Air-gapped
A deployment with no network connection to the outside world. Genuinely air-gapped software must operate without phoning home, which requires offline licensing and an offline update process. It is testable: disconnect the network and see whether the system still answers.
Citation
A link from an answer back to the exact source it came from. In regulated workflows citations are not a nicety: they are the audit trail that makes an AI-assisted decision defensible.
Co-sell
A Microsoft programme in which a partner's solution becomes visible to Microsoft's own sellers, who are compensated for supporting the deal. Requires enrolment and a qualifying marketplace offer.
Data residency
The requirement that data remains within a defined jurisdiction or environment. Residency is the most common blocker to enterprise AI adoption in regulated sectors, because most cloud AI architectures move or copy the data to be processed.
Data sovereignty
The principle that data is subject to the laws of the country in which it is held. Broader than residency: it covers who can compel access, not only where bytes sit.
DORA (Digital Operational Resilience Act)
EU regulation on operational resilience for financial entities, including requirements to monitor and document ICT third-party providers. It has made vendor security evidence a procurement gate rather than a formality.
Egress
In infrastructure, data leaving a controlled environment. It appears as a line item on cloud bills, a clause in security reviews and a risk in AI rollouts. Reducing egress to zero is the design goal Zegress is named after.
Entity extraction
Identifying the people, organisations, locations and topics inside documents, and the relationships between them. Useful for due diligence, investigations and mapping an unfamiliar estate before a migration or merger.
EU AI Act
EU regulation on artificial intelligence. High-risk obligations under Annex III apply from 2 December 2027 and Annex I from 2 August 2028, while transparency, governance and penalty provisions apply from 2 August 2026. Record-keeping and technical documentation duties were not relaxed by the deferral.
Grounding
Constraining a model's answer to specified source material. A grounded answer can be checked; an ungrounded one is a guess presented with the same confidence.
Hybrid retrieval
Combining lexical (keyword) and semantic (meaning-based) search. Keyword search matches strings and misses paraphrase; semantic search matches meaning and can miss exact identifiers. Together they outperform either alone.
MACC (Microsoft Azure Consumption Commitment)
A customer's contractual commitment to spend a set amount on Azure over a term. Purchases through the Azure Marketplace draw down against it, which turns an AI purchase into spend of an existing, expiring budget rather than a new budget request.
MCP (Model Context Protocol)
An open standard for connecting AI assistants to external tools and data sources. Because it is open, a retrieval layer that speaks MCP can serve multiple assistants without lock-in to any one model vendor.
Offline licensing
Licensing that requires no call home to validate. Necessary for air-gapped operation and verifiable by the customer at the network layer.
Permission trimming
Filtering retrieval results against the asking user's entitlements before anything reaches the model. Trimming at query time, rather than at index time, is what keeps the access model single-sourced.
Permission-aware retrieval
Enforcing existing access controls at the moment a question is asked, so a user never sees content from a document they could not open directly. The alternative — re-permissioning into a new system — creates a second access model that drifts out of sync.
Private offer
A negotiated marketplace offer restricted to a specific customer tenant and subscription, with agreed price, term and support. How enterprise deals above list capacity are transacted.
RAG (retrieval-augmented generation)
An approach where a model answers using passages retrieved from a document set at query time, rather than only from what it learned in training. Retrieval quality, permission handling and citation are what separate a usable enterprise RAG system from a demo.
Tenant isolation
Ensuring one customer, team or client cannot access another's data, enforced at the data layer rather than in the application. Essential for professional services firms handling multiple clients.
Version history
Retaining superseded versions of a document's extracted text so that a system can answer what a document says now, what it said before, and precisely what changed.
Zero egress
An architecture in which data does not leave the organisation's environment at any point in a workflow. In AI, it means the document estate, the search index and permission enforcement all remain inside the customer perimeter, and any model that is used reaches in for a specific passage rather than receiving a copy of the corpus.
Zero-copy
Connecting to data where it already lives, without migrating or duplicating it. Avoids both the migration project and the governance burden of a second copy of sensitive material.
Something missing, or a definition you would sharpen? Tell us at [email protected].