The Missing Data Layer in Business AI: Capturing the Conference Room

For decades, companies have been very good at storing the numbers and surprisingly bad at storing the thinking around the numbers. ERP systems can tell you what was invoiced, what was shipped, what a customer bought, what the gross margin was, and what hit the P&L. But when a group of experienced people gets into a room and decides what those numbers actually mean, most of that interpretation has historically walked right back out the door with them.

That may be the part of AI that matters most. Not the chatbot. Not the prompt box. Not even the ability to automate a task that a person used to perform. The bigger opportunity may be the ability to capture the narrative that companies have always created around their structured data and turn that narrative into something the business can reuse.

That was the central idea running through a conversation between Tim Rodman and longtime ERP executive Andy Nunez. Nunez described AI as being as revolutionary as the database itself, which is a pretty big statement from someone who has spent much of his career around ERP systems. His argument was that databases gave companies a way to institutionalize operational data. AI now gives them a shot at institutionalizing the interpretation of that data.

But there is a catch. A company cannot just dump every ERP table, every Teams transcript, every CRM note, every email, every sales call, and half the Internet into a frontier model and assume something magical will happen. The hard problem is increasingly looking like a data architecture problem: ingestion, context, summarization, source quality, structure, and the ability to breadcrumb an answer back to the evidence that produced it.

In other words, the interesting part is not merely getting an answer from AI. It is giving AI enough of the business to understand the question, then being able to ask the very human follow-up: Why?

From Archaeology to ERP to AI

Nunez did not exactly take the normal route into ERP. He grew up in Plano, Texas, attended the University of Texas at Austin, and studied art history and archaeology. He was a Mayanist and spent time at archaeological sites in Central America, which, as he joked, naturally prepared him to become a SaaS software CEO.

The career turn came through friends from college. Eric Frank and his partners were starting a Solomon consulting practice and wanted Nunez to join them. His initial reaction was that a “Solomon practice” sounded a little too much like a cult, so he passed. They called again after the business had customers. He passed again. When they called a third time and offered to start an office in New York, he finally said yes.

That decision pulled him into the mid-market ERP world for the long haul. He helped build the New York and New Jersey offices of Nextech, later moved to SWK, and spent roughly a decade there. During that run, the company went public, Nunez helped build the X3 practice into a multimillion-dollar business, took on Acumatica, briefly took on NetSuite, and helped grow the team to more than 200 people.

Then came Scanco. Nunez was recruited as CEO as the owners prepared for an eventual exit. The business had a five-year plan and sold to private equity in three years, ahead of plan. He stayed on afterward for acquisitions and the transition to the new ownership group.

His attempt at taking time off was less successful. After leaving, Nunez told his wife he was going to take all of 2025 off. By his telling, that plan lasted about two hours.

Consulting work led him back to people from earlier stages of his career, including Rob Kofsky, whom Nunez had originally hired years earlier when ERP implementations involved literal servers and coaxial cable. Kofsky had gone on to work in AI long before the current wave. Their conversations eventually came back to a problem Nunez had seen firsthand in private equity: traditional diligence could collect a tremendous amount of data without necessarily producing a deep understanding of the operating business.

That became the seed for BDE, the Business Development Engine. The original idea was focused on private equity diligence, but the underlying problem turned out to be much broader. How can a company combine what its systems know, what its people know, and what the outside world knows into a more useful narrative about the business?

The Numbers Were Never the Whole Decision

Executives have always made decisions with more than data. The numbers might start the conversation, but somebody still has to decide whether a pattern is meaningful, whether a customer is healthy, whether a salesperson’s forecast is believable, whether a product problem is temporary, or whether something has fundamentally changed.

Nunez used the “conference room” as the shorthand for where that interpretation happens. The room can be physical or virtual. It could also be a hallway conversation, a Teams chat, an email thread, or a quick exchange between two people. The important thing is that multiple people bring their experience to the same set of facts and create a narrative around them.

A spreadsheet by itself does not do that. It can contain the data and even the logic of the person who built it, but the real business meaning often appears when other people start questioning it. Why is that customer slowing down? Why is this deal still in the pipeline? Why did margin fall? Why does the P&L say one thing while the operating team is saying another?

Rodman framed it as the difference between the “boxy, rectangular, spreadsheet-shaped” data that software has always handled well and the loose, free-flowing commentary that historically lived in people’s heads. ERP captured the first category. The second category was much harder to turn into something a computer could analyze at scale.

