Last week, our VP of Product, Grant Spradlin, joined Hannah Erlsten of Prophecy, and Raji Narayanan of CData, for an panel discussion with Stephen Faig, Director of Enterprise AI World, about what it actually takes to make enterprise AI work.
The takeaways from were practical, honest, and for anyone who has watched an AI agent confidently return a wrong answer a bit of a relief. The core message? The problem usually isn't the model.
Here is what our team took away from the conversation.
The real barrier to reliable enterprise AI isn't the model
Most enterprise AI projects start with the same assumption: get the model access to your data and things will work. It turns out that assumption is wrong, and it is costing organizations real money.
Grant opened with a story that landed immediately. Imagine your legal team asks an AI agent: "What is the liability cap with Acme Corp?" The agent has access to the contract. It reads through the documents, finds the agreement, and confidently answers: eighteen months. There is just one problem: the contract says twelve months.
So where did eighteen months come from? The contract was flattened into plain text before the agent ever saw it. When that happened, the numbers in the fee schedule ran into the liability clause. The model had no way to tell them apart, so it stitched together an answer that did not exist anywhere in the document.
"Access gets the model text," Grant explained during the panel, "but context tells the model what the text means."
This is the core problem. When content is stripped of its structure (headings, tables, columns, reading order) the model loses the meaning that holds it together. Grant illustrated this with another example from his own preparation for the webinar. He gave an AI productivity tool the link to the registration page. The speakers were listed on the left, the moderator on the right. On screen, it was obvious. But when the page was converted to text, the column structure disappeared and the word "moderator" landed right before Grant's name. The tool told him, with full confidence, that he was the moderator.
"The model was fine," Grant said, "but the structure was lost before the agent ever even saw the page."
This distinction, between access and context, was the thread the entire panel kept coming back to. Raji shared benchmark data from CData Labs that made the scale of the problem concrete. When AI agents ran enterprise tasks without business context, only 0.25% runs succeeded. With context, 82% did. From an accuracy standpoint, approaches that gave access without context scored 75% at best. With a context-aware connectivity layer, accuracy reached 98.5%.
"It's rarely ever the model that is the problem," Raji said. "It always goes back to the data and the context."
How structure, governance, and data quality work together
All three panelists agreed on the solution: build context as a foundation, not as an afterthought.
Turn your content into infrastructure
Vertesia's approach is to treat context as infrastructure. Rather than copying documents into prompts or relying on similarity search alone, the platform processes every piece of content (PDFs, presentations, spreadsheets, images) through Semantic DocPrep, a patented process that preserves reading order, headings, tables, and figures so the model sees content the way a person does.
From there, business-specific fields are extracted with citations attached, and everything is indexed for multi-vector search. The result is not a pile of text chunks. It is a structured, reusable context layer that every agent and workflow can share.
"We didn't switch models. We didn't write a different prompt," Grant said, describing the difference a context layer makes. "We used the same model, the same question. The only difference here is the context."
He also drew a clear line on how organizations should not handle their documents: "You don't move files around." Pulling content out of the systems where it has access controls and relationships, just to stuff it into a prompt, is not an enterprise approach. "Getting the answer right once in a demo is one thing," Grant added. "Getting it right every day is operations."
Make every AI answer verifiable and repeatable
Hannah focused on what she called "the last mile" which is the gap between an AI answer and work your business can actually use. Even with the right context and clean data, AI-generated results are only useful if someone can check them.
Her four questions for evaluating any AI tool:
- Can it handle real, messy data?
- Can you check its work?
- Does it get smarter over time, or start from scratch every run?
- Will it produce the same result next month?
The last question matters more than people expect. If business rules get regenerated every time the model runs, this month's results will not tie out to last month's and your auditors will want to know why.
Hannah's closing thought in the lightning round was pointed: "The most valuable context in your company probably isn't in a system yet. It's probably in the head of an analyst who knows how to exclude the duplicate vendor IDs before month-end close. Every time AI is doing that data work, that knowledge is either getting captured or it's clocking out at five o'clock with your business data analyst."
Keep your data fresh, connected, and governed
Raji's message was about the plumbing underneath. Whether an agent uses RAG, a knowledge graph, or agent memory, those approaches are only as good as the data work behind them. And that work should be done once and shared, not rebuilt for every application.
She also flagged a freshness problem that is easy to overlook. In older reporting environments, a data lag of hours or even a day was acceptable. AI agents do not have that tolerance, especially if they are expected to act on information, not just report on it. Stale data leads to wrong answers, and wrong answers erode trust.
"Do the data work once and then share it across every agent," Raji said, "so you have one governed path to your systems, so every answer follows the same rules."
What to do in the next 90 days
The panel closed with a practical exercise that all three speakers endorsed in some form: before evaluating any vendor, identify a handful of real questions your business needs to answer. Run your current tools against them. Use what you learn to define the gaps.
Grant put it simply: "Take a real world process. Take your content, and see what kind of process you can put that content through. And you're going to learn a lot more from your content and your context and from those vendors in minutes or hours than you'll learn in any proof of concept or long pilot."
Then do the foundational work: agree on definitions, connect the right systems, establish a governed path, and measure the cost per correct answer.
The organizations that will get the most from enterprise AI are not the ones with the most advanced models. They are the ones that do the harder, less glamorous work of building context they can trust.
As Grant put it in his final lightning round answer: "Your organization's enterprise content is the proverbial fuel, the literal fuel for AI, and provides the context layer for trusted AI agents."
Watch the full webinar recording here.