What the dataset is, and what it isn't
IoT Analytics' Generative AI & Agentic AI Market Report 2026-2032 is a 283-page analysis published in September 2026. It includes an empirical analysis of 1,533 enterprise AI projects. The companion database is the Generative AI & Agentic AI Projects List 2026. It is up from 530 projects in the previous edition, which covered generative AI only and counted projects from 2022 to 2024.
Per IoT Analytics' product description, every entry rests on public information a vendor posted on its own website, such as a customer case study or partner success story. Projects come from Microsoft, Google Cloud, AWS, OpenAI, Anthropic, Alibaba Cloud and Oracle, among others. A project is included only if enough public detail exists to classify it. The list spans more than 1,200 named customer organizations and 60 countries.
That has consequences for how to read the numbers:
- The percentages describe how often an activity is tagged among publicly showcased projects. They do not measure what share of companies use AI, or how well deployments work.
- A project can carry up to five activity tags. The activity shares therefore overlap and can't be added together.
- The ranking itself covers the 734 projects from 2025 and Q1 2026.
- The reported results, such as time saved, are company-reported. They are not controlled comparisons.
The top 10 enterprise AI applications
IoT Analytics ranks the activities by share of the 734 projects from 2025 and Q1 2026:
| Rank | Activity (department) | Share of projects |
|---|---|---|
| 1 | Process optimization (Operations) | 52.2% |
| 2 | Content creation (Marketing) | 30.1% |
| 3 | Inquiry handling (Customer service) | 22.2% |
| 4 | Software development (IT) | 16.5% |
| 5 | Reporting (Operations) | 13.2% |
| 6 | Collaboration (IT) | 11.6% |
| 7 | Issue resolution (Customer service) | 10.8% |
| 8 | Product & research analytics (R&D) | 10.4% |
| 9 | Regulatory compliance (Legal/compliance) | 8.4% |
| 10 | Quality control (Operations) | 7.6% |
Process optimization held the top spot, rising from about 30% in 2022–2024. Quality control made the biggest jump, climbing 12 places from 3.5% to 7.6%. Issue resolution fell the furthest in the top 10, from 23.8% to 10.8%. By department, operations (22.8% of projects) has overtaken IT (20.2%), which led in 2022–2024.
Agentic AI: growing, but from a small base
IoT Analytics says generative AI is entering mainstream adoption, while agentic AI is not there yet. Single- and multi-agent systems made up 13.4% of 2025–Q1 2026 projects, up from 0.5% in 2022–2024. Single agents were 11.3% and multi-agent systems 2.0%. Non-agentic chatbots and assistants still made up 60.1%, down from 65.0%.
The autonomy figures need careful reading, because the report uses two different comparisons:
- Year-over-year framing: Projects that executed multi-step tasks autonomously within a defined scope rose from 1.1% in 2025 to 17.8% in Q1 2026. Output-only projects fell from 59.5% to 32.2%. IoT Analytics warns that the 2026 numbers rest on only 174 projects, so they show direction, not a stable share.
- Period-over-period framing: Across 2025–2026 as a whole, 53.0% of projects only generated outputs, down from 66.8%. Workflow assistance rose from 33.2% to 42.0%. Bounded autonomous execution reached 5.0%, against none in 2022–2024.
Don't read 17.8% as the real autonomy level in enterprises today. The 5.0% figure is the steadier one.
By industry, agentic shares were highest in technology (16.8%), professional services (16.2%) and finance (13.7%). They were lowest in manufacturing (4.3%) and education (none). Technology companies still account for 49.7% of projects, down from 54.5%.
IoT Analytics says its taxonomy is meant to stop a system being labeled agentic just because a vendor calls it an agent. That matters, since "agent" is a heavily used marketing word right now.
The Microsoft-stack examples
Several of the case studies involve tools Windows and Microsoft 365 admins already manage. All results below are as reported by the companies and relayed by IoT Analytics.
- NielsenIQ (process optimization): NIQ used Microsoft Foundry and Azure OpenAI to turn product-packaging images into structured data. It reports coding 32,000 products in 10 hours, against 300 hours before, and a 90% cut in item-coding time. Data scientists still oversee quality.
- Carvana (inquiry handling): Carvana built two tools with Microsoft. One is the Sebastian customer agent. The other is CARE, a conversation-review platform on Azure AI Foundry that analyzes all customer conversations rather than a sample. Carvana reports inbound calls per sale down more than 45% over two years.
- PepsiCo (collaboration): The company standardized more than 250,000 employees on Teams and moved 3,300 conference rooms to Teams Rooms. It then deployed Microsoft 365 Copilot, which runs on the existing Microsoft 365 access controls, so it only reaches content a user can already see. Daily active use is reported at 90–95%. The source doesn't say how that is measured.
- Hertz (issue resolution): A Copilot Studio agent called Manny gives roadside-assistance representatives vehicle-specific instructions. An early pilot reportedly cut resolution time by more than 15%. This is a pilot result.
- Estée Lauder Companies (R&D analytics): ConsumerIQ, built on Copilot Studio and Azure OpenAI Service, answers product developers' questions using internal repositories. The company reports data-gathering time falling from weeks to minutes.
- Dunaway (regulatory compliance): A Power Automate workflow collects code documents from Texas cities weekly and stores them in SharePoint. A Copilot Studio agent loads them, keeps each city's codes separate, and cites the source page. Unclear cases go to subject-matter experts. The firm reports a 90% cut in manual research time.
The non-Microsoft examples include Delivery Hero's Herogen coding agent. It runs on Google Cloud with Anthropic's Claude models and picks up Jira tickets, writes code, runs tests and submits changes. It reportedly merges more than 170 pull requests a day, over 10% of the company's total. Initial reporting cited more than 100 a day and an 85% success rate on assigned tickets. The source doesn't define "success." Kraft Heinz, Copel and Penda Health round out the list.
What admins and developers can take from this
The Dunaway and PepsiCo cases show a pattern that's easy to miss in the headline percentages. They depend on unglamorous plumbing: scheduled document ingestion, existing permission models, phased rollouts with training. They don't rely on a clever model alone. In PepsiCo's case, Copilot's reach is bounded by the Microsoft 365 permissions already in place. Any permissions sprawl in SharePoint or Teams therefore becomes an AI exposure problem. This is general industry reasoning, not a finding from the report.
Some questions to ask before copying any of these projects:
- What is the measurement baseline? "90% less time" means little without knowing how time was counted.
- Is the result a pilot or a production rollout? Hertz's figure is explicitly a pilot.
- Who reviews the output? The compliance and quality-control examples keep human experts in the loop.
- Does the tool inherit your access controls, or does it create a new data path?
Caveats
IoT Analytics sells the underlying database, starting at $800 for a team license. It benefits from a narrative of fast-moving adoption. The data is also skewed toward vendors that publish glossy case studies, which favors success stories and large, named customers. Failed or quietly shelved projects don't appear. The sample is therefore a useful map of where companies are showing AI results. It is not a measure of how common each use is, or how often it works.
Even so, the direction is plausible and consistent with the shift IoT Analytics describes: from drafting text toward carrying out workflow steps. For IT teams, that means agent governance, identity and permissions are becoming day-to-day operational concerns rather than future ones.