AI in Workplace Investigations: What an HR Investigator Can Delegate

Yes, an HR consultant can use AI in a workplace investigation for supporting tasks: building a chronology from documents, summarising transcripts, indexing evidence and checking a report for gaps. The investigator must keep the findings, the credibility assessments, the evidence handling and the confidentiality decisions. Check every AI output against the source, and tell the commissioning organisation which tools you use.

The dividing line is whether you can check the output against a source. In a five-witness matter with 38 documents, an assistant can draft a 41-row chronology in minutes and you still open every cited document; it cannot tell you which account to believe. That holds for an independent HR consultant and for a small HR team running a few investigations at a time. Procedures and legal standards differ between jurisdictions and between employers, and I am not giving legal advice; where the point is about the law, ask an employment lawyer. The method below holds up wherever you work, because it rests on one principle: the machine arranges, you decide.

Follow me on Instagram@sagnikteaches

Why the split matters more here than elsewhere

Most of the vendor writing on AI in investigations promises speed. For an investigator, the risk runs the other way. An investigation can lead to dismissal, a tribunal or court case, or damage to a person's career, and the process is judged on fairness as much as on outcome. The report may be disclosed. Your notes and method may be examined. If a finding rests on a summary that distorted a witness's words, the damage is to the person accused, to the complainant, and to you.

Connect on LinkedInSagnik Bhattacharya

So the test for any AI task is: can I show exactly where this output came from, and did a person verify it against the original? Chronologies and transcript summaries pass that test when you check them. Findings and credibility do not, because they are not summaries of anything. They are your reasoned judgement, and the reasoning has to be yours.

Subscribe on YouTube@sagnikteaches

Tasks you can delegate, and the check each one needs

Everything on that list produces material you can verify against something. Nothing on it asks the assistant to decide what happened, who is telling the truth, or what should follow.

Task 1: Building a chronology from the documents

A chronology is the spine of most investigations and a tedious job: emails, messages, rota entries, minutes and statements, all dated, all to be placed in order with sources. An assistant can read a set of documents and list dated events, which saves hours. The danger is invention: a date it infers from context, or an event it blends from two documents.

Worked example: a complaint about a manager's conduct

Say a print-shop employee complains that her manager repeatedly made comments about her work in front of colleagues over four months. You hold 38 documents: 19 emails, 9 chat messages, 6 rota extracts and 4 meeting notes. Ethically and practically, you remove names and replace them with codes (Complainant, Manager, Witness A, Witness B) before using the tool, and keep the key in a separate secure file. You then give each document an ID and ask for a table.

Below are documents with IDs (D01 to D38). Build a chronology table with columns: date, document ID, what the document records (one sentence, using the document's own words where possible), and a quote of up to 20 words that supports it.
Rules: use only information in the documents. If a date is not stated, write "undated" and do not infer it. Do not interpret, assess or comment on credibility. If two documents conflict, list both entries and add "CONFLICT" in a final column.

A realistic output (illustrative) gives 41 rows. You verify it by opening every cited document and comparing the quote. In a test like this, expect a handful of errors: an entry attributed to the wrong document ID, a date taken from an email's send time when the text referred to an earlier meeting, and perhaps one summary stronger than the source. In an illustrative run, you might correct six rows out of 41. The saving is that you spend three hours checking rather than eight building, and you end up with a table that has a source on every line.

The rule that keeps it safe

Never accept a row you have not opened the source for. If a conflict appears, you decide what it means, and you may need to ask a witness. For the wider workflow of using AI with documents in compliance work, see using AI in compliance investigations, which sets out a general safe process.

Task 2: Summarising transcripts without changing what was said

Interview notes are the heart of an investigation, and a summary that flattens a witness's hesitation or hedging changes the evidence. Use an assistant for a navigational summary, a map of where topics were discussed, not as a substitute for reading. Ask for topic headings with timestamps and verbatim quotes of key passages, and not for a narrative.

