Exit interview and feedback analysis

The exit interview and feedback analysis process runs four times a year, triggered at the end of each quarter when the HR manager and HR assistant sit down to make sense of written feedback collected from departing employees. The raw material arrives in a mix of emails and spreadsheets with no single system of record, and the team works through it manually: reading every response, organising the content, checking that no individual can be identified, spotting patterns, and eventually writing a summary report for leadership. From start to finish, each quarterly run takes around ten hours, meaning the process consumes roughly forty hours of HR time annually.

The work is harder than it sounds. Almost every run surfaces something that does not fit neatly into a standard theme, whether that is a response that risks identifying someone, feedback that touches on a sensitive situation requiring works council involvement, or simply ambiguous commentary that resists easy categorisation. The HR team has to make judgement calls throughout, and those judgements are not formally documented anywhere. The knowledge of what counts as a standard case versus an exception lives in the heads of two or three people.

The end product is a quarterly summary of themes and trends, delivered to leadership to inform retention and culture decisions. The problem is that the manual effort required to produce it leaves little time for the kind of deeper analysis that would make the report genuinely useful. The same exit reasons tend to surface quarter after quarter, but without time to interrogate them properly, the reports rarely generate the insight that would prompt meaningful organisational change.

An AI Workflow is the right recommendation for this process, and the reasoning comes down to the nature of the work itself. The core task, reading unstructured free-text responses, clustering them into themes, and drafting a structured summary, is exactly what a large language model does well when given clear instructions. The quarterly cadence and consistent output format make it straightforward to design a repeatable pipeline: responses go in, anonymisation is verified, themes are extracted, exceptions are flagged, and a draft report comes out. The HR manager then reviews the draft, resolves the flagged items, and approves the report before it reaches leadership. That human review gate is not a compromise on the automation; it is a deliberate design choice that keeps the workflow compliant with GDPR and works council obligations without adding meaningful overhead to a standard run.

The client has no stated technology preference, which is genuinely helpful here. A low-code platform such as Make or n8n connected to an LLM API requires minimal in-house engineering to set up and almost none to maintain on a quarterly cadence. The single IT generalist can support the initial platform configuration, and the HR manager can take ownership of the workflow logic and prompts after completing the recommended AIRIES training. This matters because a tool that depends on ongoing developer involvement would be a liability for a team of this size running a process four times a year.

The Hybrid path scored one point below AI Workflow and is worth addressing directly, because in practice the two approaches are closely related. The distinction is that a formally architected hybrid with explicit escalation routing and multiple decision-layer handoffs would be over-engineered for a quarterly, low-volume process with a two-person HR team. The AI Workflow with a single human review gate achieves the same compliance and quality outcome with far less structural complexity. The client does not need a sophisticated orchestration layer; they need something reliable that runs every three months and takes less than an hour of human time to complete.

Education as a standalone path also deserves honest consideration. At four runs per year and forty hours of total effort, there is a credible argument that simply training the HR manager and assistant to use AI tools interactively for theme analysis and report drafting would deliver a meaningful time saving with almost no build cost. The assessment scores reflect this: Education scores a six and is listed as viable. The reason the AI Workflow is preferred is that interactive AI tool use still requires the person to be present and directing the process each time, which means the knowledge remains in individual heads and the quality of output varies with whoever is doing it on a given quarter. A structured workflow captures the logic, makes it reproducible, and reduces the risk of things going wrong when personnel change.

The ROI picture is modest and the assessment is honest about that. At current cost levels, payback on even a lean build will take two to three years, and the business case rests more on insight quality and leadership decision-making value than on raw cost reduction. That framing should be communicated clearly to any stakeholders approving the investment. The process is worth automating, but the compelling reason to do it is that the current manual approach systematically produces lower-quality analysis than the organisation needs, not that it costs a great deal of money to run.

