An AI assistant can produce a convincing answer while missing the policy, customer record or operational detail that determines whether it is useful. For enterprise leaders, that gap affects trust, review effort and the cost of adoption.
Context engineering addresses how an AI system receives and uses the information needed for a task. At InfraTech, our view is that this deserves attention alongside model selection. A capable model still needs relevant evidence, clear boundaries and access to the right business systems.
What does context engineering mean?
Context engineering is the design and management of the information available to an AI model when it works. Prompt engineering focuses on the instructions you write. Context engineering extends that work to source material, conversation history, memory, tools and output requirements. Anthropic describes this broader, iterative approach in its engineering guidance.
Consider an assistant answering a product warranty question. It needs the correct model, market, purchase details and current warranty policy. A well-written instruction cannot substitute for missing evidence. Context design determines which records are retrieved, whether they apply, and what happens if the answer cannot be established.
What market evidence tells leaders
UK government AI Adoption Research, published in 2026, found that 29% of surveyed businesses cited AI projects being too complex or difficult to integrate and scale as a barrier. Another 21% cited data complexity, including missing or unstructured data. The research covered 3,500 businesses, with fieldwork conducted between February and May 2025. These figures describe reported current or past adoption barriers, rather than failure rates.
More recent ONS evidence also points to limited depth of adoption. Among UK businesses with 10 or more employees that used AI, only 10% reported extensive use in June 2026. The ONS cautions that this newly introduced measure is an early indication.
Neither study proves that context engineering causes better business outcomes. InfraTech’s interpretation is practical: integrating AI into real work requires attention to the information, systems and controls surrounding it. Buying access to a model is only one part of that task.
The information an enterprise AI system needs
A useful context design brings several elements together. The requirements will differ by workflow, but each should have a clear purpose and owner.
- Instructions: define the task, boundaries, priorities and response to missing evidence.
- Retrieved evidence: supply relevant, current information from approved documents and business records, with source identifiers and dates.
- Working state: retain the task’s decisions, constraints and unresolved questions across multiple steps.
- Persistent memory: retain approved information between sessions where useful, with ownership, access and update rules. Persistent memory is an application feature, not a property of every model.
- Tools: provide controlled access to searches, calculations or business systems, with clear permissions and failure handling.
- Output requirements: define the fields, citations or structure downstream users need, and validate them before use.
How retrieval helps and where it can fail
Retrieval-augmented generation, often called RAG, finds relevant material and supplies it to a model before it answers. It can connect an assistant to enterprise information without relying entirely on knowledge learned during training.
Retrieval is one component of context engineering. It does not guarantee a correct answer. A system can retrieve an outdated policy, miss a critical exception or combine records that apply to different products or customers. The model can also misinterpret accurate material.
For the warranty example, searching for similar wording is insufficient. The retrieval process should also check product identity, region and policy validity. If applicable evidence is missing or conflicting, the assistant should request clarification or route the question for review.
Access controls must operate in the application and source systems. A prompt telling the model to respect confidentiality is not a replacement for permission checks.
Why more context is not always better
Loading every available document can increase cost and response time while introducing irrelevant or conflicting information. Context selection should be tested for relevance, completeness and freshness.
The research paper Lost in the Middle, published in 2024, found that performance on retrieval and question-answering tasks could deteriorate when relevant information appeared in the middle of long inputs. This was a finding about the models and tasks studied, not a universal performance limit for every current model.
The lesson for an enterprise pilot is to test different context lengths and arrangements. Preserve critical evidence when summarising long conversations, and re-retrieve authoritative records when they may have changed. A larger context window does not remove the need for information management.
Measure reliability against the workflow
Start with one bounded workflow and a test set that reflects real operational conditions. Include ordinary requests, missing evidence, outdated documents, conflicting policies and restricted information. Treat instructions embedded in retrieved material as untrusted content, and test whether they can divert the system from its task.
Compare the existing process with the proposed AI-assisted process using measures such as:
- Answer accuracy and whether cited evidence supports the answer.
- Whether retrieval finds the necessary, current records.
- Appropriate escalation when evidence is missing or conflicting.
- Respect for access boundaries and controls on external actions.
- Human review time, response time and cost per completed task.
Validate structured responses for both format and meaning. A response that fits the required data structure can still contain incorrect values. Clear instructions, grounding and human review can reduce risk, but cannot eliminate hallucinations or replace evaluation.
Where InfraTech AI Transformation Advisory fits
Context engineering becomes useful when it is connected to a business decision. Before selecting an implementation, leaders need to know which workflow warrants investment and what information it depends on.
InfraTech’s AI Transformation Advisory provides an evidence-led starting point across enterprise reality, AI readiness, AI economics, and governance and risk. It brings operational evidence into a transformation baseline, prioritised investment roadmap and board-ready investment case.
For a proposed AI assistant, context requirements can inform that assessment: which sources exist, what preparation is needed, which risks require controls and what operating costs should be modelled. Detailed implementation and ongoing operation should be scoped separately from the advisory engagement.
Your next step
Choose one workflow where staff spend time finding information or checking AI answers. Ask which evidence the system needs, whether that evidence is current and accessible, and how you will recognise a wrong answer.
Those questions create a practical starting point for context engineering and a stronger basis for an AI investment decision.
Book a free exploratory session with InfraTech Systems to discuss your organisation’s AI priorities. Call 020 3866 1180 or email contact@infratechsystems.co.uk.