This is a transcript of an investigation interview, with timestamps. Witness A is the interviewee; the interviewer's questions are marked INV.
Produce: (1) a list of topics discussed, each with the timestamp range; (2) for each of the three allegations in the terms of reference [list], the passages where the witness addressed it, quoted exactly with timestamps; (3) any point where the witness said they were unsure, did not remember, or changed their answer, quoted exactly.
Do not summarise in your own words, do not assess credibility, do not infer. If the witness did not address an allegation, say so.

The sample output for a 70-minute interview might list nine topics, quote seven passages on the first allegation and note two places where the witness said "I'm not sure whether that was before or after the meeting". You read those passages in the transcript and, for anything that matters to a finding, listen to the recording, because transcription errors in names, negatives ("did" for "didn't") and numbers are common. A mistranscribed negative can reverse a statement.

Check the transcript itself first

Before summarising, read the transcript against the audio for the first ten minutes. If the error rate is high (heavy accent, crosstalk, poor audio), fix the transcript or do not rely on the tool. Let the witness check their statement where your process allows, since that is the fairest way to catch both machine and human errors.

Task 3: Indexing evidence and what each person said per issue

A matrix of issues against witnesses is a standard way to see where accounts agree and diverge. Building it by hand across five witnesses and six issues means thirty cells, each with a quote. An assistant can extract the quotes and place them in a grid; you then read the grid and make the assessment.

For the example above, the rows are the six issues (the first comment, the group meeting, the email in March, the rota change, the conversation in the corridor, the follow-up) and the columns are the complainant, the manager and three witnesses. Each cell holds a verbatim quote and a document or timestamp reference, or "not addressed". The grid shows at a glance that only the manager and one witness addressed the corridor conversation, and that their accounts differ on the timing. You then decide whether to go back to anyone with a question. The assistant does not tell you what the difference means. Whether the difference is a lapse in memory, a disagreement over meaning or something more is the work of the investigator, and you must be able to explain your reasoning in the report.

What must stay with the investigator

• Findings of fact. Deciding what more likely than not happened, to the standard your process requires. A tool can lay out the evidence on each issue; it cannot weigh it.
• Credibility. Assessing consistency, plausibility, corroboration and demeanour is a human act. An assistant will produce confident language about who seems reliable, and that language has no basis. Do not ask it.
• Evidence handling. Deciding what is relevant, what is privileged, what to disclose and what to withhold, and keeping a clean audit trail of who handled what.
• Interview conduct. Questions, follow-ups, adjustments for a witness who needs support. You may use a tool to prepare a list of topics, but the interview is yours.
• The recommendation and the report's reasoning. Every conclusion should be traceable to evidence in your own words.
• Decisions about sanctions or outcomes. Those belong to the employer, not to you and certainly not to a tool.

The related tutorial on whether AI can draft a grievance report looks at the report-writing stage in more detail, and the general guide on how HR consultants use AI for policies, letters and cases covers the surrounding work.

UK and US practice: what changes for the investigator

The split between delegated tasks and judgement does not change between the two systems, but the yardstick your process is measured against does, and so does what your log needs to show.

In the UK, the Acas Code of Practice on disciplinary and grievance procedures says employers should carry out necessary investigations without unreasonable delay and, where practicable, use different people for the investigation and the disciplinary hearing. It also says there is no statutory right to be accompanied at a formal investigatory meeting, although the employer's own procedure may allow it, and tribunals can adjust awards by up to 25 per cent for an unreasonable failure to follow the Code. Acas is updating the Code and has consulted on a draft 2026 version, so check the current text. For AI use, your log should show that the tool saved time rather than added delay, and that it never blurred the line between investigating and deciding.

In the US there is no single equivalent code. For harassment, the EEOC says an employer is liable for harassment by non-supervisory employees, or non-employees it controls, if it knew or should have known and failed to take prompt and appropriate corrective action; the agency also notes that its harassment pages are under review. A careful investigation is often how the employer shows it acted promptly and appropriately, so a chronology that cites its sources earns its keep. If the employer's lawyer commissioned the investigation, ask them how material should be handled before any of it goes into a third-party tool.

