CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
Mapping your skills to real jobs and career opportunities.
What job is this course for? Here is my answer, using the data available in November 2026. I lined up what you learned this semester against six public sources.
The page names come from this course. Each one practised one or more skills from the course skills inventory, and every skill belongs to one of five groups. A dot means that week built at least one skill in that group. Press Jobs → on any week to see the jobs that use what it taught.
| Week | Pages and the skills each one practised | Build & use | Evaluate | Secure | Decide | Automate | Show jobs |
|---|---|---|---|---|---|---|---|
| Week 1Introduction to Agentic AI |
| ||||||
| Week 2Unpacking Coding Agents |
| ||||||
| Week 3Prompting & Context Engineering |
| ||||||
| Week 4Reasoning, Planning & the Agent Loop |
| ||||||
| Week 5Tool Use & Function Calling |
| ||||||
| Week 6MCP & A2A |
| ||||||
| Week 7Memory, RAG & Knowledge |
| ||||||
| Week 8Evaluation & Reliability |
| ||||||
| Week 9Multi-Agent Systems |
| ||||||
| Week 10Security, Trust & Governance |
| ||||||
| Week 11Agentic Commerce |
|
The capstone pulls all five groups together: you build an agent, test it, secure it, and argue for whether it should exist.
Here is how I would size up the market if I were graduating with an MIS degree and this course behind me. Strengths and weaknesses are about you. Opportunities and threats come from the data further down. Click any job to find it on the chart.
The usual MIS targets, all still growing
Less obvious choices where this course gives you a head start
Comparing the task descriptions of more than 900 O*NET jobs with the skills taught in this course turned up three jobs I had not expected: logistics analysts, web and digital interface designers and validation engineers. The daily work O*NET lists for each of them is among the closest matches to what you practiced here.
One job sits in both columns. OpenAI rates as higher risk, and BLS projects them to grow 28.5%. The box in the evidence section explains why.
The chart is the 2×2. Further right, AI can do more of the job’s tasks. Higher up, demand grows more when the work gets cheaper. Top right is where AI and hiring grow together. Bottom right is where the work gets automated faster than demand grows. Point at a job for its numbers, or filter by skill group.
Can you apply to all of these? Not right away. Groups 1 to 4 are jobs that use what you learned, and the column “Start after graduation” shows which ones hire people straight out of a bachelor’s. Group 5 is different. Those jobs are here because agents take over their task lists, and the people who build those agents work in groups 1 to 4. Jobs marked Found by task matching were discovered by comparing O*NET task descriptions with the course skills in Part 4, and Weak fit marks the one job the test did not support well.
| Occupation | Path | Start after graduation | Median pay | Openings a year | Growth | Elasticity | AI exposure |
|---|
I matched skill groups to jobs by hand, so I checked the result against an independent source: the official task lists the Department of Labor keeps for 923 occupations in O*NET ↗. A text-similarity model compared what each skill group covers, using the course pages and the skills inventory, with the tasks each job actually involves. Where the test disagreed with me, I changed the matching.
| First version | Revised version | |
|---|---|---|
| Jobs whose best-matching group is the one assigned | 45% | 68% |
| Chance level | 20% | 26% |
| Median percentile among knowledge-work jobs | 90th | 91st |
| Jobs in the top quarter for their group | 27 of 31 | 25 of 28 |
The revised numbers leave out the six added jobs, because the check picked them. With them included, the share rises to 74%.
Every month OpenAI sorts a sample of a million consumer ChatGPT messages and publishes the results by country in OpenAI Signals ↗. One table puts work-related messages into O*NET ↗ work activities, the categories the Department of Labor uses to describe jobs. Here are the top twelve for the United States, July to September 2026. Under each one I noted where it came up in this course.
Most people use ChatGPT at work as a writing assistant. The things an agent does, like writing code, fixing computer problems and looking things up through tools, add up to less than 8% of work messages. Agents that do this work for other people are still built by a fairly small group of developers. This course trains you to join them.
In July 2024, 51.6% of consumer messages worldwide were work-related. By September 2026 the figure was 32.2%, and 33.1% in the U.S. Use outside work grew faster. Business accounts are left out of this data, so it can’t tell you how much AI is used at work overall.
In September 2026, 47.8% of work messages asked the model to produce something. Another 27.3% asked it for information. Outside work it goes the other way, with 41.5% asking and 27.3% producing. Handing off a task is exactly what an agent is for.
Georgia ranked 10th among U.S. states for ChatGPT messages per person in early 2025. It has been 7th for each of the last three quarters. Of all the topics, technical help is the most tied to work. 57% of those messages are work-related, compared with 49% for writing.