AI changes that because the commentary is now machine-readable. A meeting can be transcribed. A call can be summarized. A Teams conversation is already digital. Strategy documents, CRM notes, customer feedback, emails, and other forms of business context can all become inputs alongside traditional structured data.

That does not mean every source should be treated equally. Nunez gave a simple example: he would trust the sales number coming from Acumatica or another ERP system more than somebody’s opinion about the company on Reddit. Both might provide useful context, but they have very different levels of authority.

The source therefore becomes part of the answer. If AI says a customer is at risk, an executive should be able to see whether that conclusion came from declining order volume, a CRM note, a support conversation, an external market event, or some combination of those things.

AI Does Not Eliminate the Conference Room

One of the easiest mistakes is to hear all of this and assume AI makes the human discussion less important. Nunez’s argument was almost the opposite. The discussion is still where people challenge assumptions, bring experience to the numbers, and decide what they believe. AI simply makes more of that discussion reusable after the meeting ends.

Historically, a company might document the final decision but lose the path that produced it. Somebody writes meeting notes. Somebody updates the forecast. Somebody changes a plan. But a lot of the reasoning remains scattered across the people who happened to be there.

With AI, that reasoning can become part of the company’s memory. An agent could eventually remind an executive that the team discussed a similar customer six months earlier, that the group had previously discounted a particular metric, or that a strategic assumption had already been debated and revised.

That changes the value of the next meeting. Instead of spending half the time reconstructing what everyone thought last month, the group can begin from a better representation of the previous conversation and move forward from there.

The informal stuff may be just as important as the formal meeting. Companies have always said that some of the best information comes from the water cooler or the hallway. Today, much of that conversation happens in Teams, email, text, or other digital channels. That makes more of it technically available to become context.

Of course, this gets uncomfortable pretty quickly. Nunez openly acknowledged the “Big Brother” feeling that comes with comprehensively capturing business conversations. Privacy and data security are real issues, and he said BDE was already encountering them with early customers.

His view was that companies will have to work through those boundaries rather than pretend the issue does not exist. Businesses already went through versions of this transition when call centers began recording calls. AI raises the scale and reach of the issue, but the underlying tension between useful institutional knowledge and employee privacy is not completely new.

COVID Accidentally Created Better Raw Material for AI

There is an ironic connection between remote work and business AI. The technology for digital meetings existed before COVID, but many companies did not use it pervasively. Once remote work forced tools such as Teams into normal business life, a much larger share of company conversation became digital by default.

That means the cultural shift may have created a data source that companies did not realize they were building. A discussion that once disappeared into a conference room is now much more likely to leave behind a recording, transcript, chat history, summary, or some other digital artifact.

For Nunez, that matters because anything that is digitized can potentially be interpreted. It does not mean a company should ingest everything indiscriminately. It does mean the raw material now exists in a way that it often did not before.

Rodman connected that to transparency inside a company. He remembered quarterly reviews at Acumatica under then-CEO John Roskill as being almost uncomfortably transparent in the amount of company information shared internally. The AI version of that philosophy is not making the company’s data public. It is creating something closer to a public conversation inside the virtual walls of the company, where useful context is less likely to remain trapped in individual people.

Data Informed Is Still Different From Data Driven

More data does not remove human judgment. Rodman has long preferred the phrase “data informed” over “data driven” because the latter can imply that the machine or the metric is running the show and the human is subordinate to it.

He brought up Jeff Bezos’s explanation of Amazon’s HQ2 decision as an example. The idea was to collect a huge amount of data, study it deeply, immerse yourself in it, and then ultimately make the decision with intuition.

Nunez thought that fit BDE’s own phrase: turning data into direction. Data may be objective. Direction is not. The gap between the two contains human intent, experience, instinct, preference, and judgment.

Knowledge can change instinct without eliminating it. An executive may come into a decision leaning one way, see new information, and change course. Or the executive may see the information and still disagree with the recommendation. Either way, the final decision remains subjective.

The useful role for AI is to bring more relevant knowledge to that decision. It can make the executive better informed without pretending the executive has become a deterministic output of the dataset.

A Sales Pipeline Shows Why Context Matters

Nunez’s sales pipeline example makes the difference pretty obvious. A CRM can say a deal is going to close next month. It can display the forecast beautifully. It can even calculate probabilities and roll everything into a nice pipeline chart.

But the sales manager may know that the same salesperson has been saying “next month” for the last six months. That is context.