Confidentiality and evidence handling in practice

An investigation file holds some of the most sensitive information an employer has: allegations about named people, health information, sometimes details of third parties. Before any of it goes near an AI tool, settle four things.

1. Authority. Does your engagement letter or the client's instructions permit the use of third-party tools on the file? If it is silent, ask and record the answer. A short paragraph in your terms of reference naming the tools and the tasks avoids later argument.
2. The tool. Use a business plan with data terms you have read, with training on your content off and a retention setting you understand. Free personal accounts may treat data differently. Check where data is stored and whether a human at the vendor can read it, and whether you can delete the conversation.
3. Minimising what is shared. Replace names with codes, remove contact details, dates of birth and identifiers, and paste only what the task needs. Keep the key to the codes separately. For a sensitive matter, consider whether the task is worth the exposure at all.
4. The trail. Keep a log: which task, which tool, which documents, what the output was, who checked it and when. If the matter is challenged, this is your evidence that the process was controlled.

Practice around disclosure, employee access to their own data and the weight given to different kinds of evidence varies by jurisdiction. If you work for clients in more than one place, check the position for each. A common trap is using a recording or transcription tool in a meeting without telling the participants; tell them, and record that you did.

A second matter: a short dispute between two colleagues

Not every case is a five-witness matter. Take a yoga studio's employer who asks you to look into an allegation that one instructor took client bookings from another. There are two interviews, eleven documents (booking-system exports and messages) and a two-week window. Here the chronology is a one-page list, and the gain from AI is smaller: perhaps ninety minutes saved, mostly on turning the booking-system export into a dated list of which instructor was assigned each class and when it changed.

The check is the same, and in a small matter it can feel excessive, which is exactly when people skip it. A sample check log for this matter would have four lines: "Chronology built by tool from D01 to D11; every row opened and compared; two corrections (dates swapped on rows 6 and 7). Interview 1 topic map checked against recording at the three passages cited in the draft findings. Evidence index checked for completeness against the client's email of the 3rd. Reviewed by me on the 14th." Short, dated, attributable, and enough to show what was done if anyone asks later.

The lesson from a small matter is proportionality. Use AI where the volume justifies it, and do it by hand where a quick read is faster than writing a prompt, checking the output and logging it. A tool that saves ninety minutes and demands sixty minutes of checking is not a win.

Drafting letters and admin around the investigation

Outside the evidence, investigations generate letters and scheduling: invitations to interview, confirmations of the right to be accompanied where your process provides it, reminders about confidentiality, and outcome letters that carry the decision others have made. These are low-risk uses because the content is largely fixed by the process, not by judgement. Give the assistant your template and the facts and ask it to fill in the blanks, not to compose the substance.

Fill in this interview invitation using only the facts below. Do not change the standard paragraphs about confidentiality and the right to be accompanied. Do not add reassurance or comment on the allegations.
Facts: interviewee first name [first name]; date and time [date]; location [room or video link]; duration up to 90 minutes; contact for questions [your name].
Template: [paste your approved template]

The sample output is your template with the blanks completed. Read the result against the template to confirm no standard paragraph was reworded, since a tool tidying your phrasing on the right to be accompanied can change its meaning. For wording that touches disciplinary outcomes, use the checklist in what HR consultants must check when AI drafts disciplinary letters.

When a model's credibility phrase reaches the report

Suppose an investigator asks an assistant to summarise an interview and add a view on whether the witness was credible. The assistant writes that the witness "appeared evasive and inconsistent". There is no evasion in the transcript; the witness hesitated twice and corrected a date. The investigator, tired, adapts the phrase for the report. The accused's representative later asks what the finding of inconsistency rests on, and there is no answer, because the phrase came from a model, not from the evidence. The credibility of the whole report is hit.

The safeguard is procedural. The prompt forbids assessment, the report template requires a source reference for every finding, and a second person reads the draft with the source documents. If a sentence in the report has no reference, it is deleted or sourced. That rule would also have caught a confident summary that contradicted the record.

