|
The Agentic Enterprise
AK · Weekend Edition · 7 min read
|
|
Sunday, July 19, 2026
Everyone has the data. Almost no one has connected it.
The bottleneck in enterprise AI was never model capability. It is that most companies have oceans of data and no connected, meaningful layer for a model to act on. This week the field started calling that missing layer by its name: ontology.
Three signals this week pointed at the same quiet truth. A Rockwell survey found that 93 percent of manufacturers own the systems that generate their data, but only 23 percent have connected them. SAP spent more than a billion euros on models built to read structured business data, an implicit admission that the data, not the model, is the prize. And the analyst consensus hardened around a single diagnosis: the reason most enterprise AI pilots stall is not a weak model but a missing layer of meaning between the model and the mess. That gap, not the next frontier model, is where the next two years of enterprise value will be won or lost.
|
|
The missing layer isn't a better model. It's meaning.
| F |
or three years the enterprise AI story has been told as a race between models. Whose is smartest, whose codes best, whose is cheapest per token. This week the more useful story was about the layer underneath all of them, and how empty it still is. Start with the hard number: Rockwell Automation surveyed 1,560 industrial decision-makers across 17 countries and found that 93 percent run a manufacturing execution system, the software that captures what actually happens on the plant floor, but only 23 percent have fully integrated it across their ERP, PLM, and operational systems. The data exists. In three out of four companies, it is stranded. |
That pattern is not confined to factories. The analyst community spent 2026 converging on a single explanation for why enterprise AI keeps failing to scale. Gartner, presenting at its Data and Analytics Summit, projected that 60 percent of agentic analytics projects relying solely on tool-calling protocols will fail by 2028 for lack of a consistent semantic layer beneath them. EY has taken to calling ontologies "the missing layer in enterprise AI." Enterprises spent more than $684 billion on AI in 2025, and by one estimate over 80 percent of it failed to deliver the intended business value, not because the models were weak but because the models could not understand the business they were dropped into.
The word doing the work here is ontology, the layer that gives data shared meaning. When "account" means a checking account in one system, a trading account in another, and a cloud tenancy in a third, a model has no way to know which one you mean, so it guesses, confidently, and erodes trust. Connecting the data is step one. Giving it a common vocabulary the model can reason over is step two, and it is the step almost everyone skips. SAP's billion-euro bet on tabular models and even Google's stumble on Gemini 3.5 Pro's coding are downstream of the same reality: raw capability is abundant and getting cheaper, while connected, meaningful data remains scarce and expensive.
|
Capability is abundant. Connected, meaningful data is the scarce resource.
|
|
The Spearhead Take
This is exactly the problem we partnered with Databricks to solve, and it is why we did. The instinct when a pilot stalls is to buy a better model or another agent. The fix is almost always one layer down: get the data into a governed, connected foundation, then build the semantic layer, the ontology, that lets a model act on meaning rather than raw tables. It is less glamorous than a new copilot and it is where the return actually lives. Before you approve the next agent purchase, ask a blunter question: if we handed a capable model our data today, could it even find, join, and correctly interpret what it needs? For most enterprises, honestly answered, the answer is no, and that is the project.
|
|
|
The Obvious & The Overlooked
Three reads the market has. Four it is missing.
|
The Obvious
Most enterprise data is disconnected.
93 percent of manufacturers have an MES; only 23 percent have integrated it. PR Newswire
The money knows data is the prize.
SAP paid more than €1 billion for models built to read structured business data. SAP News
Most AI spend isn't paying off.
More than 80 percent of 2025's $684 billion in enterprise AI spend reportedly missed its intended value. Atlan
|
The Overlooked
Connecting data is only half the job.
Meaning is the other half; an ontology is what lets a model reason over the data it can now reach. EY
Agents fail at the semantic gap, not the model.
Gartner projects 60 percent of tool-calling-only agentic analytics projects will fail by 2028 without a semantic layer. Atlan / Gartner
The demo-to-production gap is a meaning gap.
Pilots pass because scope is small; they break at scale as semantic inconsistencies compound. SuperML
Better models widen the gap, not close it.
As capability gets cheaper, the disconnected, meaningless data layer becomes the binding constraint. HackerNoon
|
|
|
Moving Pieces
Five developments worth a CIO's attention.
Strategy
2026 is the year the field named the missing layer: ontology
The analyst consensus converged this year on why enterprise AI keeps failing to scale, and the answer was not model quality. It was the absence of a semantic layer, an ontology, that gives enterprise data shared meaning a model can reason over. Gartner projected that 60 percent of agentic analytics projects relying solely on tool-calling protocols will fail by 2028 without one. EY now calls ontologies "the missing layer in enterprise AI." The pattern behind the buzzword is real: when the same word means different things in different systems, a model retrieves confidently wrong answers and trust collapses. The enterprise read: the next competitive advantage is not a smarter model everyone can rent, it is a connected, well-defined data foundation only you have. That is buildable, and it is defensible.
