Workflows And Agents By Department
You learn how AI changes real processes in customer service, finance, HR, legal, sales and software, and how to describe a process before and after AI. You leave with a workflow map you can bring to your own team.
You Will Be Able To
- Describe a department process as a sequence of steps, inputs, outputs and owners
- Distinguish a fixed workflow, an assistant and an agent, and pick the right one
- Write a before and after comparison for one process in your own department
- Mark the steps in a process where a human must review or approve
- Judge which department use cases fit your pilot design from week 6
Where We Are After Week 6
In week 6 you designed a pilot: a narrow scope, a baseline, success measures, a time box and a named owner. This week we fill that design with real work. A pilot is only as good as the process it changes, so we look at how six departments actually do their work and where AI fits into each step.
A process is a repeatable sequence of steps that turns an input into an output. A ticket comes in and an answer goes out. An invoice arrives and a payment is booked. If you cannot draw the steps on one page, you cannot yet decide where AI belongs.
Workflow, Assistant Or Agent
Three patterns cover most enterprise use. A fixed workflow runs the same steps in the same order every time, and a model handles one step inside it, for example classifying an email. An assistant answers a person who asks, such as a drafting helper inside the mail client. An agent is a system that receives a goal, chooses its own steps, calls tools such as search or a database, and checks its progress until it finishes or asks for help.
Start with the simplest pattern that works. A fixed workflow is easiest to test, cheapest to run and easiest to explain to an auditor. Move to an agent only when the steps vary so much from case to case that a fixed order would break. Agents add flexibility and also add cost, delay and new ways to fail.
- Workflow: fixed steps, a model does one of them
- Assistant: a person asks, the model answers, the person decides
- Agent: a goal goes in, the system plans steps and uses tools
- Rule of thumb: pick the pattern with the least autonomy that solves the problem
Customer Service And Sales
Customer service before AI: an agent reads a ticket, searches three systems for the order and the policy, writes a reply from memory, and pastes it. After AI: a model classifies the ticket and fetches the order record and the policy text, then drafts a reply in the company tone. The human agent reads the draft, fixes it and sends it. Routine questions with a clear policy answer can go out with a light review. Refunds, complaints and anything with legal tone stay with a person.
Sales before AI: a representative spends the morning reading news and CRM notes to prepare a call, and the afternoon writing follow-up emails. After AI: a workflow gathers the account history and recent public news into a one page brief, and after the call a model drafts the summary and next steps for the representative to edit. The representative still owns the relationship and every message that leaves the company.
| Department | Before AI | After AI | Human role |
|---|---|---|---|
| Customer service | Agent searches systems and writes reply | Model drafts reply with order and policy | Reviews and sends |
| Finance | Staff types invoice data | Model extracts fields, rules match order | Handles mismatches |
| HR | Recruiter answers policy questions | Assistant answers from handbook | Decides on people |
| Legal | Lawyer reads full contract | Model flags clauses against playbook | Verifies and decides |
| Sales | Rep prepares call by hand | Workflow builds account brief | Owns the message |
| Software | Developer writes tests last | Assistant drafts code and tests | Reviews and merges |
Each row keeps a named person accountable for the result.
Finance And HR
Finance before AI: accounts payable staff type invoice data into the ledger, match it to a purchase order and chase exceptions by email. After AI: a model extracts the fields from the invoice, a rule engine matches them to the order, and only mismatches go to a person with a short explanation of what differs. Payment approval limits stay exactly as they were. Month end gains a drafted variance commentary that the controller reviews and signs.
HR before AI: a recruiter reads hundreds of applications and answers the same policy questions every week. After AI: an assistant answers policy questions from the approved handbook and cites the page, and a drafting tool prepares job descriptions and interview guides. Hiring, promotion and dismissal decisions carry high risk. Under the EU AI Act, systems used for recruitment and worker management fall in a high risk group with extra duties, so keep a person in charge of every decision and involve your legal team early.
- Invoice arrives
- Model extracts fields
- Rules match purchase order
- Mismatch goes to person
- Payment approved by limit
A model reads, rules decide the match, and a person handles every exception.
