: Independent advisory. Moderate credibility. https://www.aixccelerate.com/blog/the-ai-vendor-selection-guide-red-flags-green-flags-and-questions-that

  14. RSM — “Middle Market AI Survey 2025,” March 2025. n=966 decision-makers. 91% use GenAI; 70% need outside help. Source credibility: Independent industry survey of mid-market firms. High credibility. https://rsmus.com/insights/services/digital-transformation/rsm-middle-market-ai-survey-2025.html

  15. Xenoss — “Total Cost of Ownership for Enterprise AI,” 2025. License is 25-40% of actual year-1 cost. Source credibility: Services firm; data is well-cited. Medium credibility. https://xenoss.io/blog/total-cost-of-ownership-for-enterprise-ai

  16. METR — “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” July 2025. n=16 experienced developers, 246 tasks, randomized crossover. 19% slower with AI; perceived 20% faster. Source credibility: HIGH — independent RCT, pre-registered, open-source methodology. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

  17. CMU / GitHub repository analysis — Analysis of 807 AI-assisted repos vs. 1,380 control repos, tracking through August 2025. 39% increase in cognitive complexity in AI-assisted repos. Source credibility: HIGH — independent observational study with control group. (Reported via Rob Bowley / DevOps.com, December 2025.)

  18. Denis Atlan / SSRN — “AI ROI Analysis: Evidence from 200 B2B Deployments (2022–2025).” Median +159.8% ROI over 24 months; 27% failure rate; success concentrated in projects where training consumed 25%+ of budget. Source credibility: MEDIUM — independent, dataset publicly available under CC BY 4.0; French mid-market context, not directly U.S.-comparable.

  19. BCG — “AI at Work 2025” (n=10,635, June 2025) and “Widening AI Value Gap” (n=1,250, September 2025). 72% regular AI use; only 5% generating substantial financial gains. Source credibility: MEDIUM-HIGH — large sample; BCG advisory conflict noted. https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain

  20. McKinsey & Company — “State of AI 2025,” November 2025. n=1,993. 88% AI usage; only 6% high performers (>5% EBIT impact). Source credibility: MEDIUM-HIGH — large consulting survey; self-reported. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

Brandon Sneider | brandon@brandonsneider.com March 2026