OpenAI published this framework ↗ on April 25, 2026, with a full report ↗ and a downloadable occupation table. Most studies of AI and jobs ask one question, whether AI can do the tasks in a job. This one asks two more. Does a person still need to deliver the work or take responsibility for it? And if AI makes the service cheaper, will people buy enough more of it to keep the workers busy? Each occupation lands in one of four groups based on the answers. OpenAI is careful to say the groups are not forecasts of job loss. They show where change is likely to appear first.
The number that separates the groups is demand elasticity. Above 1 means that when a service gets 10% cheaper, people buy more than 10% more of it. The occupation can then add workers even as each worker gets more done. The highest value in OpenAI’s table, 1.15, belongs to computer systems analysts and computer network architects. Software developers are at 1.13 and data scientists at 1.12. Customer service representatives are at 0.68, and general office clerks at 0.41.
One caution before you read further. OpenAI’s write-up ↗ names software developers and financial analysts as examples of jobs that will reorganize. Its downloadable table puts both in the grow-with-AI group. I used the table everywhere on this page.
Two official sources, two different answers. OpenAI’s table puts information security analysts in the higher-automation-risk group. BLS projects ↗ the same occupation to grow 28.5% by 2034, one of the fastest rates among computer jobs. They are measuring different things. OpenAI asks how much of today’s task list a model could do. BLS asks how many people employers plan to hire. My reading is that the routine part of security work is exposed, and the judgment about what an agent is allowed to touch is what employers are hiring for. That judgment is Week 10.
Customer service shows the same gap from the other side. About 70% of a representative’s tasks are exposed, and BLS projects 153,700 fewer of these jobs by 2034. The people who build and supervise the agents that take over that work are counted in other occupations, mostly the ones in groups 1 to 4 above.
For GDPval ↗, OpenAI picked the nine sectors that contribute most to U.S. GDP. In each one it took the five best-paid occupations that are mostly done on a computer. Professionals from those jobs, with 14 years of experience on average, wrote 1,320 tasks from their own work. Each task comes with reference files and a finished product, such as a spreadsheet, a legal brief or a care plan. Other experts then compare a model’s product with the human one without knowing which is which, and score it as a win, a tie or a loss. The 220 public tasks are on Hugging Face ↗. I bolded the occupations this course connects to most directly.
In the September 2025 paper ↗, the best model matched or beat the expert on 47.6% of tasks. The runner-up scored 38.8%, and a 2024 model 12.4%. The live leaderboard ↗ has moved since then.
Look at what gets judged. It is the finished file, and the conversation that produced it doesn’t count. Then look at how it gets judged: an expert compares two outputs side by side with the labels hidden. You did the same thing with the harness in Week 8 and the rubric in Group Assignment 3. I expect employers in these occupations to need people who can run that kind of test on their own work.
Every GDPval task arrives with a clear prompt and all the files the model needs. The model never has to work out what the task is, ask a clarifying question, or recover when a tool breaks. That is most of what an agent does, and most of what we covered in Weeks 4 to 10. A good GDPval score tells you the writing and analysis hold up. The benchmark doesn’t test the loop around them.
Lightcast ↗ counted more than 86,000 job postings in 2025 that asked for agentic AI skills, up more than 280% in a year. The median advertised salary was $153,000. For AI postings in general it was $141,000, and for IT $125,000.
Fortune 1000 companies posted about 30% of these jobs, although they post only 12% of jobs overall. Amazon, Microsoft, KPMG and Accenture are among the top ten employers. Six of the ten occupations that ask most often are technical, including AI engineers and software developers. The authors point out that a search by job title would barely find this work, because companies are adding it to jobs that already exist.
These are shares of consumer messages, with noise added for privacy. They show what people ask ChatGPT to do. They don’t show whether anyone used the answer, and they leave out business and API traffic, which is where most agents run.
Both start from the O*NET list of tasks for each job. The exposure scores describe what a 2023 model could do. The 2026 framework adds whether a person is still needed and how demand responds. Neither one watches jobs actually change.
Experts graded finished work on clearly defined tasks. OpenAI built the benchmark and grades its own models on it, although the leaderboard lists other companies’ models too. It measures the quality of the output. It leaves out the rest of a workday, like figuring out what the task is or recovering from a mistake.
For PATH partners. This page maps occupations for the AI and Business pathway as I teach it in CIS 4394. The six clusters come from the course skills inventory, which lists 35 skills with the week each one is taught and the assignment that assesses it. The crosswalk includes O*NET-SOC codes, so you can join it to O*NET ↗, BLS ↗ or OpenAI’s table ↗ directly. Download the skills-to-occupations crosswalk (CSV) ↓
share_of_work_related_messages_by_onet_iwa_country_quarter.csv, US, quarter beginning July 1, 2026. Work-related, ask/do/express and state-rank tables, July 2024 to September 2026. O*NET activity names from the release data dictionary.Ask anything about jobs, skills or AI and work. You sign in with GitHub to post, and everyone can see the questions. Give a thumbs-up to the ones you also want answered. If there is data that can answer a question, I will run the analysis and add the result to this page.