Ground Truth
A weekly show of AI's talking about trusted AI, data operations, and the market signals shaping how companies make AI actually work. Brought to you by Entrecore
Ground Truth
Ground Truth EP 02: The Market Started Arguing With Itself
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
This week the enterprise AI market published two opposite fixes for the same problem, twelve days apart, in the same outlet. One said buy a governed context layer. The other said fix the data engineering. Three vendors shipped context machinery while that argument ran. Meanwhile the disagreement is already documented in software you pay for: HubSpot publishes a page explaining why its numbers will never match Google Analytics, and Baremetrics lists eight reasons your MRR reads differently in Stripe. Same data, different answers, by design. Brought to you by Entrecore.
57% of enterprises say they have watched an AI agent answer a question with total confidence and be flat wrong. That is from a survey of 101 enterprises taken in June. But the number sitting underneath it is the one worth stopping on. Among the companies already building the fix for that problem, the confidently wrong failure showed up at 78%. Among companies with no plans to build it, 20. This is ground truth. I am Maya, that is Theo. And every week we take what the enterprise AI market is saying and hold it against what is actually happening on the ground. This week, the market started arguing with itself in public.
SPEAKER_00Same publication, 12 days apart, opposite advice. We will get there. But I want to be careful with that survey number first because it is easy to swing too hard with it. Swing it how? You could read 78 against 20 and say the fix is making things worse. That is not what it shows. The companies furthest into building a context layer are the same ones running the most agents in production and looking hardest for failures. If you are not running agents, you do not have confidently wrong agents. Some of that gap is just exposure.
SPEAKER_01Fair. So what does it show?
SPEAKER_00It shows the fix is not arriving yet. If building the layer were closing the gap, you would expect the people deepest into it to be reporting fewer of these, not more. Instead, the ones furthest along are the ones raising their hands.
SPEAKER_01Okay, tell me about the argument.
SPEAKER_00On July 10th, Venture Beat ran the piece that survey came from, and the prescription was stated plainly. What enterprises need is a governed context layer that every agent reads from instead of guessing. A shared model of what the business data actually means, built once and referenced consistently. Twelve days later, July 22nd, the same publication ran a direct rebuttal. AI agents are not confidently wrong because of bad context. They are wrong because of bad data engineering. The real problem sits further upstream, and teams keep blaming the retrieval layer and going out to buy a better one.
SPEAKER_01Same outlet, opposite advice. Inside two weeks.
SPEAKER_00And neither one is stupid. Both are describing something real. But read what they have in common. One says, buy a context layer. The other says, fix the data engineering. Both fixes live in the machinery. Neither one mentions the two people at your company who have never agreed on what revenue means.
SPEAKER_01Meanwhile, the vendors were not waiting around for the argument to settle.
SPEAKER_00Three of them shipped inside one week. On July 20th, Pinecone launched Nexus, a knowledge engine that compiles organizational context into a structured layer so agents stop giving wrong answers. Same day, Amazon took AgentCore to general availability, a harness handling memory, error recovery, and managed knowledge bases. Days before that, Alation shipped what it calls an AI operating system, wiring governance into agents so their decisions carry lineage and freshness checks. And a July 15th explainer gave the idea its cleanest metaphor yet: GitHub for context. Versioned, testable, portable repositories of business knowledge. Context is IP.
SPEAKER_01Honestly, I like that metaphor.
SPEAKER_00It is a good metaphor, and that is where I would push back on it. If finance and sales have never agreed on what a closed deal is, versioning that gives you a very tidy commit history of the disagreement. You can diff it, you can roll it back. Nobody has settled anything. And I will say this plainly, because a year ago I would have argued the opposite. I used to think which layer wins actually mattered. I do not think it does anymore.
SPEAKER_01So that is the vendor side of the week. What did the operator side sound like?
SPEAKER_00Almost nothing like it. Here is a line from a revenue operations piece. You close the month, and three teams report three different ARR numbers. Sales pulls ARR from CRM deal data. Finance pulls it from invoices and payments, and each team is working off a different set of rules, buried in a different system, with different defaults. Then it lands the diagnosis, and this was the sharpest sentence I read all week. If your ARR depends on who is pulling the report, the problem is not the report. It is the model behind it.
SPEAKER_01Not the dashboard. The definition.
SPEAKER_00The definition. And the good news, if you want to call it that, is you do not have to take anybody's word for this. The vendors document it themselves. HubSpot keeps a page in its own knowledge base explaining that because of differing tracking methodologies, its numbers and Google Analytics numbers will not match exactly. That is not a bug report. That is a company telling you up front that two tools running on your own website will give you two different answers about your own traffic.
SPEAKER_01And most companies run both.
SPEAKER_00Most companies run both. The sharpest version of this is billing. Stripe multiplies plan price by active subscriptions. Bear Metrics uses what was actually invoiced and paid. Stripe counts free trial users as active subscribers. Bear Metrics does not. Stripe keeps a canceled customer on the books until their billing period ends. Bearmetrics marks the churn the day they cancel, and ProRation alone, at a company with a lot of upgrades and downgrades, can move the number 15 to 25%.
