Most consultancies and agencies already ask clients for feedback. The problem is not usually collecting it. The problem is reading it properly.
A partner sends out an NPS survey. A project manager adds a Typeform link to a completion email. A few dozen responses arrive over the quarter. Scores are skimmed, one or two striking comments are mentioned in a meeting, and the spreadsheet is quietly left until next time.
That is understandable. Reading 80 free-text comments properly takes time. Someone has to group similar issues, separate one-off grumbles from repeated patterns, and work out whether the same theme is appearing across several clients or just one difficult project.
But those comments often contain the early signs of a service issue. Slow turnaround. Unclear billing. Weak handover after a project ends. Clients rarely leave without warning; they often tell you in small ways first.
An example AI workflow: quarterly feedback theme extraction
This is an example workflow, not a real client case study. The idea is simple: use AI to do the first pass over the free-text survey comments, then let a human decide what matters.
A small script pulls the latest survey responses from a Google Form, Typeform export, Google Sheet or CSV folder. Each row might include the client name, project or service line, score, and free-text comments.
The assistant then clusters the comments into practical themes that make sense for the business. For a consultancy or agency, those categories might include communication, pricing clarity, speed, staff attentiveness, handover, reporting, or scope management.
For each theme, it tags the sentiment as positive, negative or neutral. It then produces a one-page quarterly digest showing:
- the top recurring positive and negative themes;
- the top three negative themes ranked by frequency and severity;
- a small number of representative anonymised quotes;
- which clients, project types or service lines the comments relate to;
- a simple trend against the previous quarter.
The output can be a PDF, a markdown report, or a page in the firm’s internal knowledge base. The important point is that it is short enough for a partner or operations lead to actually read.
What this changes
The value is not that AI “understands your clients” in some magical way. It does not. The value is that it can do the boring first pass quickly and consistently.
Instead of relying on whoever happens to have time to scan the spreadsheet, the firm gets a repeatable quarterly process. If communication is mentioned negatively by ten clients across three teams, that is harder to miss. If billing comments improve after a process change, that can be seen too.
It also helps separate noise from pattern. One blunt comment may still matter, but five similar comments across different clients usually deserve more attention.
Where the human still matters
The assistant should not decide what to fix. It should flag patterns.
A partner or operations lead reviews the digest before it is shared more widely or used to justify a process change. That review matters because survey comments can be subjective, incomplete, or tied to context the AI will not know.
There should also be a clear rule around anonymity. Raw survey data should only be visible to the people who already have permission to see it. Quotes used in the wider digest should be anonymised unless there is a good reason not to. Staff-specific conclusions need particular care; the tool should not become an automated blame machine.
What you need to make it work
You do not need a huge AI platform for this. A sensible first version needs:
- read access to the survey export or spreadsheet;
- a short list of theme categories relevant to the business;
- an LLM to cluster and summarise the comments;
- a basic template for the quarterly digest;
- a human review step before anything is treated as a conclusion.
For many firms, this could save three to five hours per quarter of manual reading and tallying. More importantly, it can bring service issues to the surface before they become client-loss issues.
A practical first step
If you already collect feedback, open the last quarter’s spreadsheet and ask one question: “What recurring problem would we be embarrassed to admit we missed?”
If the answer is buried in the comments, this is a good candidate for a small, low-risk AI workflow.
If you would like help designing your own version of this — with the right data access, privacy rules and human review points — book a short call with Stuart Cole Consulting and we can map the workflow before you build anything.
