Key takeaway
- Generative artificial intelligence in 2027: how it works, where it fits, NIST and EU AI Act obligations, and how to govern agentic use at enterprise scale.
What is generative artificial intelligence, and what should enterprise leaders do with it in 2026?
Generative artificial intelligence is a class of models that learn patterns from large datasets and use them to produce new text, code, audio, images, video and 3D content. Predictive AI classifies or forecasts from existing data; generative AI creates new output. In 2026 the useful question for a CTO or CIO is no longer whether to use it, but where it may act on its own, under which policy, and with what evidence.
The adoption picture supports that shift. In its survey published on 5 November 2025, McKinsey reported that 88% of respondents said their organisations use AI in at least one function, up from 78% a year earlier, and that 62% were at least experimenting with AI agents. Only 23% said they were scaling an agentic AI system somewhere in the enterprise. The gap between pilots and governed production is where most leaders now need help. For the model layer, see our guides to generative AI models and foundation models.
How does generative AI work?
Generative models use neural networks to learn the structure of their training data, then sample new outputs from what they learned. Foundation models are trained on broad data, often without labels, and then adapted to many tasks through prompting, retrieval, fine-tuning or tool use. Two research milestones explain most of what enterprises use today.
| Milestone | What it introduced | Source |
|---|---|---|
| Generative adversarial networks, 10 June 2014 | A generator that tries to make a discriminator err, so it learns to produce samples the discriminator accepts as real | arXiv 1406.2661 |
| Transformer, 12 June 2017 | A sequence model built on attention alone, without recurrence or convolutions, now the basis of most large language models | arXiv 1706.03762 |
For the adversarial approach in detail, read our GAN architecture guide.
Which generative AI modalities matter for the enterprise?
Six modalities cover most enterprise use. The table shows what each does, where it fits and the main control to put around it.
| Modality | Typical enterprise use | Control to put around it |
|---|---|---|
| Text | Summarising documents, drafting customer-facing material, explaining complex topics | Grounding in approved sources; human review of external content |
| Code | Creating code from natural language, documenting and maintaining code | Code review, testing and supply-chain checks before merge |
| Audio | Call-centre assistance and field troubleshooting support | Disclosure to callers; retention rules for recordings |
| Image | Product visualisation and damage-scene reconstruction for claims assessment | Provenance labelling; human sign-off on claims decisions |
| Video | Marketing videos and safety-training scenarios | Disclosure of synthetic media; rights clearance |
| 3D and specialised | Virtual environments, AI-assisted prototyping and design | Design review against engineering constraints |
Where does generative AI create value, and where is it a poor fit?
Value usually appears in six places: cost reduction, process efficiency, growth, innovation, discovery and insight, and citizen services in government. Use the table to decide where to start, and treat the last column as a gate, not a footnote.
| Value area | Where it fits | When not to use it |
|---|---|---|
| Cost and process efficiency | Repetitive drafting, triage, summarising and documentation | Decisions with legal effect and no human review |
| Growth | Personalised marketing and customer service content | Claims about products or prices that nobody checks |
| Innovation and discovery | Faster prototyping, idea generation, literature review | Outputs treated as evidence without verification |
| Government and citizen services | Drafting responses and navigating services | Eligibility or benefit decisions without accountable review |
| Healthcare and life sciences | Support for imaging review and research drafting under clinical oversight | Autonomous treatment decisions |
Healthcare, drug discovery, translation, education, supply chain and urban planning are often cited use cases. Treat each as a candidate for a governed pilot with an accountable owner, not as a settled outcome.
How has generative AI changed from answering to acting?
The main change since the first chatbots is autonomy. Models now call tools, read enterprise data and take multi-step actions, which is the agentic pattern. That moves the risk from a wrong sentence to a wrong action in a business system. Our guide to agentic workflows covers how those loops are built and bounded.
| Pattern | What it does | Main enterprise question |
|---|---|---|
| Generative assistant | Produces content for a person to review | Is the output accurate and safe to use? |
| Retrieval-grounded assistant | Answers from approved enterprise sources | Is the source current, permitted and traceable? |
| Agentic workflow | Plans and takes actions across systems | Which actions are allowed, and who approves them? |
| Governed agent estate | Many agents under shared policy and audit | Can we show why each action was allowed? |
What does governance of generative AI need to cover?
The NIST Generative AI Profile (AI 600-1), published on 26 July 2024 as a companion to the AI Risk Management Framework, names 12 risks that are unique to or made worse by generative AI. They are CBRN information or capabilities, confabulation, dangerous, violent or hateful content, data privacy, environmental impacts, harmful bias or homogenisation, human-AI configuration, information integrity, information security, intellectual property, obscene, degrading or abusive content, and value chain and component integration. The profile is voluntary.
