She Never Sees a Row of Your Data. The Answers Still Clear 95%.

Inna Tokarev Sela is the founder and CEO of Illumex, a company she started in 2021. The homepage line Nathan Labenz reads back to her is the whole pitch: get your data speaking the way your employees do. She came up through data products at SAP and Sisense, trained as a mathematician, and has been in love with graphs since she was coding them by hand in MATLAB on her first degrees. She has collected five degrees in all. The golden age of startups, she jokes, was a fine year to begin.

The original investor pitch was not that sentence. She walked into deep-tech meetings talking about automated context and reasoning for agents, and after two or three attempts it was obvious nobody knew what she meant. She switched to plain language: an application-free future for knowledge workers, and self-service analytics over structured sources. GPT-3 was only a glimpse then. What convinced her the timing was right was the 2017 wave of semantic models and the Hugging Face ecosystem. Graphs were context. Semantics were content. Together they could finally become a fabric that connected data to people and workflows.

The problem she wanted to kill is older than the models. After all the money companies spend on data and analytics practice, she says, the majority of business decision-making is still based on guesstimation. She does not blame the analysts. Humans simply have too many questions, and until recently there was no helper that could answer them with data automatically. Illumex is her bet that the helper can live in Slack, Teams, and the tools people already have open, rather than in another dashboard they have to learn.

Labenz has been talking to companies that generate apps on demand, the Bolt and Lovable vision of a software supernova: custom, personal, even disposable interfaces for everyone. Illumex sounds like the opposite, a collapse of interfaces into one point. Tokarev Sela does not treat the two as enemies. Tech people love new tools. Business people have a day job, and the day job is not learning software. Multiple apps, multiple logins, and constant context-switching are too much. Analytics can still live inside Salesforce. Niche agents can still exist. Eventually, she thinks, it boils down to a launcher. You ask a question or narrate a task, it happens, and you do not remember which clicks, which system, or which workflow made it so. The workstation recedes into the background.

Labenz wants that future for himself: untethered from the chair, outdoors, still able to pull information and take digital actions. OpenAI's advanced voice mode is not there yet, he says, for mundane reasons like missing features and thin context. Tokarev Sela's demo is more ordinary and more immediate. On her phone, in Slack, she asks how many products are in stock after a low-stock ping. Autopilot mode, "let Omni decide," matches the prompt to a semantic ontology that Illumex built from warehouses, databases, BI tools, and SAP. Nothing in that chain is hidden. The answer cites a semantic entity you can open.

Click through and you get lineage, attributes, the business ontology, and the SQL. "This is a platform for control freaks." Most people take the number and act. A minority want the definitions and will dig to the code. She wears Ray-Bans with a speaker in them and finds the lack of headphones liberating. The same idea applies to a data copilot: stay in Teams or Slack, chat with something that feels like an analyst, and get factual. She keeps using the word augmentation, not automation. Everyone wants smarter decisions. This is how you stop guessing.

Labenz's sketch of the architecture is the right one, and she confirms it. You cannot walk into a hybrid enterprise cold. Illumex first builds a corpus of how each industry works: terminology, processes, metrics, dependencies, from IoT and manufacturing through insurance, pharma, and finance. That knowledge is a business ontology, architecturally a knowledge graph of semantic embeddings, relational models mixed with semantic ones. Onboarding a customer does not mean reading their rows. The system looks only at metadata: schemas, query logs, and application APIs.

The companies she serves are regulated and data-heavy. A typical stack mixes on-prem Oracle, SAP, Teradata, and Vertica with Redshift, Snowflake, Databricks, Tableau, and Power BI. There are dozens of systems per department and no single source of truth. Some customers have millions of tables. Illumex brings the industry ontology as a benchmark for what a single source of truth should look like, then retrains it on that customer's metadata. The job usually takes a few days, depending on size. There is no manual labeling, no moving data into a vector database. The result is a virtual layer: a semantic single source of truth over a federated or centralized estate, including the knowledge that order ID in one system is vendor ID in another.

Labenz calls those starting models the platonic form of an ecommerce company or a pharmaceutical company. She agrees. Every customer is special, and some name tables after people. Palantir already maps organizational data into business ontologies; Illumex's claim is that it does the mapping automatically. When Labenz asks how you map a column named VARX to "product variation," she refuses to trust the name. Even self-explanatory semantics are not trusted at 100%. The system also builds taxonomies of usage: aliases in a dbt pipeline, a channel-attribution formula in a BI report, proximity of contexts. If a nonsense column appears in two calculations that both map to industry metrics, you can deduce what the column is. A blank warehouse with no logs falls back on semantic proximity and pipeline feeds. Usage is the evidence. Labels are clues.

