Enterprise AI, week 7
Workflows And Agents By Department
- Week 7 of 12: AI integration for companies
- Promise: you map one real process before and after AI
Outcomes
What You Can Do By The End
- Describe a process as steps, inputs, outputs, owners
- Choose between workflow, assistant and agent
- Write a before and after comparison
- Place a human on purpose
- Connect the result to your week 6 pilot
Where we left off
Recap Of Week 6: Pilot Design
- A pilot has narrow scope and a named owner
- Baseline measured before anything changes
- Success measures and a time box agreed up front
- Stop rules written down before the start
- Today we choose the process the pilot will change
Start here
A Process Is Steps, Not A Department
- Input comes in, steps happen, output goes out
- Every step has an owner and a duration
- If you cannot draw it, you cannot improve it
- AI changes steps, not whole departments
Pick the least autonomy that works
Three Patterns: Workflow, Assistant, Agent
- Workflow: fixed order, model does one step
- Assistant: person asks, model answers
- Agent: goal in, system plans and uses tools
- More autonomy means more cost and more failure modes
| Pattern | Who decides steps | Typical use |
|---|
| Workflow | The designer | Classify and route email |
| Assistant | The person | Draft a reply |
| Agent | The system | Resolve a small ticket |
Department 1
Customer Service: Before And After
- Before: search three systems, write from memory
- After: model drafts with order and policy attached
- Routine answers get a light review
- Refunds and complaints stay with a person
- Measure handling time and customer satisfaction
Accountability
Case: The Chatbot That Made A Promise
- A tribunal held the airline responsible for its chatbot
- What the system says is the company speaking
- Source every answer from approved text
- Log outputs so you can reconstruct what was said
Department 2
Sales: Before And After
- Before: manual call prep and follow-up writing
- After: account brief built from CRM and public news
- Model drafts the call summary and next steps
- Rep edits and owns every outgoing message
- Measure prep time and follow-up speed
Department 3
Finance: Before And After
- Before: type invoice data, match by hand
- After: model extracts, rules match, people fix exceptions
- Approval limits do not change
- Month end variance commentary drafted for review
- Measure cost per invoice and error rate
Department 4
HR: Before And After
- Before: answer the same policy questions weekly
- After: assistant answers from the approved handbook
- Job descriptions and interview guides drafted
- Recruitment is a high risk use under the EU AI Act
- People decide on hiring, pay and dismissal
Department 5
Legal: Before And After
- Before: read the full contract line by line
- After: model flags clauses and compares to playbook
- Every flag shows the source passage
- Lawyer verifies and decides
- Confidential text follows week 5 data rules
Department 6
Software: Before And After
- Before: boilerplate, doc search, tests last
- After: assistant drafts code, tests and pull request summary
- Agents can open pull requests for small tickets
- Review, tests and security scans remain gates
- Measure cycle time and defect rate, not lines of code
Method
Map The Process On One Page
- List steps, inputs, outputs, owners, minutes
- Tag each step: rule, language or judgment
- Language steps are the best model candidates
- Judgment steps with impact keep a human decision
Step | Input | Output | Owner | Min | Type
1 Read ticket | email | category | agent | 2 | language
2 Find order | id | record | agent | 3 | rule
3 Write reply | record | draft | agent | 6 | language
4 Approve refund | draft | decision | lead | 2 | judgment
Design choice
Place The Human On Purpose
- Approve every output before release
- Sample outputs after the fact
- Review only cases flagged as uncertain
- Choose by the cost of a mistake
Prioritise
Choosing The First Process
- High volume and stable pattern
- Text as main input
- Measurable baseline
- Controllable cost of error
- Data you are allowed to use
| Candidate | Volume | Repeatable | Risk | Data ready |
|---|
| Ticket triage | High | Yes | Low | Yes |
| Invoice matching | High | Yes | Medium | Yes |
| Contract review | Medium | Partly | High | Partly |
| Candidate screening | High | Partly | High | No |
What to remember
Recap
- Work at the level of processes and steps
- Use the least autonomy that solves the problem
- Language steps fit models, judgment steps keep humans
- Each department keeps a named accountable person
- Record time per step now for the week 9 cost case
Week 8
Next Week: Change Management And AI Literacy
- Good workflows fail when people do not adopt them
- Adoption curve, training plans, role redesign
- Champions network and handling resistance
Your turn
Questions And Exercise
- Map one weekly process before and after AI
- Mark language, rule and judgment steps
- Place the human and say why
- Bring the page to week 8