The AI Workflow and Hybrid paths score closest together, at 8 and 7 respectively, and it is worth being clear about why they are not the same thing. A Hybrid approach, as scored here, implies a more formally orchestrated system with distinct automated lanes and structured human escalation routing. For a process that runs four times a year and involves a two-person HR team, that level of architectural complexity would cost more to build, take longer to maintain, and deliver no meaningful improvement over a simpler design. The AI Workflow recommendation achieves everything the Hybrid label promises, including a human review gate before the report reaches leadership, but with far less moving parts. The single checkpoint where the HR manager reviews the AI draft, resolves flagged exceptions, and confirms anonymisation is both the compliance control and the practical boundary between what the machine handles and what the person handles. It does not need a separate escalation engine to work.

Education sits at score 6 and deserves serious consideration as a fallback rather than a lesser option. Training the HR manager and assistant to use an LLM directly for theme analysis and draft report writing would cost almost nothing to implement, could halve the manual effort with a few hours of AIRIES Level 1 and Level 3 upskilling, and carries no integration risk. If budget does not stretch to building a workflow, or if the IT generalist's capacity is genuinely unavailable during the build window, Education is a credible path that delivers real value. The reason it sits below AI Workflow is that it leaves the process dependent on individual skill and discipline each quarter rather than encoding the logic into a repeatable, documented pipeline. Personnel change is already flagged as a vulnerability in this process, and upskilling alone does not address it.

RPA, Traditional Code, and AI Agent all score 2 and are not viable here for distinct reasons. RPA is built for structured UI interactions and cannot engage meaningfully with unstructured free-text responses. A custom-coded solution would be disproportionately expensive to build and maintain for four annual runs with a single IT generalist available. An autonomous AI Agent capable of making decisions across the full process would undermine the compliance controls the organisation actually needs: GDPR and works council obligations require a named human to sign off at key points, which removes the core justification for full agent autonomy. Staying Manual scores 3, and the honest case against it is not just efficiency but strategic cost. Exit interview data that goes unanalysed properly perpetuates the exact retention problems it was meant to surface.

The first thing to do before touching any technology is to get the current process out of people's heads and onto paper. The HR manager and assistant between them hold all the working knowledge about what counts as a standard case, what triggers an escalation, and how anonymisation judgements get made in practice. A few hours spent documenting this, covering every step, every decision point, and every exception type encountered in recent quarters, creates the specification the workflow will be built against. Without it, there is a real risk of automating an undocumented process and discovering the gaps only when something goes wrong in a live run.

Once the process is documented, the next step is standardising the input. Exit interview responses currently arrive across email and spreadsheets with no consistent structure. The practical fix is a shared Google Sheet or Microsoft Excel template, agreed across the HR team, into which all responses are entered before any AI processing begins. The anonymisation rules, specifically which fields must be removed or generalised before data leaves the organisation's own systems, should be defined at this stage and applied to the input file manually by the HR assistant before the workflow is triggered. This is the GDPR control point that must be built in from the start rather than added later.

With clean, anonymised input in place, the workflow itself can be built using Make or n8n, both of which support low-code pipeline configuration with no significant programming required. The workflow reads the anonymised spreadsheet, sends the response text to an LLM API such as OpenAI's GPT-4o or Anthropic's Claude via a structured prompt, and receives back a clustered set of themes with supporting quotes and a flagged list of responses the model could not confidently categorise. The prompt design is critical here: it should instruct the model to group responses into thematic categories, note any response that appears potentially identifiable, and produce a structured JSON output that feeds directly into a report template. The IT generalist can handle the Make or n8n platform setup and API key configuration; the HR manager with AIRIES Level 3 training can own the prompt logic and workflow testing. Historical data from the last two or three quarters should be used for testing before the workflow goes anywhere near a live run.

The final structural element is the human review gate. Before any output reaches leadership, the HR manager receives the AI-generated draft report along with the list of flagged exceptions and a confirmation checklist covering anonymisation, scope, and works council relevance. This review step should take under an hour for a standard run and is the formal compliance control point documented for GDPR purposes. Once the manager approves the report, it is sent to leadership through the existing channel, whether that is email or a shared document. After two live quarterly cycles, the workflow should be stable enough that the assistant can operate it independently, with the manager's role limited to the review and approval step. Prompt refinements and exception routing adjustments should be captured in a simple changelog so the process does not re-accumulate undocumented knowledge.