Telling the parties and the client what you do

Put a short statement in your terms of reference or opening letter, for example: "I may use software tools, including AI-based tools, to help organise documents and prepare summaries of transcripts. Every output is checked against the original material. Findings and conclusions are my own. No personal information is shared with a tool except as described in the data handling note." Parties are more likely to trust a process that is described openly than one they discover. If a party objects, you can discuss it, and sometimes the answer is to do that part of the work manually.

A second statement belongs in the final report's method section: which tools were used, for which tasks, and how outputs were verified. It need not be long. It shows the reader that the methods were controlled.

Estimating the time you save, honestly

For a mid-sized matter, such as the manager-conduct example with five witnesses and 38 documents, the supporting tasks might take an investigator about 30 hours by hand: eight on the chronology, ten on transcripts and extracts, six on the evidence index and issue grid and six on admin. With AI doing the first pass and you verifying each line, they might take 16 to 18 hours. Illustrative savings of 12 to 14 hours per matter are plausible, but only when verification is strict. If you shorten the checking, you save more time and add risk, and that is the wrong direction.

Time one matter yourself and keep a note of the corrections you made, because that evidence is better than any vendor claim. Most vendor material on this topic is marketing and sits behind a form. A good test of a tool is to ask whether it cites a source for every line it produces and whether it lets you delete the data.

Starting without risking a live matter

1. Use a closed, old matter or invented documents. Build a ten-document dummy file with a few deliberate traps: an undated email, a conflicting date, a witness who hesitates.
2. Run the chronology and transcript prompts. Score what the tool gets wrong and write the corrections into your prompts.
3. Write your protocol. One page: approved tool, what may be shared, the coding of names, the verification rule and the log.
4. Brief the client. Agree it in writing before the first live matter.
5. Use it on a low-complexity live matter first. Review at the end: time, corrections and any concerns raised.

For the routine HR work around investigations, such as staff questions and letters, answering staff HR questions with an AI helpdesk and AI for HR in small businesses list tasks that carry far less risk. If you would rather map this with someone, that is what my AI implementation consultation covers: choosing the supporting tasks worth automating in your practice and handing over a protocol your team can run.

Questions HR investigators ask about AI

Can AI decide whether a complaint is upheld?

No. The finding rests on weighing evidence and assessing credibility, which is the investigator's responsibility and may be challenged later. AI can organise the evidence so you weigh it faster, but the reasoning, the conclusion and the report must be your own and written so you can defend every sentence.

Is it safe to put witness statements into a chat assistant?

Only if you have the authority and a tool that suits the confidentiality of the matter. Check the vendor's data terms, your client's instructions and your engagement letter. Use a business plan, turn off any training or sharing option, redact names where you can, and tell the commissioning organisation what you use.

Should I tell the parties that AI was used?

Transparency is the safer position. Say in your terms of reference what tools you use and for what, and that every output is checked against the source. If a party asks, you can answer plainly. Hiding it risks the credibility of the whole process if it comes out later.

Can AI draft the investigation report?

It can help with structure, tidy neutral background sections and check the report against your terms of reference for gaps, but the findings and reasoning must be yours. Drafting the analysis with a tool invites unsupported conclusions, so write that part yourself and use AI only to check consistency afterwards.

Further reads

• How to Get Your Staff to Actually Use AI Tools — Getting a team to use AI tools with the right rules.
• AI for Employment Lawyers: Policies, Grievance Bundles and Letters — How employment lawyers handle grievance bundles.
• What to Check in an AI Tool's Privacy Policy and Terms — What to check in an AI tool's privacy terms.
• Can Solicitors Use ChatGPT Without Breaching Confidentiality? — Confidentiality questions from another profession.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Acas Code of Practice on disciplinary and grievance procedures; EEOC harassment page (both checked 4 November 2026); general investigation practice. Employment law and procedures vary, so take advice for your own matters.