Deploying AI across the research life cycle

Last summer, I gave a workshop to the staff on augmenting the research proposal process with AI. Thought I’ll sketch out what an AI-assisted research workflow might actually look like for both quantitative and qualitative work. One caveat before we get into it. This is a map of possibilities, not a prescription. What you can actually do depends on your data, your project, and what UNC Charlotte currently permits. Before trying any of this, check what the university’s Office of OneIT has published on approved tools and data classification. That guidance sets the real boundaries. This is more of a documentation of experiments that hopefully will be part of how our AI strategy takes shape.

The quantitative research pathway

Brainstorming and question development

Figuring out what question you’re actually asking tends to be a messy, iterative conversation early in a project. I’ve started using AI as a thinking partner at this stage, not to generate the question but to pressure-test it. Describe the policy problem, and Claude or ChatGPT will usually surface the research questions embedded in it, sketch the data you’d need, etc.. What it can’t do is tell you whether that if these questions actually matter for the partner/client and it’s easy to replace your thinking with that of AI at this stage. I wouldn’t want a language model to lull me into projecting ideas without careful deliberation and thought stemming from expertise. Use these tools to think faster, not to replace the thinking.

Literature review and background research

Tools like Elicit, Consensus, and Research Rabbit have changed how I approach literature reviews. What used to take a week of database searching now takes a day. You can upload a set of PDFs and ask for themes, methodological patterns, or gaps. Claude and ChatGPT are useful for synthesis once you have sources in hand. The fabrication problem is not a footnote here. I have had AI tools confidently cite papers that do not exist. Every source from an AI-assisted literature process needs to be checked against the actual publication before it goes anywhere near a report.

Research design and analysis planning

Once you have your question and your data, AI can help you think through the analytic options. Describe your outcome, your data structure, your main variables, and Gemini will walk you through the tradeoffs between estimation strategies and flag confounders worth considering. The analyst still makes the calls. AI has no way of knowing whether a propensity score model is defensible for your specific dataset, or whether a comparison group will hold up to a client’s questions. That’s judgment built from being in the work, not something a tool can supply.

Coding plan and data cleaning

This is where I’ve personally gotten the most out of these tools. If you’re working in R or Python, describing a data transformation task in plain English, like “I have a panel dataset with multiple rows per household and need to reshape it to one row, keeping the most recent record for each variable,” gets you working code faster than Stack Overflow. GitHub Copilot and the coding assistants in Claude and ChatGPT both handle this well. Data cleaning tasks that used to eat hours, standardizing inconsistent address strings, recoding open-ended responses, joining datasets with mismatched identifiers, can be cut down substantially. The code still needs review. I’ve had output that ran clean and returned wrong answers. Sometimes it takes a minute to catch.

Analysis assistance

For regression output, visualization, and formatted tables, AI saves real time. Describe what you want a chart to show, and you get working code most of the way there. Power BI and Tableau are now building similar querying into their own interfaces. Where I’d draw the line: AI should not be selecting your model specification, choosing which variables to include, or writing the interpretation of coefficients. Those decisions require knowing what the numbers actually represent in your study context. Delegating them to a tool is a shortcut that tends to produce analysis no one can fully defend.

Review, interpretation, and shaping results

After you have findings, AI can draft a plain-language summary and flag places where a reader will ask questions you haven’t answered. It can also help you anticipate how a result might get misread. The most useful thing I’ve found at this stage is running a draft interpretation past Claude and asking what I’m overstating. It pushes back more usefully than I expected, and it’s faster than waiting for a colleague to have time to read a draft.

Report writing and editing

AI editing has gotten genuinely good at catching passive constructions, suggesting clearer alternatives to jargon, and rewriting for a different audience. Moving a methods section from academic to plain-language takes minutes rather than an afternoon. Low risk, concrete payoff. What it shouldn’t be doing is writing the findings narrative. The synthesis of what the data shows, why it matters for this community, what a partner should actually do with it: that needs to come from someone who was in the research. Outsourcing that part produces analysis that reads fine and says nothing.

The qualitative research pathway

Brainstorming and question development

The same basic logic applies here, though the texture is different. AI is useful for drafting and stress-testing interview guides and focus group questions, catching where a question is leading or too broad, and thinking through how different framings might land with different community members. A few qualitative researchers I’ve talked to at other institutes use it to plan engagement sequences, working out how to structure a series of community conversations so each one builds on the last without making participants feel like they’re covering the same ground. That use surprised me when I first heard it, but it makes sense.

Literature and conceptual grounding

The literature review tools work the same way here. Elicit and Research Rabbit don’t care whether you’re looking for ethnographic studies or regression papers. The verification requirement is the same too. AI is also useful for tracing how a concept has been defined across disciplines, or identifying methodological debates you’d want to know about before designing a study. That kind of background work used to take a long time. Now it takes a few well-constructed prompts.

