Workflows · Example project
Research workflows, with a person in charge
Proposal An example project for an agency, with invented data
Each workflow collects the data, removes personal data, and drafts findings. Then it stops for a researcher. Nothing is stored or sent until a person has accepted, edited, or rejected every draft. The decision report shows how each decision was made.
Try it
- Choose a workflow, then select Run the workflow.
- At the human check, open the evidence and decide on each draft.
- Select Approve and store, then Decision report. Select Change decisions to reopen the review and update the results.
- In Affinity map and Prioritise, drag sticky notes and cards, or use their Move to lists.
What these workflows cannot tell you
- Why people behave as they do. The workflows find patterns. Interviews and workshops explain them.
- How many customers have a problem. Counts show what people said or did, and the people who write in are not everyone.
- Where an artificial intelligence (AI) assistant will rank a brand next month. Answers change between runs, so the reports show trends.
Safeguards
What keeps every workflow safe to run on client data.
- Personal data removed firstNames, emails, phone numbers, and order numbers are removed before any AI reads the text.
- Restricted recordings stay in the buildingRaw recordings are Restricted, so only a self-hosted model transcribes them. Findings are saved as Confidential memory entries.
- Content is data, never instructionsText that tries to give the AI an instruction is held back, shown to the researcher, and never acted on.
- Two sources or it is not an insightA draft with one source cannot be accepted. It can become a question for the next round of fieldwork.
- Low confidence means evidence firstA draft under 70% confidence cannot be accepted until its evidence is opened.
- Every rejection has a reasonThe reason is kept, so prompts and the codebook improve from real decisions.
- Nothing stored until approvedThe store step waits for a person. The audit log records who decided, and when.
Take it to the next level
What to add once the first workflows prove their hours.
- Measure every runLog the hours and the edits for each step, so the time saved and the AI accuracy are measured, not guessed.
- Learn from rejectionsReview the rejection reasons each month and update the codebook, then test it on past projects.
- Compare two modelsRun the tagging in two AI models. Where they disagree, a person checks first.
- Connect the workflowsFeed the questions from feedback into the answer engine audit, and the friction points into the journey map.
An example project. The agency, its clients, and every figure, quote, and answer here are invented.