WORKFLOWS / PROCESS IMPROVEMENT
Find the process problems behind repeated requests.
Use recurring questions, corrections, and requests to spot where a process is creating avoidable work, then choose a small improvement to test.
WHAT THIS WORKFLOW PRODUCES
A specific improvement worth testing.
A short finding with supporting cases, a possible explanation, and a practical change to try. The aim is to identify avoidable work without treating every question or exception as a problem. In the example below, one request counts once even if it produces several messages, and the example suggests a cause to investigate rather than proving one.
VENDOR REQUESTS ARRIVE WITHOUT AN APPROVAL OWNER
ILLUSTRATIVE FINDING · SAMPLE DATA, NEVER RESEARCH RESULTS
THE OPERATING METHOD
Understand the workflow without AI.
Look at a defined set of work, group the repeated problems, and check the process itself before proposing a fix. Keep enough detail to return to the original cases. AI is introduced only after the method makes sense on its own.
Choose the work to review
Define the request source, period, and types included.
YOUGroup similar cases
Identify the repeated question, correction, delay, or exception.
YOUCheck the process
Compare the cases with the form, instruction, or handoff as it stands.
YOUChoose a small test
Propose a change that addresses a specific possible cause.
YOUReview the result
Measure whether the problem changed and whether the fix created new work.
YOULook for a process explanation before blaming the people using it. A question may be reasonable because the situation is genuinely unusual.
WHERE THE FACTS LIVE
Use the systems where the work already lives.
Use an authorized operational queue or defined request set. This workflow does not require monitoring private messages or collecting everything employees write. Named tools link to their pages in the library.
REQUESTS AND OUTCOMES
The original request, follow-ups, resolution, and dates.
REVIEW AND TEST
Confirmed themes, counts, examples, and the change being tested.
Google Sheets
Excel
An existing project board
THREE WAYS TO RUN IT
Choose how much machine.
AI is useful for comparing meaning across differently worded cases. Ordinary spreadsheet calculations are better for counting them. Use both where they help. The tools named are examples. Use what your organization already runs, provided the connection is supported and approved.
LEVEL 1
Find patterns in a defined sample.
Give AI an approved export containing one record per request. It proposes a few themes, points to the supporting cases, and separates possible explanations from what the records show.
Use this version when you suspect the same issue keeps returning but do not yet know whether it is a meaningful pattern.
WHAT THIS VERSION REDUCES
Reading similar requests repeatedly and manually sorting different descriptions of the same issue.
WHAT A PERSON STILL DOES
Whether the grouping makes sense and which patterns deserve investigation.
MAIN LIMITATION
A convenient sample can overrepresent noisy cases. Findings describe the records reviewed rather than every request in the company.
Build this version, step by step 4 steps, one action each. Prompts carry a Copy button.
Pick one set of cases to look at.
ONE QUEUE, ONE PERIOD
For each case keep a stable ID, the date, the relevant text, the follow-up, and the outcome if known. Include ordinary cases as well as difficult ones. Take out names and sensitive details that the analysis does not need.
Ask for a handful of themes.
Group these requests by the repeated operational issue. Propose up to five useful themes and list the supporting case IDs. Separate observed problems from possible causes. Keep unusual or unclear cases separate. Do not invent frequency, time lost, or missing outcomes.
Open the cases and count them yourself.
THE CHECK
- A few cases from each theme have been opened and the grouping holds up.
- The count is of unique request IDs, done in the spreadsheet.
Several comments on one request are still one request.
Choose one pattern to investigate.
PATTERN AND QUESTION
Pick the pattern with enough evidence or enough consequence to be worth the time. Write down the question you will ask the process owner before you propose any fix.
LEVEL 2
Compare the pattern with the process.
Give a connected assistant access to the reviewed cases and the relevant process documents. It investigates whether the issue relates to missing guidance, an awkward step, conflicting instructions, or a genuine exception.
Use this version when you have a recurring pattern and need to understand what should change.
WHAT THIS VERSION REDUCES
Comparing several cases with different parts of the process documentation.
WHAT A PERSON STILL DOES
Whether the documented process matches reality and whether the proposed change is sensible.
MAIN LIMITATION
Written records alone rarely establish why something happened. Ask the people doing the work before calling a hypothesis the root cause.
TOOL OPTIONS
An AI assistant, the existing request records, and a spreadsheet are sufficient for many teams. If the organization already uses Dovetail, its tagging and source-linked highlights can help organize qualitative themes. Buying a separate research platform is not a requirement for this workflow.
Build this version, step by step 5 steps, one action each. Prompts carry a Copy button.
