- Services
SERVICES
- Platform
PLATFORM
- Solutions
AI AGENTS
LIFE SCIENCES AI
TECHNOLOGY
- Industries
INDUSTRIES
- Work
- About Us
CORPORATE
RESOURCES
- Contact us
Generative AI is a class of artificial intelligence that produces new content rather than classifying or predicting. Given a prompt, it generates text, code, images, audio, or structured data that did not exist before, following patterns learned from training data.
The distinction from earlier AI is what the output is. Predictive AI estimates a value or category from data: this customer will churn, this transaction is fraudulent, this machine will fail. Generative AI produces an artefact: a summary, a draft, a function, a specification.
Generative models are the reasoning component inside most modern enterprise AI systems, including agents. An AI agent uses a generative model to interpret a goal and decide what to do next, but the agent is not itself a generative system. Conflating the two is the most common category error in enterprise AI discussion, and it matters because the risk profiles differ substantially: generative output can be reviewed before anything happens, while agentic action has already occurred.
Four, and they explain most of the architecture that surrounds these models in production.
No inherent knowledge of your organisation. A model knows what was in its training data. Your policies, product details, customer history, and processes are not in it, which is why retrieval over enterprise content is standard rather than optional.
No distinction between recall and construction. The model generates probable continuations, and probability correlates with truth in general but not reliably in any specific case. This is the mechanism behind hallucination.
A fixed knowledge cutoff. Training data ends at a point in time, and anything after it is unknown unless supplied.
No ability to act. A generative model produces text. Reading a record, updating a system, or sending a message requires tools and integration built around it.
Each limit has a corresponding architectural answer: retrieval for knowledge, grounding and citation for accuracy, current data supply for recency, and tool use for action. Enterprise AI platforms exist largely to provide those four.
Wizr AI treats the generative model as the reasoning layer inside a governed system rather than as the product. The Wizr Enterprise AI Platform supplies the surrounding components: retrieval over enterprise data, the integration layer that lets agents act in business systems, and the security and governance controls that make deployment approvable.
This layering is what allows the model itself to be selected per engagement. Wizr AI describes a model agnostic approach spanning proprietary and open source models, chosen on performance, cost, and data sovereignty grounds.
Generative capability appears across the portfolio in different forms: drafting responses in customer support, producing specifications and code in Glidepath AI SDLC, and authoring regulatory narrative in Wizr eCTD Studio.
Grounding is the practice of tying an AI system’s output to verified source material, so that what it produces reflects the organisation’s actual documented position rather than the model’s general knowledge. A grounded answer can name the document it came from.
Grounding is the outcome; retrieval is the mechanism that achieves it. The distinction matters because a system can retrieve without being grounded, if the model is free to supplement retrieved content with its own knowledge, and the resulting answer blends verified and unverified material indistinguishably.
Effective grounding therefore requires three things together: retrieval that supplies the right material, instructions constraining the model to answer from it, and citation that makes the connection visible and checkable.
Deploying grounded AI reliably exposes a knowledge estate problem that predates the AI programme. Five properties determine whether content can support it.
Coverage of real questions. Documentation is usually written around what the organisation wanted to explain rather than what people actually ask. Gaps appear as model failures and are content failures.
Single authoritative version. Where the same policy exists in four documents with three variations, retrieval returns whichever is most similar rather than whichever is correct.
Retired obsolete content. Superseded material retrieves as confidently as current material unless it is removed or clearly marked.
Structure that supports chunking. Content organised in sections with clear headings chunks into meaningful units. Long undifferentiated prose does not.
Permission metadata at ingestion. Entitlements must be captured when content is indexed, because retrieval that ignores them is a data leak with a convenient interface.
The useful reframing is that grounded AI does not require a perfect knowledge base, but it does make the state of the existing one visible and measurable for the first time, usually through retrieval failure patterns that identify precisely which gaps matter.
Wizr AI’s customer support agents are grounded in the organisation’s support centre content and past customer interactions, which is the combination that allows them to handle cases documentation alone would not cover, since prior resolutions capture knowledge that was never written down as policy.
In engineering, Glidepath AI SDLC grounds generation in a version controlled single source of truth for coding standards, architectures, and reusable artefacts, together with reference BRDs, HLDs, LLDs, tested code, and integrations. The version control is the relevant property, since grounding in an uncontrolled standard propagates whatever it contains.
In regulated work, Wizr eCTD Studio drafts from clinical study reports, preclinical summaries, and CMC data and verifies every table, figure, and citation across all five submission modules, which is grounding with verification attached rather than grounding alone.
GxP is the collective term for the good practice regulations governing pharmaceutical development, manufacturing, and distribution, including good manufacturing, clinical, laboratory, and distribution practice. GxP compliance for AI systems concerns what is required when software using AI participates in a regulated process.
Three requirements dominate. Computer system validation establishes documented evidence that a system performs as intended and will continue to. Electronic records and signatures rules, notably 21 CFR Part 11 in the United States and Annex 11 in the European Union, govern record integrity, audit trails, and signature controls. Data integrity expectations, frequently summarised as ALCOA principles, require records to be attributable, legible, contemporaneous, original, and accurate.
AI systems complicate validation in a specific way. Traditional validation assumes deterministic behaviour, so a defined input produces a verifiable output every time. A language model based system may reach the same correct outcome by different routes, which means validation shifts from confirming a fixed path to evidencing consistent outcomes across a representative case set, with controls around what the system may do unsupervised.
Six properties, and the fourth is where most general purpose AI platforms fall short.
Defined and documented intended use, since validation is always against a stated purpose rather than against general capability.
Reproducible configuration, meaning versioned prompts, model versions, tool definitions, and knowledge base state, so a result can be interpreted in the context that produced it.
Complete audit trails meeting electronic record requirements, including who or what performed each action and when.
Change control that covers model updates. A model version change is a system change requiring assessment. Platforms that update models transparently without customer control create a validation problem that is difficult to resolve after the fact.
Human control points at regulated decisions, with the decision recorded against a named qualified person.
Ongoing performance monitoring, since validation is a continuing state rather than a one time exercise.
Wizr AI’s regulated solutions are built with traceability, audit readiness, and human review at defined control points, and Wizr eCTD Studio operates with complete traceability and regulatory oversight while keeping humans in control of regulatory decisions. At platform level, the Security component provides least privilege access audited to SOC 2 Type II, encryption in transit and at rest to ISO 27001 standards, and continuous oversight through real time monitoring and scheduled independent audits.
It is worth being precise about scope. SOC 2 Type II and ISO 27001 are information security assurances. They are not GxP validation, and they do not evidence 21 CFR Part 11 or Annex 11 conformance. Wizr AI does not currently publish a computer system validation position, validation documentation package, or Part 11 statement.
For any regulated deployment, validation approach, change control over model versions, and electronic records conformance should be established directly with Wizr AI as part of the engagement rather than inferred from the security certifications, which address a different set of requirements.
