Own Your Insight. AI is a Tool.

Quark Insights - Making the best use of AI

AI is a remarkable tool for gathering and organizing information. What it can't do is do the analysis for you.

I'll be direct about where I stand on this: I use AI tools regularly. Tools like Claude and Perplexity have genuinely changed how I approach research and managing information. When I need to pull together a brief quickly, scan multiple sources, or organize a sprawling set of inputs into something coherent, AI does that faster than I can on my own it's an efficiency. It's not a gimmick, and I'm not using it reluctantly. I think it's one of the more powerful resources available to analysts right now.

But here's the thing I keep coming back to, and the reason I wanted to write this piece. AI is exceptionally good as a support mechanism for menial tasks and workflows. AI is not a substitute for the judgment, context, and real-world knowledge that turns that information into actual insight. That part still belongs to the analyst. Always.

I've seen the output/content from sources using AI to formulate ideas (haven't we all?). A question gets fed in, and what comes back is a structured, confident-sounding response, but there's always something not quite right in the sound and logic - worse yet, something important is conceded. The result is a pasteurized paraphrase of generalized thought - and usually not relevant. The author did not evaluate AI's response, challenge it, and apply it to what they know about the business, the audience, and the situation in front of them. That gap, between what AI produces and what analysis actually requires, is where the real work lives.

This article is about how I navigate that gap. Specifically, where I've found AI to be genuinely useful, how I've learned to pull back and rely on my own judgment, and the approaches I use to make sure the insight remains mine.

Why This Distinction Actually Matters

Information and insight are not the same thing

There's a tendency, especially right now when AI tools are new and impressive, to conflate two very different things: collecting and organizing information on one hand, and interpreting that information on the other.

AI is genuinely brilliant at the first part. Give it a broad question and it will surface relevant sources, summarize large volumes of content, and structure what it finds in a way that's easy to work with. That's real value. In my experience, it collapses research tasks that used to take hours into something much more manageable. I'm not underselling that.

What it cannot do well is the second part. Interpretation requires knowing what a finding means in the context of a specific business, a specific market, a specific moment. It requires understanding what the data doesn't say as clearly as what it does. It requires the kind of judgment that comes from experience, from previous conversations with stakeholders, from watching similar situations play out before. That's not something you can prompt your way to.

When I use AI, I think of it the way I think of a very well-read research assistant who is fast, thorough, and organized, but who has never actually worked in the industry I'm analyzing. The information they bring back is useful. What I do with it is still on me.

Step 1: Align before Touching Your Data

The work that happens before the work is what keeps everything on track

The first thing I do on any significant analysis project has nothing to do with data or AI. It's people. Before I open a file, run a query, or prompt any tool, I make sure that everyone involved, my team and the client is clear on what we're actually trying to produce and why.

This isn't a formality. It's the most important investment of time I make on a project. Analysis that starts without clear, shared alignment on the question almost always ends up answering the wrong thing. You can have perfect data, clean methodology, and a beautiful presentation, and still miss the point entirely because nobody stopped to confirm what "the point" actually was.

What alignment looks like in practice for me is a working conversation, not a form. I talk with the client about the decision their analysis needs to support, who the audience is, what format will serve them best, and what questions, if answered, would give them what they need to act. I have the same conversation with my team so no one is pulling in different directions. The goal is that by the time anyone touches a dataset, we all have the same picture of what we're building and who we're building it for.

Data sourcing and citations are part of this orientation phase too, and I take them seriously. Knowing where your data comes from, the source, the date range, the methodology behind it, and any assumptions baked in, is not something you want to reconstruct at the end of a project. I use a source document to track this from the start, and I've made a version of that document available for free on the Quark Insights website because I think it's one of the most underleveraged organizational tools in this field. Getting your sourcing and assumptions documented before the analysis begins protects the integrity of everything that follows.

Once that foundation is in place, something else happens: your AI prompts get dramatically better. When I know the decision, the audience, and the specific questions I'm trying to answer, I can give an AI tool real context instead of a vague request. "Summarize the competitive landscape for a VP of Marketing making a channel investment decision for Q3, with a focus on emerging digital channels" is a completely different prompt than "tell me about the competitive landscape." The quality of what comes back reflects that difference.

A regional sales director once sent my team a request: "Can you pull some numbers on last quarter's performance?" Before anyone touched a single file, I went back to her with two questions, what decision is this supporting, and what format would actually be useful for that conversation? She clarified that she needed to decide whether to shift reps between territories before the next quarter, with a short presentation to her VP on the calendar. That conversation took 10 minutes and saved probably two days of work going in the wrong direction.

