Applied AI Notes · What Businesses Are Asking AI to Do · November 2026
What Moved Since October
Selected measures from the October edition, set against November.
Each month I review requirements that businesses publish online for AI systems that act through software tools. In this edition, about 200 projects published over roughly two weeks to mid-November 2026 are set against about 220 in the October edition, using identical sources, window and codebook. On most measures, November looks much like October. Where shares moved further, they are set out below, although two months of data collected in one place cannot establish a trend.
Several business areas changed share
E-commerce and retail operations now account for 8% of projects, compared to 2% a month earlier, while AI products that businesses are building to sell account for 12%, against 5%. Customer support dropped to 4% (down from 10%), matching a smaller share of chat and voice assistants (20%, against 27%), although that second difference does not reach unadjusted 5% significance. Trading and crypto projects, at 4%, remain too few to interpret with any confidence.
Autonomous operations accounted for a larger share
Among the tasks requested, autonomous operations, meaning systems intended to run part of a business with little direction, were the only task category whose difference reached unadjusted 5% significance, reaching 7% of projects compared to 2% the month before. Business workflow automation, multi-agent platforms and tool integrations through the Model Context Protocol (MCP) each stayed within a point or two of their October shares.
The calls for agents are no more agentic
Of the projects that ask for an “agent”, 18% describe a model choosing its next step, also 18% after rounding in October. Fixed workflows fell to 41% of these projects (previously 49%), while requirements too vague to tell rose to 36% (27%). A larger share was classified as too vague to tell, although that difference does not reach unadjusted 5% significance.
Checks and delivery are close to October
Projects that state no success check rose to 46% (up from 39%), while those requesting testing or evaluation dropped to 46% (53%). Neither difference reaches unadjusted 5% significance, and more editions will be needed to assess whether either pattern persists. Human review, approval or handoff, mentioned in 44% of projects, is essentially unchanged.
Claude is named more often
Claude is now named in the requirements of 39% of projects, compared to 26% a month earlier, as is Python (39%, against 26%). OpenAI held at 30%, while low-code automation platforms declined to 30% (previously 36%), a difference that does not reach unadjusted 5% significance. Because these figures count tools named in the requirements, they may not capture every tool in use.
What held steady
Most of what the October edition described still holds, with the figures below moving by less than three points.
- Human review, approval or handoff: 44% (October 46%)
- OpenAI: 30% (October 30%)
- Flagged as a possible agent case: 13% (October 13%)
- Model chooses next step (among projects asking for an “agent”): 18% (October 18%)
- Business workflow automation: 16% (October 15%)
- Multi-agent platform: 11% (October 10%)
- Tool integration (MCP): 9% (October 8%)
- Content pipeline: 7% (October 8%)
- A person hands over a job: 22% (October 23%)
- Fixing or extending an existing system: 21% (October 20%)
- Brief specification: 27% (October 26%)
- Agent frameworks: 13% (October 12%)
For a full analysis behind each measure, the October edition remains the reference: Called an agent, specified as a workflow, How will anyone know it works?, Before you commission an agent.
All figures as a table
| What the system is to do | October | November |
|---|---|---|
| Chat or voice assistant | 27% | 20% |
| Business workflow automation | 15% | 16% |
| Multi-agent platform | 10% | 11% |
| Tool integration (MCP) | 8% | 9% |
| Content pipeline | 8% | 7% |
| Research or monitoring | 4% | 7% |
| Autonomous operations * | 2% | 7% |
| Outreach sequencing | 4% | 2% |
| Knowledge Q&A | 4% | 3% |
| Document extraction | 2% | 4% |
| Testing, reliability, safety | 4% | 1% |
| Coding agent or dev tool | 2% | 3% |
| Other | 9% | 10% |
| Business area | October | November |
|---|---|---|
| Sales and lead generation | 16% | 12% |
| Internal operations | 12% | 10% |
| Marketing and content | 11% | 9% |
| AI products built to sell * | 5% | 12% |
| Customer support * | 10% | 4% |
| Real estate | 5% | 8% |
| E-commerce and retail * | 2% | 8% |
| Healthcare | 4% | 4% |
| Data and analytics | 4% | 3% |
| Software and developer tools | 4% | 3% |
| Personal productivity | 3% | 3% |
| Trading and crypto * | <1% | 4% |
| Design described, among projects asking for an “agent” | October | November |
|---|---|---|
| Fixed workflow | 49% | 41% |
| Model chooses next step | 18% | 18% |
| Plain code or one model call | 6% | 5% |
| Too vague to tell | 27% | 36% |
| How success is to be checked | October | November |
|---|---|---|
| Reference check | 49% | 45% |
| Metric or model-based assessment | 6% | 4% |
| Someone's judgment | 6% | 5% |
| No check stated | 39% | 46% |
| What else the requirements specify | October | November |
|---|---|---|
| Names the systems to connect | 70% | 75% |
| Human review, approval or handoff | 46% | 44% |
| Deployment, monitoring, maintenance | 48% | 54% |
| Testing or evaluation | 53% | 46% |
| Guardrails, privacy, compliance | 46% | 51% |
| A step that needs interpretation | 38% | 32% |
| Cost or speed limits | 22% | 16% |
| A simpler option | 13% | 13% |
| What starts the work | October | November |
|---|---|---|
| Events or schedules | 53% | 50% |
| A person hands over a job | 23% | 22% |
| Unclear | 24% | 28% |
| Tools named | October | November |
|---|---|---|
| n8n, Make or Zapier | 36% | 30% |
| OpenAI | 30% | 30% |
| Claude * | 26% | 39% |
| Python * | 26% | 39% |
| MCP | 18% | 17% |
| Agent frameworks | 12% | 13% |
| Kind of work and detail | October | November |
|---|---|---|
| Fixing or extending an existing system | 20% | 21% |
| Brief specification | 26% | 27% |
| Flagged as a possible agent case | 13% | 13% |
Method and limits
For this edition I collected requirements that businesses published online for AI projects over roughly two weeks to mid-November 2026. From these, I keep projects that ask for a system to act through software tools and that aim to build new systems or to fix or extend existing ones. A language model then classifies each set of requirements against a fixed codebook, built on the design process and task-screening worksheet I use in practice. The October figures are recomputed here with the same filters, from that edition's retained labels.
One limit is scope: because the requirements were published in one place over about two weeks, they describe these projects rather than the whole market. All shares refer to projects rather than businesses, since a single business can publish several projects. In a separate review of the success-check label in each edition, a second model labelled every project twice, and where its labels disagreed with the original, a third model settled the label under written rules, although agreement by itself cannot establish correctness. Unanimous labels were also checked against the same rules, with any proposed changes settled by the third model. Other labels were never independently re-coded, with no manual audit of any label. The classification read the full text of each set of requirements, so the result reflects what businesses state rather than the system eventually built. Changes describe the mix of requirements, not market growth.
Use of AI tools
A language model classified the requirements against my codebook and helped to draft the analysis code and this text. Other AI models re-checked the success-check labels and reviewed the method and claims. Having designed the framework and the codebook and checked the figures, I am responsible for the conclusions.