Manufacturing
The factory floor has the data and can't connect it
Rockwell Automation's new report, surveying 1,560 industrial decision-makers across 17 countries, found that 93 percent of manufacturers run a manufacturing execution system but only 23 percent have fully integrated it across ERP, PLM, and operational systems. Manufacturers expect 42 percent of their processes to be AI-supported within a year, yet 43 percent admit they are not effectively using the data they already collect. The enterprise read: the scarce resource is not more AI, it is connected data. A model can only act on what it can reach, and in most industrial settings it still cannot reach most of the plant. The integration project you keep deferring is the AI project, and the survey is a rare hard number on just how wide the gap is.
Data
SAP bets a billion that the model should meet the data where it lives
SAP completed its acquisition of Prior Labs on July 17 and committed more than €1 billion, about $1.18 billion, over four years to scale it into a lab for tabular foundation models, AI built to read the structured rows and columns of enterprise data rather than free text. Its flagship model, TabPFN, was published in Nature. The enterprise read: SAP, whose whole business is the structured data inside company systems, is betting that the highest-value AI reads your tables, not your chat logs. But note the catch that ties back to this week's theme, a model built for your data still has to reach your data, and reaching it is the connection-and-meaning problem no acquisition solves for you.
Model Race
Gemini 3.5 Pro slips again, and Alphabet sheds $200 billion
Google's most powerful model, unveiled at I/O in May with a promised June launch, is now months behind schedule after its coding performance fell short of internal expectations. On July 16 the report of the delay knocked Alphabet down about 4.4 percent and erased roughly $200 billion in market value in a session, with no firm new launch date. The enterprise read: even at the frontier, model capability is lumpy and unpredictable, which is exactly why building your advantage on top of any single vendor's model is fragile. The durable moat is not the model you rent, it is the connected, meaningful data layer you own, which no competitor can download.
Deals
Enterprise AI funding keeps flowing to the boring, useful layers
The week's rounds skipped the flashy consumer plays and landed on infrastructure and governance. 8090 Solutions, which builds enterprise software with coordinated AI agents under human oversight, raised $135 million led by Salesforce Ventures. LeapXpert, which tracks enterprise communications for compliance, closed $180 million led by Riverwood Capital. Thira, founded by Apptio's co-founders, took a $21 million seed for enterprise AI. The enterprise read: capital is concentrating on the unglamorous middle of the stack, the connective and governance tissue that makes AI auditable and safe to run at scale. That is where budgets are being approved, and it is a useful signal for where the durable vendors will be.
|
|
On the Radar
Seven signals, sharpened.
| Research |
EY calls ontologies "the missing layer in enterprise AI." The semantic foundation, not the model, is where the firm says most initiatives quietly fail. EY |
| Research |
Gartner: 60% of tool-calling-only agentic analytics projects will fail by 2028. The cited cause is the absence of a consistent semantic layer beneath the agents. Atlan / Gartner |
| Manufacturing |
Manufacturers expect 42% of processes AI-supported within a year. Rockwell pairs that ambition with 43% admitting they underuse the data they already have. PR Newswire |
| Data |
SAP's TabPFN was published in Nature. Peer review, not a launch blog, underpins SAP's billion-euro bet on structured-data models. The Next Web |
| Deployment |
Microsoft's Frontier Company is spending $2.5B on 6,000 embedded engineers. The unit exists because integration, not model quality, is where enterprise AI stalls. Microsoft |
| Markets |
Alphabet lost about $200 billion in a single session. A model delay, not a revenue miss, wiped out roughly a Netflix-worth of market value. CNBC |
| Deals |
8090 Solutions raised $135 million led by Salesforce Ventures. Enterprise capital keeps flowing to the connective and governance layers, not consumer apps. Crunchbase |
|
|
Quick Hits
The wider field, one line each.
| Rockwell surveyed 1,560 industrial decision-makers across 17 countries for its MES report. Investing News |
| 44% of manufacturers now rank integration as their top MES buying requirement. PR Newswire |
| 46% of manufacturers reported a cyber incident in the past year, pushing security up the buying list. PR Newswire |
| Enterprise AI spend reportedly topped $684 billion in 2025, with over 80% missing intended value. Atlan |
| SAP completed the Prior Labs acquisition on July 17, 18 months after the startup was founded. SAP News |
| SAP is committing more than €1 billion, about $1.18 billion, over four years to the new lab. Tech.eu |
| Gemini 3.5 Pro was unveiled at Google I/O in May with a June launch that never happened. CNBC |
| Microsoft's Frontier Company is led by Rodrigo Kede Lima and staffed with 6,000 embedded engineers. Microsoft |
| 8090 Solutions' $135 million round was led by Salesforce Ventures; Thira raised a $21 million seed. Crunchbase |
| "Active ontology" is being pitched by data-catalog vendors as the 2026 default for enterprise AI. Atlan |
|
|
The Number
23%
Of manufacturers have fully integrated their MES, though 93% own one
Everyone has the data. Almost no one has connected it.
