.png)
What an enterprise analytics leader and our own product team taught us about the real source of an AI advantage
Everyone has AI now. Same models, same chatbots, same productivity jump. So in a market where every competitor has the same tools, what is actually left to compete on?
That was the question behind our recent fireside chat, "The AI Is the Easy Part." We were joined by Diana Gowe, Director of Global Strategic Analytics at a Fortune 50 healthcare company, alongside Charlie Temkin from the product team at Northern Light. Diana brought the practitioner's view from inside one of the most demanding research environments in the world. Charlie brought the product and infrastructure side. Together they made a case that is easy to state and hard to live up to: the model is the commodity, and the advantage has moved to what the model is built on.
Here is what came out of the conversation.
Skepticism is the normal starting point
Diana was candid about where her team began. When AI was first approved in their research portal, there was real skepticism, and she counted herself among the biggest skeptics. The fear was the one everyone in this space shares: hallucinations. There is so much noise from general-purpose tools, and so much junk in the answers, that trust does not come automatically.
What changed it was not a better chatbot. It was the content underneath. Because the data in the portal was aggregated from trusted, vetted sources, the answers held up. Usage climbed as people tested it, ran summaries, and saw the results come back accurate. As Diana put it, comfort grew because the answers could actually be trusted.
Charlie framed why that arc is so predictable. Everyone got access to a tool at roughly the same moment. The differentiation is no longer access. It is being intentional about how the tool is deployed and what data it can reach. The teams winning with AI are not the ones with a different model. They are the ones who put AI on top of content they can stand behind.
Faster work and better decisions are not the same thing
One of the sharper moments came when Diana drew a line between speed and judgment. She calls AI a force multiplier. Let the machine do what the machine does best, which is combing through tens of thousands of pieces of information and quickly returning what is most relevant. That is the faster work.
But the better decision still belongs to the human in the middle. The competitive intelligence professional is not the one making the business decision. Their job is to generate the insight that lets senior leaders decide well. When a summary that used to take hours arrives in minutes, that time does not disappear. It moves to the strategic question: what does this mean for the business, and what are the implications? That is the part AI cannot do, because it requires sitting inside the business and understanding the environment you operate in.
The cost of an answer you cannot verify
The risk in all of this is concrete. Nearly half of executives act on AI outputs they never verified. In competitive and market intelligence, where a single call can carry tens of millions of dollars, that is not a rounding error.
Diana told a story that captured both the danger and the fix. A portal user ran a query tied to a high-stakes business development decision, the kind of regulatory question where being wrong is expensive. The answer came back, and the user did not believe it, so she was ready to walk away. Diana clicked the citation, which took her straight to the underlying source database, which confirmed the answer. The AI had been right. The point was not that the machine is always correct. It is that the user could check in seconds. Diana noted it took her longer to type her email into the source database than it did to click through and verify.
Charlie connected this to a shift in human behavior. For most of our lives, polish was a proxy for effort. If something looked well-researched and highly formatted, someone had probably spent real time on it, so it earned more trust. AI broke that overnight. Now anything can look good and pass the sniff test instantly. The job, he argued, is to update our sniff tests, with in-line citations that show not just that a source exists but where each specific claim came from, as you read it.
Why full documents beat fragments
This was the most technical part of the conversation, and one of the most important.
Charlie explained that most AI tools handle long documents by chopping them into chunks, roughly paragraph-sized pieces, and then retrieving whichever chunks seem most relevant to your question. For a simple task, like asking how much parking costs in an apartment lease, that works fine. But syndicated research, patents, and market intelligence reports are deeply self-referential. A passage in one section may depend on an appendix, an exhibit, and a diagram elsewhere in the document. Break the document apart and you lose those connections, which is a likely source of the incomplete answers and hallucinations people experience with general-purpose tools.
Northern Light takes a different approach. Where possible, the entire document is given to the AI model rather than fragments of it. That can cost a little more in processing, perhaps a second or two longer, but the trade is worth it when the decisions on the line are this consequential.