Step 2: Organize & Orient Data for Analysis

Constrain what AI is using, and interrogate it to stay on task

Once alignment is done, and I know exactly what I'm trying to answer, the next step is making sure the data I'm working with is actually ready to support that question. This is also where my relationship with AI gets its first real test, because messy data and an unconstrained AI tool is a combination that produces confident-sounding nonsense faster than almost anything else I've encountered.

An analysis-ready dataset has a clear unit per row, plain-language column names, consistent formats, and known limitations. If the column labeled "revenue" means one thing in your customer relationship management (CRM) export and something slightly different in your billing platform, you have a problem that no AI tool is going to catch for you. You'll find out when a stakeholder asks a question your numbers can't actually answer.

My approach to data hygiene is non-negotiable. I save every raw export in a folder labeled "Original" and never touch it again. I work from a copy labeled "Clean" or "Working." I make sure each row represents one clear unit, whether that's one campaign, one customer, or one week, and I rename columns so that any person reading the file can understand what each field actually represents without having to ask. Then I write a short note documenting what I filtered, joined, or changed. That record matters more than most people think, especially when a client comes back three weeks later asking how a specific number was derived.

Step 2: Clean Data and a Short Leash

Getting your data right is only half the job; keeping AI focused on it is the other half

Once alignment is established and I know exactly what I'm trying to answer, the next step is making sure the data I'm working with is actually ready to support that question. This is also where my relationship with AI gets its first real test, because messy data and an unconstrained AI tool is a combination that produces confident-sounding nonsense faster than almost anything else I've encountered.

An analysis-ready dataset has a clear unit per row, plain-language column names, consistent formats, and known limitations. If the column labeled "revenue" means one thing in your customer relationship management (CRM) export and something slightly different in your billing platform, you have a problem that no AI tool is going to catch for you. You'll find out when a stakeholder asks a question your numbers can't actually answer.

My approach to data organization/orientation is non-negotiable. I save every raw export in a folder labeled "Original" and never touch it again. I work from a copy labeled "Clean" or "Working." I make sure each row represents one clear unit, whether that's one campaign, one customer, or one week, and I rename columns so that any person reading the file can understand what each field actually represents without having to ask. Then I write a short note documenting what I filtered, joined, or changed. That record matters more than most people think, especially when a client comes back three weeks later asking how a specific number was derived.

Clean data gets you to the starting line. What you do with AI from that point forward is where the real discipline comes in, and where I see the most preventable mistakes happen.

AI tools do not have adequate guardrails to ensure they are working only with the data you've given them. Left without clear constraints, they will draw on outside sources, make inferences that go beyond your dataset, and wander into territory you never intended, all without flagging it. The output will still look polished and organized. That's exactly what makes it a problem.

When I bring AI into an analysis, I'm deliberate about establishing data boundaries in the very first prompt. I tell it precisely what it's working with: the dataset, the time range, the specific fields I want it to consider. I make clear those are the only inputs. Something like: "Work only from the data I've provided here. Do not reference outside sources or draw inferences beyond what's in this dataset." That instruction goes in at the start, and I reinforce it as the work develops.

More critically, I don't set those constraints and walk away. I interrogate the outputs at every step. If AI surfaces a pattern or produces a ranking, I ask it to trace that back to the data. I push back. I ask it to explain its reasoning in plain terms. I'm checking whether what it's telling me reflects the actual numbers I gave it or whether something has crept in from somewhere else. One misread field, one quiet assumption about what a column represents, and your trend data is pointing the wrong direction. If you're not close enough to catch that, the error moves forward into your analysis and nobody knows it happened until someone asks a hard question in a client meeting.

This is precisely why the human element in AI-assisted analysis is not a courtesy. It's a requirement. Clean data going in does not guarantee clean analysis coming out. You have to stay engaged, keep the tool constrained, and challenge what it gives you at every meaningful step. When I do that consistently, the collaboration works well. AI handles the mechanical work of structuring and exploring the data faster than I could do it alone, and my oversight of its outputs is what keeps the work grounded in the actual evidence.

Step 3: Keep it Simple

Ranking and trend are the two lenses I reach for first

One of the places AI has genuinely helped me is in the mechanics of sorting, structuring, and visualizing data. But I want to be clear about what I use it for and what I don't. I use AI to help me build and format the views. I don't use it to tell me what those views mean.

The two views I come back to most are ranking and trend. Ranking means sorting items by a key metric so I can see quickly what's working and what isn't. Trend means watching how that metric changes over time for the items that matter most. Together, they answer the most common question a stakeholder has: what should I pay attention to, and is it getting better or worse?

If I'm looking at 80 live campaigns, I'm not going to analyze all 80 in depth. I sort by cost per lead, identify the top 15 and the bottom 15, and then look at how those 30 have trended over the past three months. That's a manageable scope. The pattern usually becomes clear pretty quickly, and when it does, the interpretation is mine to make, not AI's.