Research design

For interview guides, sampling logic, and participant criteria, AI is an effective sounding board. You can engage in a back-and-forth to refine your thoughts on the population, the research question, and what you’re trying to learn, and if the design enables to answer the questions without serious limitations. But it lacks local knowledge and the essence of conversations you have had in the community so far. Knowing which community leaders to approach first, how a particular organization’s history shapes what they’re willing to say in a group setting, what it means that a specific partner agreed to participate: none of that is in a language model. Qualitative research in community settings runs on relationships and context that take years to build.

Data collection and transcription

Automated transcription is more reliable than ever before. Otter.ai and Whisper both handle clean audio well. The transcription features built into Teams and Zoom also have improved over the past year. I think they still struggle with heavy regional accents, overlapping speakers, and technical vocabulary. But this is certainly a promising sub-area of qualitative research that’d benefit when dealing with bilingual or multilingual participants. Any recording that contains names, addresses, or other details that could identify a participant stays out of external AI tools. This means standard consumer transcription services, unless your project is using an enterprise-licensed version with a data protection agreement covering your institution.

Coding plan

AI can help you build a preliminary coding framework before you go into the data. Describe your research question and ask for a draft set of thematic categories and potential subcodes. It won’t be right, but it gives you something to react to, which is often faster than building from a blank page. The codes still need to come from what people actually said. Use AI to think about structure and then set it aside when you go into the transcripts. The real work of grounded coding is reading closely and letting the material lead.

Assisted coding and review

Several tools now let you apply a preliminary coding scheme to a set of transcripts and flag relevant passages for review. For projects with twenty or more interviews, that kind of first pass can save a lot of time. But human review of a real sample isn’t optional. People’s tones, attitude, approach, and all relevant metadata about the conversation have to be recorded by the researcher to meaningfully analyze the start given by AI tool. And be free to discard any conversation that isn’t going according to how you planned.

Analysis and interpretation

The core of thematic analysis, finding patterns across coded material, noticing where accounts contradict each other, deciding what the data actually shows, has to stay with the researcher. AI can help you organize excerpts and draft what the themes appear to show. I’ve used the prompt “what’s the counter-evidence to this interpretation?” enough times that it’s now a standard step in my process. The insight work is human, though. There’s a moment in qualitative analysis where you realize that what people are saying about housing insecurity is actually about something else, family disruption, the sense that institutions don’t keep their promises, a neighborhood that’s changing faster than people can adjust to. That doesn’t come from a tool. It comes from reading carefully and knowing what you’re listening for.

Report writing and editing

What I said about editing in the quantitative section applies here too. AI is good at clarity, catches passive voice, and helps you translate between audiences. Use it. The findings narrative is different. A qualitative report lives or dies on whether the synthesis of what you heard rings true to the people who were in those conversations and to the community whose experience it’s describing. That credibility comes from the researcher’s judgment exclusively.

Checks, guidelines, and ethical framework

Data governance first

Your data classification determines your tool options. Publicly available data, aggregate indicators, Census tables, published datasets, can go into approved external tools. Internal drafts and non-sensitive project materials can go into enterprise-licensed tools like Microsoft Copilot or Google Gemini through your institutional account; both have data protection agreements with UNC Charlotte. Identified data, anything covered by FERPA or HIPAA, data shared under a restricted use agreement, and all Data Trust datasets do not go into external AI tools.

Verification is not optional

Every factual claim from an AI-assisted step needs to be verified before it appears in a report, a client communication, or a public product. Every citation needs to be checked against the actual source. Every piece of generated code needs to be reviewed for logic, not just whether it runs. Make this a habit at each stage, not a quality check at the end.

Disclosure and documentation

We don’t yet have a field-wide standard for how to document AI-assisted work in methods sections. I think we should document it the same way we’d document any other methodological choice: what tools were used, for what tasks, how outputs were reviewed. That seems like the minimum. For community-facing research, the bar is higher. Partners and residents whose data we hold have a legitimate interest in knowing when AI was part of the process that produced findings affecting their communities. That’s not a legal requirement right now. I think it should be our practice anyway.

Human judgment stays in the work

AI should make researchers faster and expand what they can do on their own. It should not be making the calls that define what the research says. Choosing your research question, specifying your model, deciding what your coded themes mean, writing what the findings imply for policy or practice: those stay with the person who knows the context.

Community and ethical obligations

The communities we work with shared their data because they trust us to handle it carefully. That trust is the precondition for everything else we do. Using AI in ways that weren’t part of how that trust was established, even where it’s technically permitted, erodes it. Transparency with partners about when and how AI is involved in research that affects them isn’t optional, even when no policy requires it.

Follow university policy

Everything in this document is a set of observations and possibilities. UNC Charlotte’s Office of OneIT and the university’s AI guidance set the actual boundaries: which tools are approved, which data classifications they apply to, how research data policies and AI policies interact. Check that guidance before adopting any workflow described here.

What comes next

The technology will change. Some of what I’ve described here will look dated in a year. But the questions underneath it, what good research requires, what we owe the communities we work with, where the line is between using a tool and letting it do your thinking for you, those aren’t going anywhere. I’d rather work those out together than discover after the fact that everyone made different calls.