Ask an AI agent why sales fell this week, and the first challenge is deciding what to measure.
Which stores should the analysis include? Does “sales” mean gross sales or sales net of refunds? Are we comparing equivalent periods—for example, week-to-date versus the same period last week?
These definitions shape the tables an agent selects, the filters it applies, and the calculations it performs. They also determine whether its results agree with the reports the business already uses.
MIP—MirrorShip Intelligence Platform—connects enterprise data, business semantics, and supporting context to support this process. It provides a shared foundation for MirrorPilot, MIP’s built-in Data Agent, and for agents enterprises already use.
Business definitions belong in queries
For this sales investigation, the agent needs more than order records.
Orders, inventory, traffic, and campaign records provide the data needed for the analysis. Before querying them, the agent needs the business definitions that make those records meaningful.
Gross order value and net sales after refunds can both be valid measures. They answer different questions. An agent needs to establish which measure applies before selecting fields and calculating a result.
Combining sales and inventory introduces further requirements. Product identifiers need to match across systems. Inventory snapshots need to be aligned with the relevant order periods. Store inventory must be distinguished from warehouse stock.
These rules need to become inputs to the query itself. When metric formulas, relationships, and aggregation levels are explicit, teams have a basis for checking both the calculation and the result.
Context connects facts and methods
For enterprise analytics, context brings together the data, business definitions, knowledge, memory, and skills an agent needs for a task.
MIP organizes this context around five types of information:
These categories describe how information is used. A refund-policy document is knowledge; the rule that affects net sales also needs to be expressed in an executable metric definition. A validated investigation can become a memory record, and its repeatable steps can help refine a skill.
Retrieval and semantics therefore serve different purposes. Retrieval can help an agent find the applicable refund policy. A semantic definition establishes how the corresponding sales metric is calculated.
The task also determines which context is appropriate. A past investigation may apply only to a particular channel or period. A definition may have changed. Access permissions and validation status still need to be considered.
A foundation agents can share
Much of this context is useful across tasks and agents. A sales agent and a supply-chain agent may both use the same product, store, and regional definitions.
If each maintains those definitions separately, a change in regional coverage has to be replicated across applications. If the updates take effect at different times, the agents may produce conflicting regional totals even when each query runs correctly.
MIP places reusable data access and business semantics at the platform layer. Different agents can use that shared foundation while retaining their own objectives, tools, and workflows.
Shared semantics does not mean forcing every team to use one formula. Finance and operations may legitimately use different sales measures. The purpose is to make each definition explicit, including when and where it applies.
MIP brings this approach together in three parts:
One Engine. Unified Semantics. Every Agent.
One Engine: MirrorData
MirrorData is MIP’s unified data foundation, combining object storage, open lakehouse capabilities, and a real-time analytics engine.
It supports analysis across an enterprise’s existing data infrastructure, including real-time analytical stores, lakehouse tables in open formats such as Apache Iceberg, and connected business systems.
For the sales investigation, recent orders and inventory updates may come from a real-time analytics system, while historical records reside in an open lakehouse. Campaign documents and operational logs can supply additional background. The actual scope depends on source integration, update pipelines, and permissions.
Depending on the data and task, analysis can combine SQL queries, full-text search, and vector search. These capabilities can work together—for example, by joining retrieval results with structured business data.
Images, audio, and video can contribute through metadata and processed representations such as labels, text, or vectors. Their use depends on the preprocessing and indexing configured for the task.
Unified Semantics: MirrorMind
MirrorMind organizes business semantics, knowledge, memory, and skills to provide relevant context for analysis.
Semantic views encode business definitions in the data engine, including metrics, dimensions, and relationships. Agents query the defined metrics and dimensions, while the engine expands and executes those queries using the declared relationships and calculation rules. Analysts can inspect the expanded SQL with EXPLAIN INLINE.
In the sales example, the team first establishes:
The formula for net sales.
The relationship between stores and regions.
The applicable time range and order-status filters.
The comparison periods, ensuring they cover equivalent durations.
The agent references these definitions in its query, and the engine applies the declared relationships and calculation rules during execution. The resulting calculation can then be checked against reference queries and business reports.
Supporting material has a separate role. Campaign plans may help explain a change in demand. Historical corrections may clarify why a definition changed. Calculation rules establish the numbers; supporting records help people assess the explanation.
Every Agent: MirrorPilot and beyond
MirrorPilot is MIP’s built-in Data Agent for asking questions and analyzing enterprise data.
It uses the applicable metrics, dimensions, and relationships to query data, then draws on relevant context to explain the results.
Custom agents, industry-specific agents, and copilots can also connect to MIP through interfaces, subject to integration and authorization. Enterprises can keep their existing user interfaces and workflows while maintaining shared definitions at the platform layer.
A sales agent may investigate revenue changes while a supply-chain agent examines replenishment. Both can reference the same product and store definitions without needing identical analytical workflows.
Deliver answers with supporting evidence
Return to the question: Why did sales decline this week?
Once the relevant data, definitions, and tools are configured, the analysis can establish which stores to include, which sales metric to use, and which periods are comparable. It can then break down the change by channel, store, and product before checking inventory, traffic, campaigns, and previous records.
The analysis should distinguish findings supported by the data from hypotheses that require further investigation.
A decline concentrated in a few products is a finding supported by the data. Stockouts during the same period are a relevant observation, but estimating their contribution to the decline requires further analysis, including checks on product availability, traffic, and other contributing factors.
A useful result should make four things clear:
Data sources: Which records were used, how current they were, and which stores they covered.
Business definitions: Which metric, regional scope, and comparison basis applied.
Calculations: How data was joined, filtered, aggregated, and compared.
Limits of the explanation: Which statements are observed facts, which are inferences, and which need further validation.
This is what “Beyond Answers. Toward Understanding.” means in practice: an answer with enough supporting information for people to examine it, challenge it, and continue the investigation.
Keep what the enterprise learns
Models, agent frameworks, and applications will change. An enterprise still needs to maintain its definitions of customers, products, revenue, and risk.
Each validated analysis can leave behind something reusable: a corrected metric definition, a reference query, a time-bounded record of an anomaly, or a tested investigation procedure.
For example, a team may confirm that internal orders should be excluded from a particular metric. It can update the definition and check the change against a reference query. Future agents can reuse the corrected definition rather than rediscovering the rule.
That accumulation requires maintenance. Definitions need effective dates. Historical findings need a clear scope. Outdated material needs to be reviewed or retired.
MIP connects these enterprise assets to the agents that use them. MirrorData provides the data foundation, MirrorMind organizes definitions and supporting context, and MirrorPilot and external interfaces bring them into analytical tasks.
The aim is to help enterprises carry their existing data investments and business knowledge forward as AI evolves.
See MIP in action
On October 20, MirrorShip will introduce MIP’s architecture and core capabilities, with product demonstrations showing how enterprise data, business semantics, and supporting context work together in agent-led analysis.
The Next Begins Here
MirrorShip Intelligence Platform AI Product Launch
Date: October 20, 2026
Time: 2:00 p.m. Beijing / Hong Kong time (UTC+8)
Format: Online livestream
Explore the launch event: https://www.linkedin.com/events/7515230959584305152?viewAsMember=true