Why AI needs context to produce useful work

Artificial intelligence can generate convincing answers without really knowing your company. Discover how to provide the context it needs to work coherently: relevant information, current decisions, instructions, limits and quality criteria.

Imagine asking an AI tool to write a page presenting one of your company’s services. You briefly explain your offering and request professional, persuasive copy.

Seconds later, a proposal arrives with an attractive title, organised structure, seemingly reasonable arguments and even a call to action. At first glance, the work is finished.

But reviewing it reveals problems. AI presents your company as a low-cost alternative although your positioning centres on specialisation. It promises delivery times you never guaranteed, uses commercial language unlike your way of working and addresses a different audience from the one you prioritised.

The text is well written.

But it does not represent your business properly.

This is not necessarily a linguistic capability problem. Nor does it mean finding a secret formula for better instructions. Something more important may be missing.

AI knows the assignment but not the business decisions that should guide it. It does not know which promises you can sustain, which customers have priority, which arguments reflect your value proposition or which information is approved.

When these elements are absent, the system may fill gaps with common patterns, assumptions or plausible claims. The result can seem professional yet remain unsuitable.

This matters because using AI in a company is more than producing more texts, reports or proposals. It means developing work that fits processes, respects decisions and helps achieve specific objectives.

For that, the tool needs more than a good question.

It needs business context.

That context must be selected, current and accompanied by instructions, limits and verification criteria.

AI can produce an answer without knowing your company. Producing useful work, however, requires understanding what it must do, which reality to respect and how to check whether the outcome is suitable.

1. A correct answer can still be useless to your company

Generative AI tools can write, summarise documents, compare alternatives, organise information and assist tasks of varying complexity. Their capability is considerable.

But we must distinguish an answer that seems correct from a result usable within a business.

Formal quality does not guarantee business usefulness

Imagine preparing a commercial proposal. AI structures it correctly, with an introduction, benefits and an invitation to purchase. Yet it uses a price from an old offering version. Perfect writing can still be commercially incorrect.

Now consider strategic analysis. The tool identifies growth opportunities, suggests services and organises priorities. But one opportunity was already examined and rejected because the company lacks operational capacity. The recommendation may seem intelligent yet ignore a relevant decision.

Even apparently simple tasks can fail: a customer response promises an unavailable condition; an article uses unverified data; an internal document mixes criteria from different projects.

The problem is not necessarily writing.

It is the distance between what the tool produced and what the business actually needs.

AI does not automatically know your company’s decisions

A tool may have general knowledge of marketing, architecture, restaurants or other disciplines. It can know common practices and suggest reasonable solutions.

But general knowledge must not be confused with an organisation’s specific reality.

It does not necessarily know approved commercial decisions, rejected projects, available services or operational providers. Even with access to business documents, information may be incomplete, contradictory or outdated.

We must therefore provide references needed for the work. This does not prevent new ideas: it helps distinguish current business reality from alternatives still needing assessment.

A proposal can be creative without presenting a hypothesis as an approved decision.

The assignment’s purpose changes what makes an answer good

Exploring ideas for a new service differs from writing the final page for an approved one.

Exploration can allow alternatives and experimentation. Final copy requires greater precision: commercial terms must match the current offering, arguments need support and public claims require review.

Likewise, preliminary research can present hypotheses, while a decision report must distinguish facts, uncertainties and recommendations.

Context should therefore explain how the result will be used. Knowing the subject is not enough: AI needs to know its work’s function.

2. Information, context, instructions and judgement: four different elements

Giving AI context can sound like simply attaching more documents. But plentiful information does not guarantee correct interpretation.

Four elements should be distinguished.

The four relate, but serve different functions.

Information: what we know or have available

A company may have commercial presentations, strategy documents, website pages, emails, quotes and conversations. All contain potentially useful information.

But not all are relevant to every assignment or represent current decisions.

An old quote can explain an offering’s evolution without its prices remaining valid. A rejected proposal can explain a previous decision without automatically being an available service.

Context: what matters for this task

Context selects and organises information needed to interpret the assignment properly.

For a commercial page, useful knowledge includes the service presented; its intended people, helped by our guide to ideal customers and buyer personas; the problem solved; differentiation; actual terms; language representing the business; and the action to facilitate.

We do not need every document produced over ten years. We need references that materially change work quality.

OpenAI’s technical documentation identifies relevant context as an element helping models work with specific information, including business knowledge absent from training data. [1]

Instructions: what the tool must do

Instructions define the task. They can establish format, recipient, depth, restrictions and conditions the result must meet.

For example:

“Prepare a service page for owners considering a comprehensive renovation. Use only approved services and flag any claim requiring verification.”

This is more precise than:

“Write professional copy about renovations.”

But even excellent instructions can be insufficient if the approved offering is unavailable or documents contradict.

Good instructions guide the work. Context provides the reality it must work with.

Judgement and authority: who decides what is valid

This fourth element is particularly important.

AI can compare options, apply defined criteria and propose decisions. With suitable tools and permissions, it can also execute certain operational decisions.

That does not mean unlimited authority over the business.

Someone must establish the current strategy, permissible commitments and who approves particular consequences.

A tool may propose a positioning change. That recommendation may be interesting, but should not automatically become official strategy. The same applies to prices, contract terms and public claims.

The ability to generate a recommendation is not the authority to approve it.

Responsibility must remain clearly assigned even where operations can be delegated.

3. Practical example: the same assignment with and without business context

Let us apply this distinction.

Imagine an architecture studio specialising in renovating traditional homes in the Pyrenees. It works with owners restoring existing properties through carefully considered architectural projects.

Its proposition centres on specialisation, knowledge of the built environment and attention to each property’s particular characteristics. It does not compete on lowest price or provide minor household repairs and general maintenance.

