Personal Agent: an AI agent for employee feedback

Spradley’s Personal Agent answers leaders’ questions about what employees said, cites the evidence for each answer and drafts a follow-up when it is thin.

Miguel Cabral Oliveira

Co-founder & CTO

Nov 4, 2026

Personal Agent is Spradley’s AI agent for employee feedback. A leader asks a question in plain language, such as “why is the night shift unhappy with scheduling?”, and it answers from the interview reports that leader is allowed to see, with every point traced to the area, claim or quote it rests on. When the reports can’t answer, it says so and drafts a short follow-up interview for the leader to send.

Why does a static report stop working?

A report answers the questions its author thought of. The leader reading it has others. Does this affect the depot or only head office? Was it there last round too? Is the workload complaint the same thing as the overtime complaint? A finished PDF has no way to answer those, so they go to HR as an email, wait a week, or never get asked.

Gallup’s guide to employee surveys warns that asking for feedback without acting on it “can backfire”. Often the gap between results and action is made of questions nobody had time to answer. That is the gap Personal Agent fills. Each leader gets their own, and it works on what their people said.

What can you ask Personal Agent about your employee feedback?

Anything about the patterns in what people said. Here are four questions a leader typically asks after a round of interviews, and the shape of the answer that comes back. The setting is invented: a distribution company with three depots and a head office, after its autumn interviews.

The leader asks	What comes back
“Why are the depots unhappy with the new rota?”	The causes people described, in order, with the area and claim each one comes from. Where departments lean differently on the same area, both sides. Quotes only from those the report already published.
“What changed between the spring and autumn rounds?”	What persists, what is new, and what no longer appears in the published results. Something that dropped out is never reported as solved.
“Is the overtime problem the same as the workload problem?”	The cause chains people stated, followed from one claim to the next, so you can see whether two complaints share a root. An inferred link is marked as the agent’s own.
“Help me prepare managers to talk to their teams.”	A conversation guide: what to share, open questions to ask, what you may hear, and what never to do. It quotes no one.

Two things hold across all four. Sizes come in words such as “several” or “most”, never as a count of people. And the agent says what the research does not show, as plainly as what it does.

An invented example: the rota question

Say the operations director opens the autumn report and asks why the depots are unhappy with the rota. Personal Agent reads the report’s areas and goes to the one on scheduling. It sets out what people described there, and where depots lean differently, it shows both sides and any claims that explain the gap. Each sentence of the answer carries its reference, so the director can open the report at the exact area.

Then she asks whether drivers feel the same. If the interviews didn’t cover drivers as a separate group, the agent says the reports cannot answer that. It offers to draft a follow-up interview for the driver teams instead of guessing.

How does it show where each answer came from?

Every factual sentence points at its source: a report, an area of that report, a claim in that area, or a quote the report published. After the answer is written, a check runs over it. A quote that doesn’t match a published one, or a figure that didn’t come from what the agent read, is removed before the leader sees it.

For questions that need counting or comparing, the agent writes a small program and runs it over the reports in a sandbox with no network and no database. It can group claims by kind, follow cause links, or set two rounds side by side. Only the result comes back into the conversation, and that result goes through the same check.

When a report didn’t go deep enough on something, the agent can ask Spradley’s analysis engine to read the answers again for that question. That takes minutes. The new claims are judged for privacy before they reach the agent, and they arrive in a later turn of the chat.

What happens when the evidence isn’t there?

Personal Agent drafts a follow-up interview. It hands the question to Spradley’s interview builder, which writes a title, a short brief and one to three open, neutral themes while the leader watches. The leader can ask for changes in plain words: drop a theme, soften the angle, shorten the introduction.

Nothing goes out on its own. The leader sends it from the draft, after one confirmation, to the whole company or to named departments, by email. A follow-up must reach at least five people. When the answers come in, the new report joins the others and the agent reads it like any other.

Can it draft the next step?

Yes. For anything longer than a chat answer, the agent starts a write-up and builds it in the background. There are six kinds:

Readout: where to act first and why, what to protect, the cost of waiting, and one or two commitments.
Deep dive: one area followed across departments and rounds.
Comparison: two rounds or reports side by side.
Tensions: claims that pull against each other, and departments that lean apart.
Root cause: the causes that sit upstream of several problems.
Conversation guide: for a manager’s talk with their team.

Every factual sentence in a write-up carries its reference and is checked before you read it. You can ask for one section to be rewritten, then accept or reject the change. Download it as PowerPoint, PDF or Word, or open it in Google Docs. A chat answer opens in Google Docs too.