Legal And Software
Legal before AI: a lawyer reads a thirty page supplier contract line by line to find liability, termination and data clauses. After AI: a model highlights those clauses, compares them to the company playbook and lists deviations with the source passage next to each. The lawyer verifies every highlight against the contract and decides. Models can state wrong things confidently, so legal work always requires a human check of the cited text. Confidential documents also need the data rules you set in week 5, including KVKK and GDPR duties.
Software before AI: a developer writes boilerplate, searches documentation and writes tests last. After AI: a coding assistant suggests code in the editor, drafts unit tests and summarises a pull request for the reviewer. Agents can now take a small ticket, change several files and open a pull request. The team keeps code review, automated tests and security scanning as gates, because generated code can contain defects and unsafe patterns like any other code.
Mapping A Process And Placing A Human
Use one page for each process. List the steps, and for each step write the input, the output, the owner and the time it takes today. Then mark each step as rule based, language based or judgment based. Language based steps such as reading, summarising, drafting and classifying are the best candidates for a model. Judgment steps with legal, financial or personal impact keep a human decision.
Then place the human on purpose. There are three common positions: the human approves every output before it leaves, the human samples a share of outputs after the fact, or the human handles only the cases the system flags as uncertain. Choose the position by the cost of a mistake. A wrong marketing draft costs little. A wrong payment or a wrong hiring message costs a lot. Write the choice in the process map so reviewers know their role.
Illustrative scale only. Higher cost means a person approves before release.
Choosing What To Automate First
Pick processes that run often, follow a clear pattern, have text as the main material, and have a measurable baseline. Customer service triage, invoice matching and contract review all pass this test. Avoid starting with a process that nobody has documented, that depends on one expert's intuition, or that touches personal data without a legal basis.
Score each candidate on volume, repeatability, risk and data readiness from week 3, then choose one or two for the pilot you designed in week 6. Keep the before and after comparison as a living document. In week 9 you will turn the same comparison into a cost and benefit case, so record the time per step now.
Keep the first version of any redesigned process small. Run it beside the old process for a few weeks, compare results case by case, and let the people who do the work say what feels wrong. They notice problems that metrics miss, such as a reply that is correct but sounds cold, or a summary that skips a detail the customer cares about. Fix those points before you widen the scope, and write each fix into the process map so the next team inherits it.
- High volume and a stable pattern
- Text or documents as the main input
- A baseline you can measure today
- Low or controllable cost of error
- Data you are allowed to use
Key Terms
- Process
- A repeatable sequence of steps that turns an input into an output.
- Workflow
- A fixed sequence of steps in which a model handles one or more of the steps.
- Assistant
- A model-based helper that answers a person who asks and leaves the decision to that person.
- Agent
- A system that takes a goal, chooses its own steps and uses tools until it finishes or asks for help.
- Tool call
- A request from a model to an outside function such as search, a database query or a calculator.
- Human in the loop
- A design in which a person reviews or approves model output before it takes effect.
- Exception handling
- The path a process takes when a case does not match the normal pattern.
- Baseline
- The measured time, cost or quality of a process before AI is introduced.
- High risk system
- A category in the EU AI Act for uses such as recruitment that carry extra legal duties.
Exercise: Map One Process Before And After
- Choose one process from your own department that runs at least weekly.
- Write its steps on one page with input, output, owner and minutes per step.
- Mark each step as rule based, language based or judgment based.
- Redraw the process with AI in the language based steps and place the human on purpose.
- Write two sentences on what could go wrong at each AI step and how you would notice.
- Share the page with a colleague and ask which step they would trust least.
Further Reading
- Building Effective Agents, Anthropic engineering blog
- Regulation (EU) 2024/1689, the Artificial Intelligence Act, Official Journal of the European Union
- Prediction Machines, Ajay Agrawal, Joshua Gans and Avi Goldfarb, Harvard Business Review Press
- Reengineering the Corporation, Michael Hammer and James Champy, HarperBusiness