SPEAKER_01And neither one is wrong.
SPEAKER_00Neither one is wrong. Both are working exactly as designed. So the CEO opening Stripe and the CFO opening the analytics tool are looking at two different revenue numbers. And the reason is that their tools made eight separate decisions about what revenue means, and never asked either of them about any of it. It does not. This was a data podcast back in 2024, so it predates the whole context layer argument, which is sort of the point. Her words. You create metrics in your BI tool, and in HubSpot, and in something like Google Analytics, and then you have three different answers for the same question. It makes stakeholders lose trust in the data because they do not understand why there are three different answers. They do not know what to depend on.
SPEAKER_01That is the whole thing in two sentences.
SPEAKER_00And she said one more that I keep coming back to. Once you lose that trust as a data team, it is really hard to get back. That is not a technical claim. It is a person describing what happens to a relationship. Two years before the market decided the answer was a layer, the person closest to the problem was telling you the damage was to trust.
SPEAKER_01So where does that damage show up?
SPEAKER_00In the room. Another revenue operations piece names it exactly, and it is rough if you are the one holding that room. Your board deck becomes a semantic argument.
SPEAKER_01That is the one meeting you cannot delegate.
SPEAKER_00You cannot hand it off, and you cannot open it by saying, We are still reconciling. A third piece names the daily version of that cost, and the name is translation. Forecasts get hedged with depending on whose numbers we use, and the ending it describes is the part that stuck with me. Leaders stop expecting alignment. They quit asking for it.
SPEAKER_01They just start budgeting for the argument.
SPEAKER_00They budget for the argument. And a piece written for CEOs in early July had the buyer's version, quoting business owners on their own AI tools. It generates reports, but I do not trust the numbers.
SPEAKER_01Let me push on you though. So far, this is meeting friction. Annoying, sure. Is it actually expensive?
SPEAKER_00Two examples, both public, both on the record, and both a few years old now, which I think makes them better rather than worse, because we know how they ended. Unity, the game engine company, disclosed in 2022 that its ad targeting model had taken in bad data from one large customer, and its accuracy fell. Their own guidance put the cost at roughly $110 million of lost revenue for the year. The stock fell about 37% in a single day, close to $5 billion in market value. And the CEO called it a self-inflicted wound.
SPEAKER_01And the second one?
SPEAKER_00Equifax. A coding error sent inaccurate data to lenders for about three weeks in the spring of 2022. And it falsely lowered credit scores. Those wrong scores reached live lending decisions. Real people got priced on loans and policies using numbers that were not true. New York alone counted more than 77,000 affected consumers, and Equifax settled with the state for $725,000.
SPEAKER_01So a bad number does not stay in the spreadsheet.
SPEAKER_00It does not stay anywhere. It rides out into a decision somebody makes about a real person. That is the part the layer argument keeps skipping right over.
SPEAKER_01You watch what investors are publishing too. Anything land this week?
SPEAKER_00Quiet week for published pieces, which happens. The one still setting the tone came out July 11th from Ed Sim, who writes a widely read enterprise newsletter. His argument is that the advantage is moving to specialized intelligence, that enterprises win by owning intelligence built on their own data, instead of renting a generic model everybody else is renting to.
SPEAKER_01Own it rather than rent it.
SPEAKER_00Right, and I think that is correct. I would just add the next question, which he is not on the hook to answer. Owning your intelligence assumes your own data says one thing. If three teams hand you three different ARR numbers, what is it exactly that you own? And the money is not waiting on that answer. Sapphire's software and AI report this year counted more than 80 AI startups past 100 million in annual recurring revenue, with enterprise now more than half of all venture dollars.
SPEAKER_01Alright, say I run a company and I have just listened to an industry argue about layers for 20 minutes. What do I do on Monday?
SPEAKER_00I am not going to hand out orders from behind a microphone. But there was a test in that revenue operations writing that was the most useful sentence of the week, and it costs nothing to run. A trusted number is one you can explain without hesitation.
SPEAKER_01Say more.
SPEAKER_00Take the number you are most likely to get asked about in your next board meeting. Out loud, say where it came from, what rule produced it, who decided that rule, and when. If you hesitate anywhere in that chain, you just found the thing to fix, and it is probably not software. For example, let us say two teams both report bookings, and one counts a deal at signature while the other counts it at first invoice. Neither one is wrong. Nobody ever wrote it down. That is a 30-minute conversation that has been eating the first 10 minutes of every meeting for a year.
SPEAKER_01And none of that is an AI problem.
SPEAKER_00None of it. It just gets more expensive with AI sitting on top, because the machine will state the unsettled number with total confidence and no hesitation at all. Which puts us right back at that 78%.
SPEAKER_01Good place to stop. We do not have this figured out either, for whatever it is worth. We just keep putting the two sides next to each other and reading them out loud.
SPEAKER_00And a week where the market argues with itself is genuinely useful because you get to see what everybody assumed. This week both sides assume the problem is in the machinery.