Diana had lived the difference without knowing the cause. She ran the same prompt in two tools. One kept surfacing the same passage repeatedly and left her underwhelmed. The same prompt in her portal returned far more references and noticeably better answers, even though a similar dataset existed in both. When the difference between chunking and full-document retrieval was explained to her, it clicked. She got insights she would otherwise have missed.
Traceability is what earns adoption
Citations did more than build trust. In one case they saved real money.
Diana ran a deep research report to test a business partner's hypothesis. The output was strong enough that she built a few slides from it. Weeks later the partner asked for the references in case the findings were challenged. She highlighted what she had used and sent it over, and he was struck by how many sources backed the work. Because the insight was trusted and traceable, the team did not need to commission a primary research study they had been planning, which saved a couple hundred thousand dollars.
Charlie's point on why this matters: a tool that cites everything only at the end of a forty-page report leaves you roughly where you started, with more to verify, not less. Traceability has to live where the claim lives, in line, so the human in the loop can confirm a statement the moment they read it and keep moving.
Governance is a feature, not a hurdle
It is easy to treat governance as the painful gate at the end. Diana reframed it. Governance is what aligns an AI initiative with organizational goals, balances innovation with responsible use, and minimizes legal, regulatory, and reputational risk. Going through the process also lends credibility, because she can tell users the application is approved, the sources are approved, and the summaries have been reviewed.
A few things mattered most in getting that approval, and they double as a checklist for anyone facing the same process:
- Every statement carries a visible citation you can click through to the original source, not just one or two references at the end.
- Explicit permission from data sources, including proprietary suppliers, to run generative AI against their content. If permission cannot be secured, the AI can be switched off for that source.
- Zero retention, with assurance that prompts and questions are not saved or fed into any training model.
- No training set being built from the licensed content.
Charlie underscored the prompt point. Asking a chatbot how long to bake a cake is harmless. But the questions a team like Diana's asks are strategy. If those prompts leaked, they would reveal what a company is investigating, which is competitive intelligence in its own right. The data matters, and so do the questions.
Build versus buy, honestly
The build-it-ourselves question comes up constantly, and Diana does not dodge it. Her platform has existed for more than twenty years, evolving into modern AI capabilities on top of data sources already aggregated. Adding AI to an existing, governed foundation was faster, easier, and more cost effective than starting from scratch. Her broader point: this technology moves so quickly that by the time you finish building brick by brick, what you built may already be behind.
Charlie laid out what the internal build actually requires, and none of it is trivial:
- Content licensing, including the right to use each source with AI, which is a significant ongoing job on its own.
- An enrichment layer of taxonomies, governance, and machine learning to make the data AI-ready and industry-specific.
- A way for agentic AI to navigate that enriched context.
- Compliance, including SOC audits and annual security reviews of every vendor.
- And then maintaining all of it permanently, while staying current as new models arrive.
His summary was blunt. Given the choice between hiring two or three people to maintain an internal system or adding people who find insights, most teams choose the insights team every time. The value of AI is in freeing humans to do what they are uniquely good at, not in standing up infrastructure.
Four questions to ask any enterprise AI
The session closed with a simple framework you can use whether you are evaluating a new tool or pressure-testing the one you already have:
- What is it allowed to read? Good, vetted, relevant content is what produces answers you can trust. Garbage in, garbage out.
- Can I verify it? And can you verify it in line, as you read, without launching a weeks-long process of checking claims after the fact?
- Will legal and IT sign off? Come with a strong business case, source permissions, visible citations, and zero retention.
- Is it current? The tool needs current data flowing in and the reasoning to know whether a source is from 2010 or 2026, because a model on its own does not even know what today's date is.
If you cannot answer all four, you do not have an intelligence advantage. You have a productivity tool that everyone else has too.
The takeaway
Fast is easy. Anyone can produce something that looks convincing in seconds. Trustworthy is the hard part, and in competitive and market intelligence, trust is what separates an AI you can demo from an AI you can defend.
The model was never going to be the differentiator. The content beneath it, the ability to verify it, the governance around it, and the discipline to keep a human in the loop are where the advantage actually lives.
Want the full conversation? Watch the recording, and read our whitepaper on the infrastructure behind trustworthy enterprise AI.