So the team gets into a pipeline meeting and starts asking questions. What changed? Why is the salesperson more confident now? Did the executive sponsor go cold? Is there a new obstacle? What should the salesperson try next?

The CRM record usually does not contain the full interpretation created by that conversation. The forecast might get updated, but the discussion that led to the update often disappears.

AI creates the possibility of keeping it. The next time the same customer or a similar situation comes up, the system could surface what the team said before, what happened afterward, and perhaps an outside event the team never noticed.

That last part is where the combination gets especially interesting. An LLM may know a lot about the outside world while knowing almost nothing about the specific business using it. Rodman referenced a line from Rob Collie’s book that LLMs have “a PhD in everything but your business.” The job is to connect that broad outside knowledge to the company’s inside knowledge without losing track of which is which.

A customer may have purchased during one market condition and then purchased again years later under a similar condition. The internal sales team might never notice the pattern. An AI system that can connect external events, transaction history, CRM activity, and prior internal discussion at least has a chance to surface it.

Then the sales manager still decides what to do. That is the recurring point. Better context does not eliminate direction. It improves the information available before a human chooses it.

The Hard Part Is Not the Million-Token Prompt

This is where the conversation moved from executive philosophy into architecture. “Comprehensive” sounds great until somebody asks how an LLM is actually supposed to consider years of ERP transactions, thousands of meetings, CRM history, strategy documents, customer conversations, and external market information at the same time.

Rodman’s concern was the context window. He used roughly a million tokens as the working example and pointed out that even if the window gets much larger, a real company can generate vastly more information than any single prompt can reasonably hold.

So the solution cannot simply be “put everything in the prompt.” That is the AI equivalent of solving every database problem by creating a bigger spreadsheet.

Nunez described the current stage as the DOS-prompt era of AI. People are still thinking in a very linear pattern: input something, get an output, then move to the next prompt. That can be useful, but it is not the architecture required for a company-wide memory and decision layer.

The more important layer is ingestion. BDE was being built to pull information continuously from multiple sources, partition it efficiently, preserve the context required for the current initiative, and then bring the right material forward when a question is asked.

BDE was also using RAG and multiple frontier models rather than assuming one model should do everything. Nunez said the models were moving quickly enough that the system could bring in different models where they were most appropriate, while keeping the work focused on the specific initiative and its context. He was also careful to admit that his development team understood some of the technical details better than he did.

Rodman described the structure as something like a spider web. The LLM does not need every piece of company information loaded into active context at once. It needs a path through the information: enough higher-level structure to know which branch to follow, then the ability to dive into the detailed source when needed.

That starts to look a lot like old-fashioned data architecture again. How should information be joined? What is the relationship between one object and another? Which summary points to which underlying detail? Which source is authoritative? What is the smallest useful representation that allows the next layer to find the right evidence?

That is why this is not only an AI problem. It is a data structure problem with an LLM sitting on top of it.

Build Little Strategy Documents Inside the Database

One of the most useful phrases in the conversation was “strategy documents in your database.” Nunez compared the process to what a company already does when it takes a long strategy discussion and turns it into a shorter document that people can actually use later.

The summary is not valuable because it replaces the original discussion. It is valuable because it makes the original discussion easier to retrieve.

Imagine a meeting transcript with 20,000 words. An AI system does not need to reread every 20,000-word transcript from every meeting every time somebody asks a question. It can create a smaller representation of the important decisions, assumptions, themes, and unresolved issues from each meeting.

Those summaries become another layer of company data. The AI can scan the higher-level representations, decide which meetings appear relevant, and then drill back into the detailed transcript when more evidence is needed.

This also means AI embedded in individual applications is not necessarily competitive with a broader AI layer. If Gong can summarize its own sales calls well, a cross-company system can use that summary as an input instead of doing the entire job again. The same principle could apply to ERP systems, meeting tools, CRM applications, and other specialized software.

Each application can “chew the food first,” as Nunez put it. The company-wide layer can then combine the useful output from those separate islands with other data that no individual application can see.

Breadcrumbs May Matter More Than Confidence Scores

The moment AI starts influencing executive decisions, “trust me” is not a sufficient explanation. Nunez kept returning to the need to breadcrumb the answer back to the data.

If the system highlights a problem in the P&L, the executive should be able to trace the recommendation back to the relevant line in the P&L. But a comprehensive system might also point to a strategic goal document, a prior meeting, a customer conversation, and a correction that an executive gave the system three weeks earlier.

