An Application-Free Future? Speaking Directly to Data with illumex CEO Inna Tokarev Sela
Summary
- illumex’s core bet is that metadata—not raw enterprise data—can become a shared control layer for AI work. Its system maps schemas, query logs, API signatures, formulas, and usage patterns onto canonical industry ontologies, then creates a customer-specific semantic layer without moving or viewing underlying values. Tokarev Sela says onboarding can take days even when customers have millions of tables.
- The “application-free future” does not eliminate software; it hides its orchestration behind familiar conversational interfaces. Employees could ask questions or initiate workflows through Slack, Teams, voice, or another launcher while specialized applications operate in the background. The aspiration is that users “don’t really care which application you need to invoke,” reducing context switching while allowing niche agents to coexist.
- illumex claims accuracy above 95% overall, while accuracy can fall a little below 95% on underutilized data. In those cases, the product may return no answer because values are missing, duplicated, or corrupted. The commercial wedge is not replacing analysts but reaching employees who might otherwise wait nine months for a BI dashboard.
- Trust is treated as a product surface rather than a promise. Users can trace an answer through definitions, lineage, calculations, source selection, and generated SQL—making illumex, in Tokarev Sela’s phrase, “a platform for control freaks.” Built-in audits cover conflicts, duplication, and PII, while governance also requires attention to access rights, bias, ethics, and skewed data. Prompts cannot rewrite organization-approved business logic.
- Shared context is the prerequisite for multi-vendor agentic workflows, and it also drives the cost thesis. Tokarev Sela argues that “if each of your models has a separate context, they will never be aligned”; illumex therefore provides shared context and connectors for runtimes such as AWS Bedrock or NVIDIA NIMs. Reusing that context can, she says, save up to 80% of tokens.
- The go-to-market signal is a rapid shift from technical users toward business self-service. Six months earlier, users were mainly data-management teams, governance teams, and analysts; Tokarev Sela said all recent inbound requests had been about business-user self-service. Pricing follows connected data sources—one source for starter, three to five for midsize, and an enterprise “all-you-can-eat” tier—with no separate surprises around seats or consumption.
- The labor call is augmentation now, followed by a structural shift from producing analysis to supervising it. Public-company and board reporting still require human validation, but routine generation can move to machines while analysts resolve conflicts, tailor experiences, and interpret alerts. Tokarev Sela’s formulation is “fewer content creators and more content moderators,” implying a shift away from routine production tasks but continued demand for deep domain judgment.
Deep dive
1. illumex began with an application-free thesis before the generative-AI boom
Tokarev Sela founded illumex in 2021 around “automated context and reasoning for agentic AI,” language that proved difficult to pitch at the time. She simplified the message to a mission: enable an “application-free future for knowledge workers” through self-service access to structured enterprise data.
Her starting puzzle came from SAP and Siemens: after heavy investment in data and analytics, why were so many business decisions still based on guesswork? Employees generate more daily questions than analyst teams can answer, leaving factual decision-making constrained by organizational throughput rather than data availability.
Graphs supplied relationships and context; semantic models supplied content. Tokarev Sela saw their combination as a fabric connecting “data to people to workflows,” with natural-language access augmenting employees rather than automating whole jobs away.
2. Applications may proliferate while their interfaces collapse
Nathan Labenz contrasted illumex with AI software builders promising abundant personalized, disposable applications. Tokarev Sela’s answer was that the visions are compatible: specialized agents and applications can proliferate, but business users should not have to learn them, switch contexts, or memorize “the order of clicks” required to complete a task.
Her end state is a launcher interface where someone narrates a question or task and the orchestration disappears into the background: “You don’t really care which application you need to invoke,” what data it needs, or which workflow produces the result.
Labenz pushed the idea beyond desktop chat: voice could let knowledge workers interact with information and execute digital actions while walking rather than remaining “locked in at the workstation.” Tokarev Sela compared the liberation of Ray-Ban glasses with built-in speakers to avoiding the need for additional headphones.
3. The Slack demo makes governed analytics look like an ordinary conversation
In Slack, Tokarev Sela asked, “How many products do you have in stock right now?” The “Let Omni decide” mode matched the prompt against illumex’s semantic ontology, selected the relevant business logic, generated the query, and returned the answer in the same environment.
Analysts can choose a more exploratory mode, while most business users simply consume the result. The same response exposes its semantic entity, lineage, attributes, definitions, source data, and SQL for anyone who wants to inspect “how the cake is baked.”
That layered experience matters commercially: users can receive an immediate answer, while data specialists retain the ability to validate every step. Tokarev Sela called it “a platform for control freaks,” but argued most users will inspect deeply only once or twice before trusting routine results.