Human review is not optional. BI already inserts a gap between business users and data, she says, by handing subjective modeling to data teams. RAG makes that worse if you trust a data scientist to represent the whole logic corpus. Illumex generates the ontology, then puts certification in a web app or in Slack and Teams so non-technical domain experts can verify definitions. That is the playground: humans and AI checking the same objects. Labenz wonders whether this could even have worked in 2021, when models still looked like stochastic parrots. She still does not fully trust off-the-shelf semantic reasoning. Reasoning is picked up from the canonical ontologies the company keeps enriching, and from rule-like cues already sitting in APIs, such as thresholds in a procure-to-purchase flow.

Labenz has been burned by embeddings at his own company, Waymark. Template matching kept locking onto placeholder text instead of pace and vibe, so they ripped the embedding component out. Structured data, Tokarev Sela replies, is almost bare of context unless you mine metadata and usage. Documents are easier. Illumex's hygiene rule is harsh: leave no leftover embeddings from the input ontology in the output. If a concept in the industry template does not match the customer's sources, it is removed. That usually means the customer never connected a system that would support that workflow. The platform will not introduce a business metric other companies use if this company has no data for it.

There is no one model. Explicit semantic labeling is one model, SQL query analysis another, entity resolution a third, then graph neural nets and older subgraph-matching algorithms on clusters. They benchmark every task and have not found a single open-source model that is good at all of them. On the academic Spider text-to-SQL set, with about sixty thin domains from baseball to cooking, they auto-built ontologies and hit 91%. The miss to 99% was garbage in the open data, including wrong labels. Real enterprises have far more usage context, so she expects the number to improve. Customer-specific gaps are usually connector formats, not missing logic: if the API swagger is new, that is engineering. Ontology creation itself has no manual step. Certification by domain experts is the human one, and she is proud of it, because you will not trust an agent you cannot inspect.

Pre-building that context also cuts the bill. A prompt that hits a pre-built filter never goes to an LLM; only the unmatched remainder becomes a runtime call. She puts the saving at up to 80% of tokens. The same packaged context is meant to be pluggable into any LLM runtime, including AWS Bedrock and NVIDIA NIMs. The industry, she says, is leaving niche agent apps for agentic workflows. "You absolutely have to have the shared context. If each of your models will have separated contexts, they're never aligned." Customers can compose agents from different vendors and keep them on one map.

New foundation models get benchmarked as they land. They are getting better at public-domain reasoning and at user intent: "show me" can now imply a count. They have not, in her testing, made a significant breakthrough on proprietary business corpora, because the labs optimize for the widest denominator. A whole ensemble that used to decompose SQL into filters and measures has collapsed into one semantic model trained on Illumex's own corpus, and that single model beat the ensemble. Each customer still gets an automatically fine-tuned "uber" graph-and-semantics stack, closer to a super RAG than to a generic LLM. Cross-industry projection is the newest trick. A supply-chain employee who asks how many "pieces" they have today is understood as deliveries, because jargon now pollinates across domains. The team rotates rather than splitting canonical-ontology people from mapping people. Data scientists get bored, and keeping every employee relevant is part of the compensation.

People tacitly treat the human analyst as 100% reliable. Labenz has become radically skeptical of one-off requests in his own companies. Illumex benchmarks against analyst-built BI reports in POCs, not against ChatGPT-plus-schema. Analysts are irreplaceable for anything with legal liability, she says: she cannot imagine an SEC filing or a public-company board report shipping from a copilot without a human signature. What the product actually serves is the underserved queue. One department had waited nine months for a BI dashboard that was not a company priority and was expensive to build. A data copilot was something they could have tomorrow.

Labenz's gut for a typical one-off at a big company is 90 to 95% fully accurate. Tokarev Sela puts Illumex above 95% on data people actually use, and a little under 95% on unused data. When the system returns no answer, she says that call is 100% because the data is corrupted: missing values, duplicates, a supposed single source whose value distribution does not match. A naive query on a flawed foundation will look reasonable and still be wrong. Because Illumex learns from prior human interactions, it behaves like an uber-analyst who has sat with every analyst the company ever had. Logic drift is not stuffed into the model on contact. Conflicting definitions get flagged for the data-domain owner. A 20% deviation heuristic could auto-absorb changes; in practice companies want to see the conflict first. The work that remains is review, not construction, and she puts the effort saved at more than 90%, because context and reasoning are already built.