Give the AI the cases you have already checked.
CONFIRMED THEME
The cases that show the pattern, plus a few similar cases where the process worked fine. Keep the IDs or links so every claim can be traced back.
Give it the process as it is written today.
CURRENT FORM AND GUIDANCE
Through an approved document connection or by uploading the files. Make sure it has the actual form or procedure text rather than a title or a search result.
Ask for more than one explanation.
CAUSE TO INVESTIGATE
Ask what evidence supports each explanation and what would rule it out. A missing field, unclear ownership, or an unusual request can all produce the same kind of follow-up message.
Check with the process owner.
THE CHECK
- You know which version of the process people are using.
- You know whether they can open it.
- You know about the exceptions the records do not explain.
Agree one change and how you will measure it.
CHANGE AND MEASURE
Ask the AI to draft the smallest practical change, a way to measure the original problem, and a way to spot any new work the change creates. Agree who tests it and when you will look at the result.
CAPABILITIES CHECKED AGAINST OFFICIAL DOCUMENTATION
Claude's Google Workspace connection documents access to reference material, subject to authorization and file support. Source, checked September 6, 2026.
Dovetail documents source-based qualitative theme organization with tags; it does not establish causation or remove the need to check AI classifications. Source, checked September 6, 2026.
LEVEL 3
Keep a recurring process review running.
A configured workflow brings relevant new or changed requests into a simple review table. AI proposes classifications and investigates selected patterns against approved process documents. Counts and comparisons come from the records, and a person approves any process change.
Use this version when a shared service, administrative queue, or recurring intake process produces enough repeated work to justify ongoing review.
WHAT THIS VERSION REDUCES
Repeated sorting and finding evidence for an emerging operational issue.
WHAT A PERSON STILL DOES
Theme definitions, significance, investigation priorities, and changes to the process.
MAIN LIMITATION
Classification mistakes and source changes affect the findings. Keep reviewable examples, an uncertain category, and an explicit record of what was included.
TOOL OPTIONS
Make can pull Zendesk tickets and coordinate the AI and spreadsheet steps; its agent product is in open beta, and n8n is another option. Use whichever source the team already has; this page is not recommending a new help desk. Check the counts against a sample a person has reviewed before using them for decisions.
IMPROVE THE PROCESS, NEVER SCORE PEOPLE
Use this to improve a process. Do not score employees, infer their motives, or treat the number of questions a person asks as a performance measure.
Build this version, step by step 6 steps, one action each. Prompts carry a Copy button.
Choose the queue and the boundary.
SELECTED REQUESTS
One source and one type of request to begin with. If the work already arrives through Asana or a form, keep it there. Zendesk is the example for teams that already handle requests as tickets.
Pull each case with enough context.
REQUEST AND RESOLUTION
In Make, use the ticket and comment operations to get the request, the follow-up, and the resolution where it exists. Titles on their own are not enough to analyze.
Keep a small review table.
ONE ROW PER CASE
Each row holds the source ID, the date, the current theme, a short reason, the outcome, and a link. When a case changes, update its row. Leave sensitive text in the original system, and process only the cases that changed since the last run.
Give the agent a fixed set of choices.
KNOWN, NEW, OR UNCERTAIN
It sorts each case into the agreed themes, may look at the allowed case context, and may propose a new theme when nothing fits. Uncertain cases go to review. It cannot rewrite the theme definitions on its own.
Count with formulas, investigate with the agent.
VERIFIED COUNTS
The unique-case totals come from spreadsheet formulas or plain workflow logic. At the review interval you chose, the agent looks at the chosen patterns and the relevant process documents and comes back with the supporting cases, a possible cause, and one small test.
Approve the test and measure it.
THE CHECK
- The process owner has reviewed the proposal and made the approved change.
- The same measure is compared before and after.
- The request source, period, and inclusion rules are written down.
If the volume or mix of requests changed, say so before crediting the fix.
CAPABILITIES CHECKED AGAINST OFFICIAL DOCUMENTATION
Make's Zendesk integration documents supported ticket and comment operations; the specific queue and permissions need account-level validation. Source, checked September 6, 2026.
Make AI Agents can use tools inside a scenario; the agent product is documented as open beta. Source, checked September 6, 2026.
n8n documents a tool-using agent node as an alternative orchestration approach. Source, checked September 6, 2026.
WORKS EVEN BETTER WITH
The right fix might be a clearer procedure, an easier way to find an answer, or a small tool. Choose it based on the problem the requests revealed.
Standard Operating Procedure → | Answer recurring internal questions from approved information → | When is a small AI tool worth building? →