4. Canonical ontologies turn fragmented metadata into company-specific context
illumex first built domain knowledge spanning manufacturing, insurance, pharmaceuticals, retail, finance, IoT, and other functions. It encodes terminology, processes, metrics, analysis, and dependencies as business ontologies combining knowledge graphs, semantic embeddings, relational models, and semantic models.
Labenz described these as the “platonic form” of each enterprise type: idealized representations against which a customer’s inconsistent implementation can be compared. Tokarev Sela agreed, noting that organizations customize systems, invent personal column names, and develop structures that diverge substantially from any benchmark.
Customer environments may mix on-premises Oracle, SAP, Microsoft SQL Server, Teradata, or other traditional systems with Redshift, Snowflake, Databricks, Tableau, and Power BI. Departments can have dozens of systems and “absolutely no single source of truth”; illumex rechains its benchmark ontology against their collective metadata.
The inputs are schemas, historical queries, application APIs, pipeline logic, and related metadata—not table contents. Tokarev Sela said onboarding generally takes days, even though some customers have millions of tables, and supports centralized or federated architectures without moving data into a vector database.
5. Usage patterns resolve meanings that names and embeddings cannot
Labenz’s hardest case was an e-commerce field named something meaningless like “VRX,” possibly in another language. Tokarev Sela said illumex “does not trust semantics” even when labels appear self-explanatory; it builds taxonomies that capture how a field is actually used.
A column might feed a dbt transformation, receive an alias, and participate in a BI formula calculating channel attribution. If both meaningful and opaque columns recur in calculations that map to known metrics, the system can infer the opaque field’s role from formulas, proximity, pipelines, and repeated usage.
Labenz said his company had found standard embeddings unsuitable for matching video templates because placeholder text overwhelmed the desired match on “vibe,” pace, and energy. Tokarev Sela acknowledged that structured data is similarly bare of context, making combined semantic, graph, and usage analysis essential.
illumex removes canonical concepts that lack grounding in connected customer systems: “absolutely no traces or leftovers” remain from the input ontology. This prevents an industry-standard workflow from entering the customer model merely because comparable companies use it.
6. Automation lies in many specialized models, not one omniscient LLM
Separate models handle explicit labeling, SQL-query usage, entity resolution, relationship discovery, and subgraph matching; some components use graph neural networks and others traditional algorithms for identifying subclusters. Tokarev Sela said illumex continually benchmarks them because it has not found one open-source model that performs every task well.
On an ontology it generated for the Spider text-to-SQL benchmark, illumex reported 91% accuracy. Its review attributed much of the remaining gap to incorrect or low-quality examples in the open-source benchmark; real deployments provide richer histories, formulas, and application context than Spider’s roughly 60 sparse domains.
Manual work is limited to engineering a connector when a system exposes metadata in an unsupported JSON, schema, or API format. Tokarev Sela was categorical that ontology creation itself has “absolutely no manual task,” while business experts still review and certify the result.
She compared the approach with Palantir’s use of business ontologies and organizational-data mapping, arguing the underlying concept is not novel. illumex’s claimed differentiation is automating that mapping and cleaning rather than staffing a manual ontology-building process.
7. General models understand questions better than they understand businesses
Model progress has simplified bounded tasks: illumex once used an ensemble to decompose and classify SQL into filters, dimensions, and measures, whereas one domain-trained semantic model now outperforms that stack. Tokarev Sela nevertheless sees no significant breakthrough in out-of-the-box understanding of proprietary business processes.
Foundation models increasingly infer user intent—casual language such as “just show me” can imply a count—but struggle with the organizational meaning of the subject. Providers train toward broad public-domain understanding, not the private terminology, rules, thresholds, and exceptions that make a particular enterprise function.
Each customer therefore receives its own custom semantic language model and graph system, automatically tuned on organizational metadata and possible usage examples—“an ‘Uber RAG,’ in a sense.” Canonical ontologies continue improving in parallel rather than freezing at onboarding.
Cross-industry projection helps interpret employees who do not speak a single domain’s jargon. In her supply-chain example, “How many pieces do we have today?” might use “pieces” to mean deliveries; the system maps that language without assuming every speaker uses the canonical term.
8. Shared context aligns heterogeneous agents and cuts repeated inference
Agentic orchestration fails, in Tokarev Sela’s framing, when every model receives a separate representation of the business: “If each of your models has a separate context, they will never be aligned.” A governed shared context lets models from different providers reason over the same accepted concepts and relationships.
illumex provides connectors to runtimes such as AWS Bedrock or NVIDIA NIMs. Customers can combine niche agents without rebuilding, retransmitting, or recreating the entire organizational context for every runtime call.