On regulation, the European Commission states that obligations for general-purpose AI model providers under the EU AI Act applied from 2 August 2025, that its enforcement powers, including fines, began on 2 August 2026, and that models placed on the market before 2 August 2025 must comply by 2 August 2027 (see the Commission guidelines page, last updated 28 April 2026). Check your own obligations as a deployer with legal counsel.
In practice, governance means an inventory of models and agents, approved data sources, policy that is versioned and enforced at the point of action, and decision traces that an auditor can read. ElixirData Context OS keeps governed context, versioned policy and decision traces in one decision layer above AI agents, and AgentGRC keeps autonomous agent actions inside policy, with the audit trail risk teams need. For the wider programme, read our guide to responsible AI for enterprise customers and the sovereign AI guide for enterprise leaders.
What are the risks and trade-offs of generative AI?
| Risk | How it shows up | Mitigation |
|---|---|---|
| Misinformation and fake content | Convincing but false text, images and video | Provenance labelling, grounded answers, human review |
| Privacy violations and deepfakes | Synthetic media that imitates people | Consent rules, detection and clear disclosure |
| Automated cyberattacks | Faster, more varied attacks | Security testing of models and generated content |
| Bias and discrimination | Skewed training data reflected in outputs | Representative data, evaluation and monitoring |
| Intellectual property | Unclear ownership and legal disputes | Rights review, licence checks and counsel |
| Job displacement and skills loss | Pressure on content, design and coding roles | Role redesign, training and review of reliance on outputs |
How should an enterprise adopt generative AI, step by step?
Start narrow and widen as evidence builds. Use this sequence.
- Inventory and classify. List current use, data touched and owners.
- Fix the data. Use high-quality, diverse, permitted data that mirrors what the model will meet.
- Pilot with a human in the loop. Collect user feedback and measure error types.
- Add controls. Security of data, model and outputs; transparency when users interact with AI; explainability for critical decisions.
- Move to agents in shadow mode. Record what an agent would have done beside human decisions, then raise autonomy as traces show accuracy. ElixirClaw is an agentic OS platform to build, orchestrate and govern AI agents and automation workflows while keeping humans in control.
- Monitor continuously. Watch for bias, unexpected behaviour, data drift and cost, and design for scale.
To map one generative AI decision your agents should make, and the context and policy it depends on, book a Context Workshop. See what happens in the workshop.
What should you plan for in 2027?
One dated EU milestone falls in 2027 for anyone building on or buying general purpose AI models.
- 2 August 2027: the European Commission states that providers of general purpose AI models placed on the market before 2 August 2025 must comply with the AI Act obligations by this date (Commission guidelines page, updated 28 April 2026).
- 2 August 2026: the same page says the Commission's enforcement powers, including fines, took effect from this date, so supplier evidence is now worth requesting.
Ask each model supplier which date applies to the model you use, and confirm your own deployer duties with counsel.
Frequently asked questions about generative artificial intelligence
What is generative artificial intelligence?
Generative artificial intelligence is a class of models that learn patterns from large datasets and produce new text, code, audio, images, video and 3D content, unlike predictive AI, which classifies or forecasts from existing data.
How is generative AI different from agentic AI?
A generative assistant produces content for a person to review. An agentic workflow plans and takes actions across systems, so it needs policy, approval limits and an audit trail as well as accurate output.
How widely is generative and agentic AI used in enterprises?
McKinsey reported on 5 November 2025 that 88% of respondents said their organisations use AI in at least one function, 62% were at least experimenting with AI agents and 23% were scaling an agentic AI system.
What risks does NIST name for generative AI?
The NIST Generative AI Profile (AI 600-1), published 26 July 2024, names 12 risks, including confabulation, data privacy, information integrity, information security, intellectual property and harmful bias or homogenisation.
When do EU AI Act obligations for general-purpose AI models apply?
The European Commission states that obligations applied from 2 August 2025, its enforcement powers began on 2 August 2026, and models placed on the market before 2 August 2025 must comply by 2 August 2027.
When should an enterprise not use generative AI?
Avoid it for decisions with legal effect or clinical consequence without accountable human review, and wherever outputs would be treated as verified evidence without checks.
Follow XenonStack on LinkedIn: XenonStack on LinkedIn for new guides on agentic AI, data and platform engineering.
Explore the XenonStack platform family: ElixirData · ElixirClaw · ElixirHub · NexaStack · MetaSecure · Neural AI · AgentGRC · AgentSRE · AgentAnalyst · AgentInstruct · AgentRAI · AgentScrum · AgentSearch · AgentForce · AI Aviator · SAIF Aviator · CohortIQ · XenonStack AI · XenonStack UAE