33 min read

An Assistant Can Enter Every App and Still Not Find Your Entities

Opening every app to an AI assistant solves calling, not identity. Once assistants surface any entity in a semantic index, entity resolution becomes the enterprise constraint — and the data team's bill.

An Assistant Can Enter Every App and Still Not Find Your Entities

A sales rep asks the assistant on her phone to pull up "the Meridian contract." The assistant has permission to reach the CRM, the contract repository, the billing system, and the support desk. It opens all four and returns three Meridian entities: a prospect marketing created in 2024, a customer record renamed after an acquisition, and a billing account keyed to a legacy ERP number. She wanted the second one. The assistant did exactly what it was built to do — it called every app — and still could not answer the question, because nothing in the chain had decided which Meridian is the Meridian.

That failure mode is arriving at scale. On November 3, 2026, Apple released a Siri AI developer preview across the iOS, iPadOS, macOS, and visionOS 27.3 betas, letting any app integrate through App Intents, adding skills that combine several actions into one request, and stating that Siri AI can now surface any entity in the Spotlight semantic index — not just entities that fit an existing schema. App Store submissions open in the spring. The same week, Google said K-12 accounts will start with Search, Gemini, and Gemini Notebook all off early next year, and its education vice president told the New York Times that "Search and AI are deeply integrated," as reported November 4, 2026.

Read the two together and the direction is clear: the assistant layer is opening to every app, and the app layer is fusing with the assistant. Neither announcement says whether your organization can tell the assistant which entity it is looking at. That question is answered on your side of the API. For a CDO or data platform lead, it is now the constraint that decides whether an assistant deployment produces answers or plausible noise.

Opening every app solves calling, not identity

App Intents is an integration mechanism. It lets an app declare its content and actions so the assistant can invoke them, which removes the need for a bespoke integration every time a user wants to check a delivery or book a reservation. That is a real unlock. But integration is plumbing, and plumbing is the part platform vendors are built to solve. The harder problem sits one layer down: when the assistant calls your app and asks for "the Meridian contract," your app has to return one thing.

Apple's schema-free claim sharpens the point. Previously an assistant could surface only entities that fit a predefined schema — a contact, a calendar event, a message. The preview states that Siri AI can now surface any entity in the Spotlight semantic index. A schema is a contract about what an entity is, and it did quiet disambiguation work on the assistant's behalf. Remove the schema requirement and the assistant will surface whatever your index contains, including the three Meridian records, the duplicate vendor rows, and the customer whose identifier changed when the company was acquired.

The causal chain runs in one direction. The assistant's reach expands to every app and every entity type. The number of candidate matches for any given request therefore grows, not shrinks. The only thing that can collapse those candidates to one is an identity layer that knows which records refer to the same real-world thing. Without that layer, greater reach produces more confident wrong answers, not fewer. Opening every app to an assistant solves calling. It does nothing for identity.

The binding constraint has moved to entity resolution

For a decade the constraint on enterprise AI was access: could the model reach the data? That constraint is dissolving. The assistant can reach the app, and the app can reach the database. The remaining question is whether one database agrees with another about who the customer is.

Entity resolution is the discipline of deciding when two records refer to the same real-world entity. It is unglamorous, expensive, and almost always under-resourced. It depends on keys, aliases, merge history, and the rules that govern when a merge is safe. A customer in the CRM may carry a Salesforce ID; the same customer in billing may carry an ERP account number; the support desk may know them only by email domain. None of those systems is wrong. They never agreed on a shared identity, because until now nobody asked them to answer a question in real time on behalf of an assistant.

Schema-free surfacing turns that background chore into the binding constraint. When the assistant could surface only schema-fit entities, the schema absorbed part of the disambiguation. Now the assistant will surface any entity the index contains, which means the index's quality is the assistant's accuracy. A semantic index built over unresolved entities is a fast, confident way to retrieve the wrong record.

Model improvement does not dissolve this. A better model can be more persuasive about a match, but it cannot know that the ERP account and the CRM record are the same customer unless something in your data says so. The model reasons over what you expose. If what you expose is three Meridian records with no merge history, fluency becomes a liability rather than an asset.

Availability conditions are a data-governance signal, not a footnote

Apple's release notes are unusually candid about the conditions attached to Siri AI, and each one matters to a data leader. Siri AI is unavailable in the EU on iOS, iPadOS, and watchOS, though it is available on macOS 27 and visionOS 27 for EU account holders. It is not available to users under 13. Server-side features carry daily usage limits, and expanded access will be available for a fee in the future.