I'll use AI to suggest chart formats or help me think about how to structure a summary for a specific audience. But the question of what the trend means, whether it's a signal worth acting on, and what the right recommendation is still mine to define, that's where I need to be in the driver's seat. AI can tell me a cost per lead went up. It can't tell me whether that's a problem, an anomaly, or a signal that a specific channel is maturing. That requires knowing the context.

Step 4: Always Show Your Work

Transparency is what turns a finding into something people will actually use

One of the things I feel strongly about, and try to practice consistently, is making sure anyone who reads my analysis can see exactly how I got there. Where the data came from, what I filtered or changed, what assumptions I made, and yes, where AI tools played a role.

This matters for a few reasons. It builds trust with the people who use your work. It makes your analysis reusable. And increasingly, it answers a question that stakeholders are starting to ask out loud: how much of this came from a machine, and how much came from a person? If you're working on forecasts, you know exactly where I'm coming from on this point.

My answer to that question, in my own work, is that AI helps me gather and organize. The interpretation and the recommendation are mine. I make that explicit in what I produce.

I add a small sourcing/citation note to every output, usually just a few lines at the bottom of a slide or the end of a document. It covers the data source and time range, the major steps I took to clean or structure the data, any assumptions that are factored into the numbers. It retains my credibility and can serve to deflect severe grilling. It's a lot easier to defent your work when you make what you've used and done known.

It doesn't take long to write. And it changes how people receive the work.

Step 5: Keep AI in Its Lane and Stay in Your Own

The clearest guardrail I use: AI organizes, I interpret

When I prompt an AI tool, I'm always explicit about what I want it to do and what I'm going to do myself. I ask for structure, summaries, options, and draft language. I don't ask it to reach conclusions or make recommendations. That distinction has held up well for me across a lot of different projects.

A prompt that works: "We're supporting a budget reallocation decision for next month. Here's campaign-level data , Spend, Leads, Channel, Cost per Lead, last 90 days. Suggest three ways to structure a brief summary for a marketing VP." What I get back is a set of structural options I can evaluate and choose between. The framing of what matters, and what to recommend, comes after I've reviewed the data myself.

What doesn't work: "Look at this data and tell me where we should move the budget." That produces an output that sounds confident and may even be directionally right, but it's built on the AI's pattern recognition, not on my understanding of the business, the team, the history, or the competitive context. I don't trust that. I've seen it miss things that any analyst with six months of context would catch immediately.

The version of this workflow that I've found most reliable: I use AI early and often to gather information, structure what I'm working with, and think through how to present my findings to a specific audience. Then I step back, apply what I know, and write the actual interpretation and recommendation myself. AI does the setup. I do the analysis.

Putting It Together

The five steps I've shared aren't complicated, and they don't require any special tools beyond what most analysts already have access to. What they require is a willingness to stay in charge of your own thinking, which sounds obvious, but is genuinely harder than it looks when you have a tool that can produce polished, confident output in seconds.

Here's how the steps connect:

1. Prioritize Alignment. Confirm the question, the audience, and your data sources with your team and client before anything else.

2. Organize/orient your data. Build an analysis-ready structure you understand and can describe clearly.

3. Use simple views. Ranking and trend to surface what matters and focus your attention.

4. Document everything. Sources, steps, assumptions, and how AI contributed.

5. Keep the interpretation yours. AI organizes. You evaluate, apply, and conclude.

What I've found is that working this way actually makes AI more useful, not less. When I know what I'm trying to answer, when my data is clean, and when I'm clear on what I want AI to help with versus what I'm going to do myself, the outputs are better, the analysis is faster, and the work is something I can stand behind.

Try This on Your Next Project

Pick one analysis project you're working on now, a client report, an internal request, a research summary, and start with alignment instead of data. Define the decision, the audience, and the three questions your work needs to answer. Document your sources before you begin. Then track where you use AI and where you make the calls yourself.

I'd be curious what you find. The line between what AI does well and where human judgment is non-negotiable tends to get clearer the more deliberately you draw it. Drop your thoughts in the comments, this is the kind of thing that gets sharper with real examples from real practitioners.

The Part That Doesn't Change

AI tools will keep improving. The volume of information they can synthesize and the speed at which they can organize it will only grow. What won't change is the value of someone who can look at that organized information and understand what it actually means, for this client, this market, this moment.

That's the skill worth protecting. AI can do a lot of things, but it can't replace the analyst who knows the business, understands the context, and can translate a data pattern into a decision that people will act on. Ultimately, I fall back on the old sniff test. Read your analysis after you write it, and imaging yourself in your client's position - does it make sense?