Research edition · 14 September 2026
The report’s observations, prices and forecasts are a dated snapshot. Examples of savings are estimates unless explicitly identified as study results. This is not a fresh verification of every claim.
| Builder group | Strategic position | Representative products |
|---|---|---|
| Major model providers | Own frontier models, distribution, APIs, and assistant surfaces | OpenAI, Anthropic, Google, Microsoft, Meta, xAI, Mistral |
| Productivity-suite companies | Own the system of record and user permissions | Microsoft, Google, Salesforce, Atlassian, Slack, Notion, Coda, Asana, ClickUp |
| Enterprise AI/search companies | Normalize private context, permissions, and enterprise actions | Glean, Coveo, Sinequa, Rovo, Gemini Enterprise |
| Independent AI startups | Win a narrow workflow through better UX or domain data | Perplexity, Fathom, Granola, Fireflies, Reclaim, Motion, Mem |
| Automation platforms | Turn models into repeatable cross-app processes | Zapier, Make, n8n, Workato, Power Automate |
| Vertical software companies | Embed AI where records, policies, and revenue already live | Salesforce, HubSpot, Intercom, Zendesk, ServiceNow, Workday, Intuit |
| Open-source and local ecosystem | Reduce lock-in and support local/private deployment | Obsidian plugins, Ollama, Open WebUI, Llama-family deployments, n8n self-hosting |
| Creative software companies | Combine generative models with editability, rights, and asset libraries | Adobe, Canva, Figma, Descript, Runway |
| Universities and labs | Establish benchmarks, causal evidence, safety methods, and new interfaces | Stanford, MIT, Harvard/BCG, Microsoft Research, Princeton, METR, NBER, OWASP |
Case studies and counterexamples
- Customer support: The NBER field study is a useful positive case because it measured real agents, real conversations, and a concrete outcome rather than clicks or sentiment alone.[38]
- Email and documents: Microsoft’s randomized field experiment suggests individual time savings can appear before teams change coordinated behaviors such as meetings.[40]
- Enterprise context: Glean, Microsoft Copilot, Gemini Enterprise, Slack AI, and Notion compete on permissioned context and workflow position more than on model novelty alone.[5][8][13]
- Coding: The evidence is mixed by task and population. Combined company experiments found more completed tasks, while METR found a slowdown among experienced open-source developers working in mature repositories. A credible directory should show both, not select only the positive result.[41][42]
- Product churn: Google’s July 2026 rename of NotebookLM to Gemini Notebook shows why product pages need an “as reviewed” date, aliases, redirects, and historical pricing rather than a static feature list.[7]
- Marketing-driven claims: Vendor pages frequently use claims such as “10x productivity,” “finish 137% more work,” or “save four hours every week.” These are leads for investigation, not evidence. The website should quote the claim, label it as vendor-supplied, and seek a methodology or independent validation.
- Operational failure is possible even in polished institutions: The Associated Press reported that a Deloitte Australia report delivered to the government contained apparent AI-generated errors and nonexistent references, leading to a partial refund.[48] The lesson is not that professional firms cannot use AI; it is that source verification and accountable authorship remain necessary.