FAST-08 Lab Guide: AI for Workflow Improvement and Process Automation

Provider: 7formations | Course: FAST-08 | Instructor: Wadood Chaudhary | Five labs, one per session (45 to 75 minutes each) Audience: trainees (steps, worksheets) and instructors (setup, answer keys, timing). Instructor-only notes are marked Instructor.

All companies, people, requests, invoices, and data in this guide are fictional. Builds run only in the 7formations training sandbox (Microsoft track: Power Automate with AI Builder or Copilot Studio; Zapier track: Zapier with AI by Zapier). Nothing in the labs connects to an employer's production systems. Product menus change often; the instructor confirms current screens before each cohort.

0. Lab environment and safety rules (read before Lab 1)

Sandbox. Each trainee gets a sandbox environment (Microsoft) or a training workspace (Zapier) with a test mailbox, a test form, a tracking list or table, and a test team channel. Sandboxes are spending-capped and disabled at cohort end. Never connect a personal or work production account.

Data.

• Fictional data only. Never paste or route a Massachusetts resident's name with an SSN, driver's license, or financial account or card number (G.L. c. 93H, 201 CMR 17.00), PHI, or confidential business data.
• Connections use the sandbox test accounts with the minimum access each flow needs (least privilege). Never share passwords.
• Treat every input as untrusted: request emails and invoices can contain text that tries to instruct the AI step.

The rule for every design. A person approves every step that affects customers, money, or personal data.

Fictional organizations

• Merrimack Freight & Logistics, a 70-person regional trucking and warehousing company in Lowell, MA. Labs 1 to 3 use its customer request process.
• Back Bay Accounting Partners, a 25-person accounting firm in Boston. Lab 4 uses its accounts-payable invoice intake (fictional vendors).

Completion. Submit W1 to W5, the working prototype (exported or screenshots), and the improvement proposal. Labs are scored Meets / Not yet; the proposal must score Meets on every rubric criterion (see projects.md). One resubmission within the course window.

Lab 1 (Session 1): Map and baseline a process

Goal. Produce W1 (current-state map with swimlanes, handoffs, waits, and error points) and W2 (baseline time study).

Process description (fictional). Requests reach Merrimack Freight's shared inbox by email (about 60%), web form (30%), and phone (10%), about 45 a day. A coordinator reads each request, re-types it into the tracking sheet, decides the type (quote, delivery status, damage claim, address change, other), and forwards it. Quotes wait for the pricing analyst (often 1 to 2 days). Status questions are answered by the coordinator from the transportation management system (TMS). Claims go to the claims clerk with photos. Address changes go to billing.

Steps

1. Map (20 min). Sticky notes (in person) or whiteboard (virtual). Swimlanes: Customer, Coordinator, Pricing analyst, Claims clerk, Billing, TMS. One note per step. Mark handoffs (arrows between lanes), waits (W), and error points (red dot).
2. Estimate (5 min). Before looking at the data, estimate touch time and lead time for one quote request. Record it.
3. Measure (15 min). Use the sample time log below to compute touch time, wait time, and lead time.
4. Find the waste (10 min). Label re-keying, waits, rework (misroutes), and send-backs.

Sample time log: 5 quote requests (fictional)

Deliverables. W1 map photo or export; W2 table: per request, touch time (coordinator + analyst minutes), lead time (arrival to quote sent), wait time (lead minus touch); averages; your estimate versus the measured value.

Lab 2 (Session 2): Future-state map with checkpoints

Goal. Produce W3: a future-state map that removes waste, adds AI steps (classify, extract, draft, summarize) where they fit, and places a human checkpoint at every step affecting customers, money, or personal data, with the data each step touches.

Steps

1. Redesign (20 min). Start from W1. Remove re-keying. Add AI steps only where the task is frequent and checkable.
2. Checkpoints (10 min). Mark each step that affects customers, money, or personal data with a person icon and the role that approves.
3. Data and permissions (10 min). For each automated step, list the data it reads and writes and the minimum connector access needed.
4. Exceptions (5 min). Add the path for "other," low-confidence labels, and failures.

W3 control table (fill one row per automated step)

Lab 3 (Session 3): Request routing flow (guided build 1)

Goal. Build and test a working flow in the sandbox: new request → AI classify → log to tracking list → route by label, with a review queue for "other." Complete W4 (build checklist) for this flow.

Build steps (tool-neutral)

1. Trigger: new email in the sandbox requests inbox (or new form response).
2. AI step (classify): prompt from activity F08-S3-A1 (allowed labels, definitions, "other" fallback, label-only output, injection guard). Define a single output field: label.
3. Action: add a row to the tracking list: received time, sender domain only, subject, label, status "New."
4. Condition / paths: quote → pricing channel; status → coordinator queue; claim → claims clerk; address_change → billing review; other → review queue.
5. Notification: post the request ID and label (no personal data) to the right team channel.
6. Log: run time, label, routing result.

Test set (fictional): run all ten

W4 build checklist

• ☐ Trigger fires on the sandbox inbox only.
• ☐ Classify prompt lists labels, definitions, and fallback; output is the label only.
• ☐ Connection access limited to one inbox, one list, one channel.
• ☐ No personal data posted to channels.
• ☐ "Other" and two-request emails go to the review queue.
• ☐ Every run logged.
• ☐ Ten test cases run; failures recorded and fixed.

Lab 4 (Session 4): Invoice extraction with approval (guided build 2)

Goal. Build a flow for Back Bay Accounting Partners: new invoice email with a PDF → AI extract fields → duplicate check → approval (sequential above $5,000) → write approved rows to the AP table; with error handling and a run log.

Build steps

1. Trigger: new email with attachment in the sandbox AP inbox.
2. AI step (extract): prompt from activity F08-S4-A1; output fields vendor_name, invoice_number, invoice_date, due_date, subtotal, tax, total, po_number. No bank details.
3. Error scope: if the AI step fails or returns no total, set status "Needs review," log the error, and notify the flow owner. Stop.
4. Duplicate check: if invoice_number plus vendor already exists, stop and notify the AP clerk.
5. Approval: AP clerk sees the extracted fields and a link to the PDF; options Approve or Reject with comment. If total is over $5,000, then the controller approves second (sequential).
6. On approve: write the row with approver names and times. On reject: status "Correction needed," notify the clerk with the comment.

Test invoices (fictional): text layer of each PDF

W4 build checklist additions: error scope tested by forcing a failure (disconnect the AI step or send a non-invoice attachment); approver sees the PDF link; sequential approval over $5,000; duplicate check; log stores approver name and time but no bank or card data.

Lab 5 (Session 5): Test log and improvement proposal

Goal. Produce W5 (break-it test log of at least 15 tests across both flows) and the one-page improvement proposal (template in projects.md).

Steps

1. Write tests (15 min). At least three in each category: messy input, security or injection, data protection, volume or limits, approval paths.
2. Run and log (25 min). Record expected result, actual result, pass or fail, fix, and retest.
3. Measure (10 min). Time ten prototype runs including human review. Compare with the W2 baseline.
4. Draft the proposal (20 min) and pitch it (three minutes).

W5 test log

Savings calculation (example). Baseline 9.0 min per request; prototype 3.0 min including review; 45 requests per day: (9.0 - 3.0) x 45 = 270 min per day. At the BLS private-industry average employer cost of about $46.89 per hour worked (June 2026), that is about $211 per day of capacity, before tool and support costs. State it as capacity freed, not headcount.