Rockwell's survey of 1,560 industrial leaders is the cleanest picture yet of the gap that defines enterprise AI right now. Barely one in four has connected the systems that generate its data, and connecting them is only the first step before the harder one, giving that data shared meaning a model can act on. Manufacturers expect 42 percent of their processes to be AI-supported within a year, but a model can only reason over data it can reach and understand. The integration you keep postponing is not a prerequisite for the AI project. It is the AI project.
|
|
Counter-Signal
Skepticism
"We need an ontology first" can become the new reason nothing ships.
The risk in this week's tidy consensus is that "ontology" becomes the next consultant's word, a reason to start a two-year, boil-the-ocean data-modeling project that never touches production. That failure mode is as real as the one it claims to fix. Enterprises have burned years and fortunes on grand master-data-management and enterprise-data-warehouse programs that aimed to define everything before doing anything, and most of them died in committee. Calling the same instinct "ontology" does not make it ship faster.
The discipline is to build the connected, meaningful layer in the narrow slice where a specific decision needs it, prove value, and expand, rather than modeling the entire enterprise up front. Meaning is essential, but it is earned use case by use case, not decreed in a data dictionary. The teams that win will not be the ones with the most elegant ontology. They will be the ones who connected just enough data, gave it just enough shared meaning to make one high-value decision reliably, and then did it again. Buy the thesis. Refuse the boil-the-ocean version of it.
|
|
From the Field
For three years the enterprise AI conversation has been about models. This week the more honest one surfaced, and it is not about models at all.
Which model is smartest, which codes best, which costs least. It is a comfortable conversation because it is someone else's problem to solve. You just pick a vendor and wait for the next release. The harder truth is that almost every company is sitting on the raw material of extraordinary AI and cannot use it. The Rockwell number says it plainly. Ninety-three percent of manufacturers have the data. Twenty-three percent have connected it. The rest have oceans of information trapped in systems that do not talk to each other, and no shared vocabulary to make sense of it if they did. You can rent the smartest model in the world and point it at that, and it will still guess, because it has nothing solid to stand on.
Here is what I would sit with this weekend. The advantage in the next phase of AI will not come from the model, because everyone rents the same models. It will come from the one thing a competitor cannot download: your data, connected and given meaning. That work is slow, unglamorous, and easy to defer for another quarter of pilots. It is also the whole game. The companies that win will be the ones that stopped shopping for intelligence and started building the foundation the intelligence needs.
Everyone has the data. Win by connecting it.
Let's get to production,
AK
|
|
|
|
Two conflict disclosures apply this edition. Databricks is referenced in The Spearhead Take as a Spearhead partner; the edition's connected-data and ontology thesis aligns with that partnership, and readers should weigh the take as coming from a firm with a commercial interest in helping enterprises build exactly this layer. Anthropic is a Spearhead technology partner and its Claude model produced this edition under human editorial direction; it is not a subject of this edition, and the model-race coverage runs through a competitor's (Google's) stumble. The Rockwell figures (93 percent own MES, 23 percent fully integrated, 42 percent processes AI-supported within a year, 43 percent underusing data, 44 percent rank integration top, 46 percent cyber incident, 1,560 respondents, 17 countries) are from a single vendor survey and should be read as directional and self-interested. The Gartner projection (60 percent of tool-calling-only agentic analytics projects fail by 2028) and the "$684 billion 2025 spend, over 80 percent missed value" figures are carried via a data-catalog vendor's (Atlan's) summaries rather than a primary Gartner document; treat them as reported and directional. SAP's investment figure (more than €1 billion, about $1.18 billion, over four years) and Prior Labs' details (TabPFN in Nature) are from SAP's announcement and deal coverage. The Alphabet market-value loss (roughly $200 billion, about 4.4 percent) is a market reaction reported by CNBC around July 16, not an audited figure. The July funding rounds (8090 Solutions $135M, LeapXpert $180M, Thira $21M) are from funding trackers and were not independently confirmed; Microsoft Frontier figures ($2.5 billion, 6,000 engineers) are Microsoft's own. The ontology-as-default framing is a forecast and a vendor-advanced thesis, not a settled outcome, and the Counter-Signal argues the opposite risk. No emoji, hashtags, or exclamation marks were used. No India-domiciled outlets were used. All editorial decisions are human-directed.
|
|
The Agentic Enterprise
Know more about AI than 95% of your peers. By 7 AM.
A daily AI intelligence briefing for enterprise leaders, published by Spearhead. We build AI systems that work. Strategy. Engineering. Production. Outcomes.
© 2026 Spearhead. All rights reserved.
|
|