From a write-up you can save a commitment, such as “publish the depot rota two weeks ahead”. Spradley emails a reminder at four weeks and again at eight, until you mark it done.

What can Personal Agent see, and what can it never see?

It reads the company’s completed interview reports and nothing below them. It never sees one person’s answer. What it gets is each report’s areas, the claims under them, the quotes the report already published, how departments lean, and released ratings.

Groups under five. An area held by fewer than five people reads “size not released”. A follow-up interview needs five people or more.
Who sees what. Owners and managers use it. An owner’s agent reads every report; a manager’s reads only the reports of the departments they were given.
Names. A sentence naming an employee is cut from the answer.
Hard lines. It will not help find out who said something, support a decision about an individual, choose which team to cut, or monitor loyalty, union activity or intent to leave. It declines in a sentence and offers the nearest useful thing, such as the pattern across a group large enough.
Your files. You can add a PDF, CSV or text file to a question, from your computer or Google Drive. The agent can quote it, but never uses it to narrow a group or match people to findings.

These lines sit in code, in the agent’s instructions, and in Spradley’s rules of use. Nothing said in a chat changes them. Our post on anonymity thresholds explains why five is the floor.

How is it different from a chatbot over a CSV export?

Paste survey comments into a general chatbot and it will answer. Three things go missing.

Citations you can check. A general chatbot paraphrases, and you can’t tell a quote it found from one it invented. Personal Agent cites the claim behind each sentence and removes quotes and figures it can’t match.

A minimum group size. A spreadsheet of comments has every row. Ask for the night shift in one small depot and the chatbot will happily answer about three people. Personal Agent reads reports where numbers and ratings for fewer than five people were never shown, and it won’t narrow a group with a file you add.

A way to ask again. A chatbot over an export can only reshuffle what is there. When the evidence is missing, Personal Agent drafts the follow-up interview that would collect it, and the answers come back as a new report.

There is also the matter of where the data goes. A CSV pasted into a personal chatbot account has left your organisation’s control. Spradley keeps its database in Frankfurt and runs its AI in Google Cloud EU regions.

Can you use it inside Claude, ChatGPT or Copilot?

Yes, read-only. The Spradley MCP server lets an owner or manager read the same reports from Claude, ChatGPT or Microsoft 365 Copilot. The owner turns it on for the company. The same rules apply: a manager sees only their departments, sizes come in words, and an area under five people stays unreleased. Drafting follow-ups and write-ups stays in Spradley.

Where does it fit in an HR listening programme?

After every round. Your survey shows that a score moved; Spradley interviews people one on one, by text or voice, and asks a follow-up question on each answer; Personal Agent turns the report into answers, a readout for leaders and a guide for managers. If you are starting from a score drop, read how to find out why engagement scores dropped. For the interview itself, see what an AI-moderated employee interview is. Our HR and People page shows the whole loop, from a survey score to what managers change.

Frequently asked questions

Is there an AI agent that can answer questions about employee feedback?

Yes. Spradley’s Personal Agent answers a leader’s plain-language questions about the company’s interview reports, cites the area, claim or quote behind each point, and drafts a follow-up interview when the evidence is missing.

Can Personal Agent see what one employee said?

No. It reads the finished reports: areas, claims, published quotes and department leans. It never sees one person’s answer, and an area held by fewer than five people reads “size not released”.

Who can use Personal Agent?

Owners and managers. An owner’s agent reads every report in the company; a manager’s reads only the reports of the departments they were given access to.

Can it make a deck or a document?

Yes. It writes readouts, deep dives, comparisons, tension and root-cause analyses and manager conversation guides, which you can download as PowerPoint, PDF or Word or open in Google Docs.

Does Personal Agent send anything to employees by itself?

No. It drafts follow-up interviews, and a leader reviews and sends them. A follow-up must go to at least five people.

Key terms

Personal Agent: Spradley’s AI agent for each leader. It answers questions about what people said in interviews, cites its evidence and drafts the next step, without ever seeing one person’s answer.
Claim: One statement in an interview report, such as a cause, an effect or a proposal people made, that an answer can cite.
Write-up: A longer cited document the agent builds in the background, such as a readout or a conversation guide, checked sentence by sentence.
Minimum group size: The fewest people who must stand behind a finding before its size is shown. In Spradley it is five.
MCP server: A standard connection that lets AI apps such as Claude, ChatGPT or Copilot read data from another product, here Spradley’s reports, read-only.