That history matters because business interpretation changes. A number does not become meaningful merely because it exists. People apply context to it, and sometimes they explicitly tell the system that an apparent conclusion is wrong because the system is missing a fact that experienced people know.

That correction should not disappear. It should become part of the institutional context available the next time a similar question appears.

This is also where data quality and source weighting become critical. BDE’s approach included rating the quality of data, scorecarding it, and preserving the path behind an answer. A model can sound equally confident when it is reading a financial number or a random online opinion. The architecture has to know that those things are not the same.

AI Inside ERP Is Useful, but It Knows Only Its Island

Rodman had just spent a full day at Acumatica Ascent looking at the AI capabilities coming in Acumatica 2026 R2. He found the work interesting and saw plenty of use cases, but he was still wrestling with whether the most interesting AI would ultimately live inside ERP itself.

Nunez’s answer was basically: useful, yes; comprehensive, no. He compared a generic AI with little company context to an AP clerk who started yesterday. The new employee may be smart and capable, but the person who has been in the company for years understands far more of the business around the transaction.

Adding ERP data makes the AI more experienced, but it still only knows what the ERP knows. Nunez described AI operating on a limited ERP dataset as a form of advanced BI. He immediately qualified that by noting it can go further than traditional BI because it can take action, but the core limitation remains: limited data produces limited context.

There are still excellent reasons to put AI inside ERP. Nunez referenced the goal of using AI to reduce or even eliminate parts of the finance close process. Rodman separated that kind of use case into what might be called the “worker” side of AI: automate something a person already does, streamline data capture, or remove unnecessary manual steps.

That kind of automation makes perfect sense inside the line-of-business application. ERP has the necessary local context for many accounting and operational tasks. Gong has the necessary local context to summarize a sales call. A meeting platform has the necessary local context to create a useful meeting summary.

The direction side is different. If the question is “When is this customer likely to buy again?” the ERP may not know that the customer was just acquired by private equity, hired a new CFO, changed strategy, had a difficult support conversation, or operates in a market experiencing a new external event.

That requires the AI to cross the islands. In that sense, the broader decision layer starts to feel more like business intelligence than ERP because its job is to grab useful information wherever it lives rather than force everything into one transactional application.

This May Be ERP All Over Again

Nunez kept returning to an analogy he admitted he was probably overusing: spreadsheets versus relational databases. Spreadsheets were useful before relational databases, and they remained useful afterward. ERP did not kill Excel. If anything, ERP systems eventually started embedding spreadsheet-like experiences back into the applications.

AI may evolve the same way. Prompting will not disappear. Application-specific AI will not disappear. ERP will not disappear. The new layer gets added because the previous layer cannot solve every problem by itself.

The spreadsheet is good at personal interpretation. The relational database is good at creating shared structure around business processes. AI can be good at interpreting a much broader mixture of structured and unstructured information, but it still needs a framework around that information if the company expects repeatable results.

Nunez described the end state as “reforming ERP” more comprehensively. Traditional ERP brought together accounting, sales, inventory, warehouse, support, and other operational functions. The next layer could bring in those systems plus the company’s narrative, strategy, customer conversations, external signals, and other forms of context that have historically sat outside the ERP database.

That is a much bigger idea than putting a chat box on an ERP screen. It is also a much harder problem.

And yes, dark mode still got a surprising amount of enthusiasm. Rodman joked that after sitting through all of the AI permutations at Acumatica Ascent, the feature that landed hardest with him was dark mode. Nunez was equally enthusiastic. Sometimes the revolutionary technology has to compete with the feature users can see immediately. That seems about right.

BDE Started as a Better Private Equity Data Room

BDE is Nunez’s attempt to make some of these ideas concrete. The team began coding the product the prior December with private equity as the initial use case.

The problem was traditional diligence. A buyer could get the P&L, churn spreadsheets, CRM access, and a data room full of material, then spend a relatively short diligence period interviewing people to figure out what the numbers actually meant.

Nunez wanted the information to be dynamic rather than dead on arrival. Instead of evaluating only the snapshot created during a roughly 45-day diligence process, the system could include years of context about leadership, product lines, technical debt, customer health, and other dimensions of the company.

The same information could continue to be useful after the acquisition. The buyer could establish a baseline during diligence and keep measuring the company against that baseline rather than throwing away the context once the transaction closed.

BDE then applied scoring pillars to that narrative. The original framework was built around major dimensions relevant to SaaS businesses, including areas such as customer health, leadership, and technical debt. The measurements could be changed to match the buyer’s investment thesis.