The other reason she calls it a platform for control freaks is interrogation. Human analysts get annoyed after a few rounds of "why that table, why that formula." The machine has no such limit. Ask how channel attribution was calculated and it will say the definition came from the business-metric library, and that the source is used by 90% of API calls. One or two of those rounds is usually enough for a business user to trust a decision. She wants people to see how the cake is baked, and she uses cake rather than sausage on purpose. Governance is built in rather than parked in a GRC department that never touches the agent: conflicts, duplications, and PII are audited, and the reports export. On the way in, every prompt is mapped onto part of the ontology, so a chat box cannot reprogram organizational logic. The only way logic changes is that the underlying metadata changed, a table vanished, an API appeared, and the system flags, alerts, and regenerates definitions.

Never reading the rows was an early, deliberate choice, not a later constraint. Seeing the values would help, she admits, but it is the business model and, she thinks, the future. A SaaS vendor that holds both the logic and the data is a liability she does not want. At runtime Illumex generates the query, sends it into the customer's environment, and the result comes back in Slack. Discovery calls in the early days made it obvious that generative AI touching enterprise data was a non-starter. Asked what it is like as a female founder, she says it is the only experience she has. The metadata stance now knocks whole blocks off vendor-security questionnaires. Enterprises still ask where it is hosted and who the subprocessors are, and she thinks they should. If you do not trust the vendor, you will not trust the output.

If metadata is the scaffolding of the business, Labenz wonders, should companies worry about it more than the rows? Tokarev Sela turns the question. Processes and reported metrics are not that differentiating; public companies already disclose a lot of them. The real split is the richness of data accumulated over time. Foundation models will get faster and cheaper, and the automations around them will be only as personal as the data you feed them. Firms that do not collect everything they can are missing a future revenue engine. Five years ago every company was told to become a digital company. Next come data products, data services, even data agents sold to customers and partners. Application-free software, in her longer view, also needs shared formats so companies can invoke cross-firm workflows, pay for consumption, and keep actual values unshared while the semantics stay aligned.

Who uses Illumex is moving. Six months earlier, most users were technical: data management, governance, analysts running the tool on domains they did not know. In the last six months, inbound demand has been about business-user self-service, which feels less intimidating once governance and cost controls are in the box. She expects agentic workflows themselves to become users faster than people think, AIs chatting with Illumex in Slack. The product, she says, is a playground for machines, humans, data, and applications.

Seat-based pricing breaks in that world. One executive's agent could fire thousands of queries a day. Investors want predictability, and so does she when the bill arrives. The surprise factor on cost and ROI, what the industry is calling the cost of AI success, is the biggest reason enterprise adoption slowed last year. Illumex sells sealed tiers by number of data sources: one source is starter, three to five is mid-size, then an all-you-can-eat enterprise tier, with no seat or consumption surprises. They are not, she says, the *Silicon Valley* pizza-bankruptcy episode. Pre-built context means most interactions stay inside the company, and fine-tunes hit one block of a composite architecture rather than the whole stack. Against off-the-shelf APIs and current graph RAG, she claims the same 80% cost advantage.

Labenz asks the unkind career question. Entry-level software jobs already look shaky. Would he tell a college student to go into data analytics, knowing the degree might finish after the role has changed? Tokarev Sela thinks professions will be reinvented several times in a lifetime, the way cart-pulling became steering a wheel. There will be fewer content creators and more content moderators. That is not a eulogy. Current spend on data analytics, she estimates, covers maybe 10% of what the domain could do. Upgraded analysts should handle alerts, resolve conflicts, talk to business users about the experience they actually want, and give the software instructions. Humans still catch analogies, metaphors, and customer intent, and they still recognize why a model is screaming that the data is insufficient, because they have watched semantic systems iterate. New graduates still need to go deep in something before they can moderate it.

Asked what to study, she does not point first at a syllabus. Inspiration in agents, she says, is coming from the biggest industry players and from unexpected sources such as DeepSeek, not from the academy. Read the keynotes and read between the lines. She still follows academic work on ontology RAG and on marrying context with content, but she does not see great novelty yet, only a growing admission that a flat semantic approach will not carry multimodal agents. A conference paper on ontology-based RAG for supply chain showed that an industry ontology makes RAG better and cheaper. Fine, she says. Now try building that ontology by hand for every use case. It does not scale. Labenz's own advice rhymes: the best general-purpose education in generative AI often comes from fast-moving practitioners who are not professors, because they stay against the frontier.

She has a nine-year-old. She is not sure university is the path, and she wants him to put his hands on whatever is newest, because life is short and the stack moves fast. Neural-net toys for babies did not stick; throwing a ball and talking about gravity did. She introduced him to ChatGPT early, not so he would outsource homework, but so he would learn critical thinking, check sources, and juggle tools. The opposite school, pencil and notebook only, is not her school. After the novelty, the boy decided the homework was trivial and he would just write it himself. She was fine with that. The tools are there when the judgment says to use them. Labenz, with three kids and an oldest almost six, leans her way and is still figuring it out.