Because prompts first match against prebuilt context and only the result of that matching is sent as a runtime call, Tokarev Sela claimed savings of up to 80% in tokens versus repeatedly supplying context through conventional approaches.
9. Accuracy without governance is not an enterprise product
Asked to compare ChatGPT, human analysts, and illumex, Tokarev Sela rejected the tacit assumption that humans are 100% reliable. illumex benchmarks against analysts’ production queries and BI reports, then exposes enough evidence for users to interrogate why a source, definition, or calculation was chosen.
She said accuracy is above 95% overall, while it can fall a little below 95% on underutilized data. Missing values, conflicting distributions, or duplicated sources can make “there is no answer due to corrupted data” the correct response; she claimed 100% conviction in the corrupted-data cases the company analyzed.
Analysts remain “irreplaceable” for public-company, board, or legally consequential reporting that requires human approval. The near-term opportunity is the neglected queue: one prospective department had waited nine months for a BI dashboard and wanted a copilot because the company did not consider its request a priority.
Governance sits inside the reasoning path rather than in a separate GRC exercise. illumex audits conflicts, duplicates, PII, and other issues; users can export reports, certify definitions, and inspect lineage. Tokarev Sela also identified bias, ethics, access rights, and skewed data as governance considerations.
10. Learning from history does not mean accepting every new query as truth
Tokarev Sela described the product as an “Uber analyst” carrying the combined experience of an organization’s analysts. It does not load that history into one context window; it decomposes knowledge into semantic entities, relationships, workflow embeddings, and reusable building blocks.
Top-down workflow representations are cross-validated against bottom-up component embeddings so the logic does not silently drift. A new analyst’s ad hoc query with faulty logic should not immediately retrain the organization’s accepted model.
Serious deviations are flagged as conflicts for a data-zone owner or domain expert. illumex could use thresholds—Tokarev Sela offered deviations above 20% as an example—to begin adjusting automatically, but she said customers generally prefer to inspect a logic change before incorporating it.
Human participation is review rather than construction: the system generates definitions and explanations, while experts approve exceptions. Tokarev Sela estimated this removes more than 90% of the effort or friction associated with building comparable agentic workflows manually.
11. Metadata-only execution reshapes security, pricing, and data strategy
At runtime, illumex converts a prompt into a query or tool call and sends it into the customer’s infrastructure for execution; the result returns to the originating interface. illumex works from metadata and business logic without directly handling the underlying values or sensitive data.
Tokarev Sela said keeping business logic separate from data is imperative. A SaaS provider holding both a company’s logic and all its data becomes a concentrated liability; the metadata-only architecture reduces risk, although enterprises still need answers about hosting, subprocessors, and account isolation.
Labenz wondered whether metadata itself might reveal the scaffolding of a competitor. Tokarev Sela pivoted: processes and reported metrics are often less differentiated than the accumulated data, and future automations will be “as personalized, as good as your data is.” Companies that fail to accumulate rich data may miss future revenue engines.
Product demand had shifted in six months from analysts, governance teams, and data-management teams toward business-user self-service; Tokarev Sela expects AI agents to join them. illumex prices by connected sources—one for starter, three to five for midsize, then enterprise “all-you-can-eat”—with no separate surprises around seats or consumption.
12. Analysts become moderators as frontier practice outruns formal curricula
Labenz asked whether AI is pulling up the entry-level ladder as it may be doing in software development. Tokarev Sela expects professions to be reinvented repeatedly, but framed the data transition as moving from “content creators” to “content moderators,” not eliminating expertise.
Organizations spend heavily on analytics while completing perhaps 10% of the work that could be done. Analysts can instead resolve semantic conflicts, investigate insufficient-data alerts, customize agent behavior, translate business requirements, and supervise outputs at a scale manual report production cannot reach.
Her advice to graduates was to build deep enough proficiency to moderate systems soon, because analogy, metaphor, user intention, and failure diagnosis still require judgment. She was less confident about the role’s 50-year durability: moderation itself may eventually become another profession that gets reinvented.
For current learning, Tokarev Sela recommends following major industry players and their business keynotes, while also reading research on ontology and multimodal context. She mentioned recent NIST-conference research on ontology-based RAG for supply chains, which reportedly improved performance and cost, but questioned whether manually creating an ontology for every use case can scale.
Her parenting analogue was direct access plus skepticism: let her nine-year-old use ChatGPT, but prioritize critical thinking, source checking, and asking “what could go wrong.”