The scoring itself was not supposed to end the discussion. A business owner might look at the system’s rating and say it is wrong. Good. Now the buyer and seller have something specific to debate, and the evidence behind the score can be inspected.

That is the same data-informed idea again. The system creates a more comprehensive starting point. Humans still negotiate, disagree, and decide.

The Same Engine Can Segment Existing Customers

Once BDE was pulling all of that information together, the private equity use case stopped looking like the only use case. Nunez’s team began working with resellers on customer segmentation, particularly how to re-penetrate an existing customer base more intelligently.

Traditional segmentation often uses the metrics that are easiest to consume. Customers become A, B, or C based on revenue, accounts receivable, complexity, frequency of contact, or another simple matrix.

But those metrics do not necessarily explain why the customer is likely to buy. A customer may have just been acquired. A new CFO may have appeared on LinkedIn. The company may be changing direction. Those signals can matter even though they are not fields in the ERP customer table.

Nunez’s argument was that real segmentation should be dynamic. The system can combine internal customer data with external signals and continually update the customer’s profile or propensity to buy.

That can lead to playbooks at the individual-customer level rather than only the industry level. Instead of saying every manufacturing customer gets the same campaign, the company could shape the next action around what is actually happening with that particular account.

The idea naturally moves up the funnel. If the same signals help explain existing customers, they may also help identify prospects in the outside market. Tools such as ZoomInfo or Clay can make that process more efficient, but Nunez did not describe them as mandatory. They are additional prepared inputs into the larger context layer.

Cross-Company Intelligence Creates a Bigger Privacy Problem

The longer-term possibility gets even more interesting—and more sensitive—when multiple companies contribute to the same analytical platform. Nunez used HVAC companies as the example because private equity has been actively consolidating that kind of business.

One HVAC company gives a buyer a deeper view of that one company. A hundred HVAC companies create the possibility of anonymized benchmarking across a segment.

That could help answer questions that are difficult to answer from a single company’s history. Which types of companies are performing better? Does the buyer’s investment thesis match what the broader segment appears to show? Are there outside conditions associated with stronger performance?

But anonymization and privacy are not side details. Nunez repeatedly came back to the need to keep the information private and secure. A marketplace or benchmarking layer only works if the participants can trust how their information is isolated, aggregated, and used.

The opportunity therefore expands at exactly the same time as the governance problem. More context can create better analysis. More context also means more sensitive information to protect.

The ERP Channel May Be a Natural Way to Deliver It

Nunez’s instinct for taking BDE to market is familiar to anyone who has spent time around mid-market ERP: use the channel. Rather than reinventing the ecosystem, he sees an ERP-adjacent reseller or consultancy as a natural route because one partner can apply the capability across multiple clients.

That does not rule out selling directly to companies or working with private equity firms. His description was essentially to stay focused while still casting a wider net as the product evolves.

He also sees the platform itself becoming channelized. A central publisher could provide the engine while partners verticalize the scoring, data sources, playbooks, and use cases for the industries they understand.

That is another reason the architecture matters more than the individual prompt. An ERP reseller, a private equity firm, and a SaaS operator may all ask different questions. The reusable part is the machinery that can ingest their data, organize the context, preserve provenance, and apply a business-specific framework on top.

The Bigger AI Breakthrough May Be Company Memory

The flashy version of AI is a model that gives an impressive answer. The more consequential business version may be a company that gets progressively better at remembering why it believed what it believed.

That includes the hard numbers. It includes the ERP transactions, the P&L, the CRM records, the pipeline, and the operational data companies have been storing for decades.

But it also includes the narrative. What did the leadership team think at the time? What did the salesperson say? Which assumption did the CFO reject? What strategy was the company pursuing? What did the customer complain about? What external event was occurring? Which source should be trusted most?

That is where the conference room becomes a data source rather than a place where data goes to die. The human discussion does not disappear. It becomes part of the institutional record.

Rodman’s own skepticism about the “comprehensive” part is still important. The architecture is not finished. Context windows are finite. Data has to be partitioned. Summaries can lose detail. Sources have different levels of quality. Privacy and security are difficult. And nobody should confuse a fluent AI response with a proven business conclusion.

But the direction is easier to see once the problem is framed correctly. The next stage of business AI may be less about teaching an LLM how to talk and more about teaching a company how to structure what it already knows.

Then the machine can do something genuinely useful: bring the right evidence back into the room. The people still decide what to do with it.