Treat that as a governance brief rather than fine print. A capability that is regionally unavailable, age-gated, rate-limited, and eventually metered is not a stable utility you can wire into a customer-facing workflow without a fallback. If your entity resolution depends on an assistant that is unavailable on European iPhones, your European operations need a different path to the same answer. If server-side features are subject to daily limits, a high-volume internal tool cannot assume unlimited calls. And if expanded access will eventually carry a fee, the cost of assistant-mediated lookups becomes a line item you should be able to forecast.

The Google decision makes the same point from the other direction. K-12 accounts will start with Search, Gemini, and Gemini Notebook all off unless administrators explicitly turn them on, and because Search and AI are now deeply integrated, schools that decline the AI also lose Search. That is a platform making an architectural choice about how tightly capability and dependency are coupled. The lesson for an enterprise data leader is that assistant capability is a conditional dependency whose terms the platform sets, not you. The one part of the stack whose terms you do control is the entity layer, which is an argument for owning it.

Who pays the cost, and the case for waiting

The cost of this shift does not land on Apple or Google. It lands on the data platform team that now has to make its entities externally consumable, and on the CDO who has to fund the work. The vendors have solved their part of the problem — reach — and the remaining work is yours. That is not a criticism of the vendors; it is where the boundary of responsibility falls.

The strongest opposing view is that this work is premature. Models keep improving at fuzzy matching, the argument goes, so why build a governed entity index now when the next release may resolve Meridian on its own? It is a fair challenge and deserves a direct answer rather than a dismissal.

Fuzzy matching and entity resolution are not the same operation. A model can score the similarity of two records. It cannot know, absent a governed source of truth, that a merge was deliberate, that an alias is legitimate, or that two records were kept separate on purpose because they represent different legal entities with the same name. Model improvement raises the ceiling on candidate generation. It does not supply the merge history, the alias table, or the permission boundary that decides which record a given user is allowed to see. Those are enterprise artifacts, and no model release produces them. Waiting does not avoid the work; it defers it while the assistant layer's reach keeps expanding, which raises the cost of the eventual cleanup.

Build the entity index before you integrate the assistant

Invert the usual sequence. Most teams integrate the assistant first and discover the identity problem in production. Build the externally consumable entity index first, then integrate, so the assistant has something coherent to call.

Start with one high-frequency entity — customer, order, or contract — and export three things for it: the keys each source system uses, the aliases the business has accumulated, and the merge history that records which records were combined and why. That export is the raw material of resolution, and most teams discover when they attempt it that the merge history lives in someone's memory or a spreadsheet rather than a governed table.

Then measure the share of entities that cannot be auto-matched. That number tells you whether the index is ready to expose. A high unmatched share is not a failure; it is a scoping signal that shows where manual review or rule refinement is needed before the assistant is allowed to rely on the index. Track it over time and use it to decide which entity types are safe to expose first.

The architecture question is where the index lives. If entity tables, their detail, and their permissions sit in one engine, the entity index can be a view over the existing tables rather than a new ETL copy that drifts from the source. That is the capability Singdata describes for its lakehouse: entity tables, detail, and permissions in a single engine, so the index is a governed view rather than a duplicated pipeline. The value is not a benchmark — it is that the index stays consistent with the systems of record because it is derived from them, and permissions are enforced in the same place the data lives.

FAQ

Why does giving an assistant access to every app not solve finding the right customer?

Because access and identity are different problems. App Intents lets the assistant call your app; it does not tell your app which of three Meridian records is the right one. The assistant's reach expands the number of candidate matches, and only an entity layer that knows which records refer to the same real-world thing can collapse them to one.

Isn't building an entity index premature if models keep improving at fuzzy matching?

Model improvement raises the ceiling on candidate generation, but it does not supply merge history, alias tables, or permission boundaries. Those are enterprise artifacts. Waiting defers the work while the assistant layer's reach keeps expanding, which raises the eventual cost rather than avoiding it.

What is the smallest first step before integrating an assistant?

Pick one high-frequency entity and export its keys, aliases, and merge history into a governed table. Then measure the share of records that cannot be auto-matched. That number tells you whether the entity is ready to expose to an assistant, and it gives you a baseline to improve against before you widen the integration.

Previous posts

1. Medallion Architecture: What Changes Between Bronze, Silver, and Gold? — https://singdata.com/trending/medallion-bronze-silver-gold

2. How Much Data Preparation Is Hidden in Your BI Reports? — https://singdata.com/trending/hidden-data-prep-in-bi-reports

3. How to Migrate from ETL to ELT Without Breaking Reporting or Agentic AI Freshness — https://singdata.com/trending/2026-11-04-elt-file-first-ingestion