It wants a website page for comprehensive home renovation and uses AI for an initial content proposal.

This case is entirely hypothetical. The assignments and extracts are editorial examples demonstrating the method, not tests of a particular model or results for a real client.

First assignment: correct but insufficient instructions

The studio gives AI this request:

Write a professional, persuasive page for our home renovation service in the Pyrenees. Include benefits, a service explanation and a call to action.

The request is understandable, but leaves questions open.

Which renovations does the studio actually perform? Who are its customers? Does it offer construction execution or only architectural services? Does it compete on price, speed, specialisation or something else? Which commitments can it make? Which services are excluded?

AI might fill these gaps with common renovation sales language.

For example:

We renovate your Pyrenees home quickly and affordably.

We offer turnkey solutions, competitive quotes and tight deadlines to transform your house without complications.

This might fit some companies, but we have no reason to claim it correctly represents our studio.

It introduces problems: competitive prices are not its proposition, speed is promised without known timescales, and “turnkey” could imply services or responsibilities outside the offering.

AI generated plausible commercial copy without enough information to know its limits.

Second assignment: adding business context

Now imagine the studio has basic strategic documentation. It need not provide its entire archive. It selects relevant information and prepares this context:

The tool now has a much more specific reference. It does not necessarily know every business detail, but receives decisions that should shape its work.

What can change in the result

With this context, an illustrative proposal might begin:

Renovating traditional homes in the Pyrenees

Every existing home has characteristics worth studying before planning its transformation.

We develop architectural renovation projects considering the property’s conditions, its owners’ needs and the intervention’s relationship with its surroundings.

If you are considering renovating a traditional home, we can begin by studying your project’s characteristics and the scope of services you need.

This extract fits the described positioning better. It promises neither savings nor deadlines, does not present the studio as a general repairs company and proposes a next step proportionate to the commercial relationship.

But it is not automatically correct. The studio must confirm it actually offers these services and that claims about its working methods reflect real practice.

The comparison

The second assignment has better conditions for useful work, but does not guarantee a superior answer every time. AI can still err, misinterpret sources or omit important information. Improvement must therefore be evaluated.

Context also reveals missing information

Imagine the studio wants to mention experience with protected buildings. The tool finds previous project references but no confirmation of specific expertise it may communicate.

Suitable behaviour would identify the gap:

Available documentation does not confirm the scope of experience with protected buildings. This claim requires validation before inclusion on the page.

That answer may be less striking than an invented commercial argument, but is far more useful to responsible production.

Context’s value is not only better writing. It also defines what we know, what we do not know and what must be checked before action.

4. What AI needs to know to work with your business

Not every assignment needs extensive documentation. Simple tasks, such as correcting spelling or summarising a provided document, can need only a few instructions.

Public content, strategic analysis and tasks affecting business processes require a stronger foundation.

A simple tool can organise preparation.

Context brief for AI assignments

This is not intended as a mandatory form for every conversation. It helps assignments remain coherent with an organisation.

A complete assignment without an endless prompt

Return to the architecture studio. We could formulate:

Assignment: draft a page presenting comprehensive renovation of traditional homes.

Objective: help interested owners understand service scope and decide whether to request an initial conversation.

Context: use current documentation on the studio’s offering, positioning, priority audience and tone.

Limits: do not promise unconfirmed deadlines, savings, prices or services. Do not present hypotheses as facts.

Result: propose structure and copy with a main title, sections, arguments and call to action.

Control: separate claims needing validation. Do not consider the result approved for publication.

There is no magic phrase here. We reduced ambiguity, defined work and supplied specific evaluation conditions.

When more context is needed

Information quantity should depend on the task.

Summarising meeting minutes may require only the document and desired summary type. A commercial page needs offering, audience, positioning and publication conditions. Digital investment analysis may require objectives, costs, capabilities, alternatives and risks.

Operating on a real system additionally requires appropriate permissions and controls.

We seek sufficient knowledge for a specific task, rather than the largest possible context.

5. Which documents to use and how to resolve contradictions

Many companies already possess much of the knowledge needed for AI work. The problem is that it is scattered across strategy documents, sales presentations, conversations, quotes, website pages, reports and decisions never formally recorded.

These materials may also represent different business moments: old presentations mention discontinued services; quotes contain exceptional terms; commercial proposals await approval; recent conversations suggest undecided changes.

Providing all material without organisation increases the risk of mixing incompatible realities.

Sources have different authority

Imagine the architecture studio retains three documents: an approved strategy prioritising comprehensive renovation; a draft advisory line for minor renovations; and an old commercial presentation mentioning household maintenance.

Asked to write a services page from all three, AI may infer every service is available. That would be wrong. We must explain what each source represents.

Classifying documentation helps work

A simple structure can be used:

Classification must reflect actual organisational rules. An “approved” label alone is insufficient: know who may approve and that approval’s scope.

The newest document is not always correct

Suppose someone drafted a commercial proposal yesterday, while positioning was formally approved two months ago.

Yesterday’s document is newer but may await assessment. It should not automatically replace approved strategy.

Consider date, version, status, scope and decision authority. Record replacements too, especially when several people or providers work on the same assets.

What AI should do when it finds a contradiction

Imagine two current sources give different service prices. The tool should not automatically choose the newest, average them or invent a rate.

First check whether scopes differ: one may concern individuals and another companies. If contradiction remains, flag it and request the appropriate responsible person’s decision.

A useful answer might be:

Documentation contains two different commercial conditions for this service. I have not identified an authorised source resolving the discrepancy. The current condition must be confirmed before use.

This behaviour is not guaranteed merely by requesting it. The process also needs review mechanisms.

Traceability makes claims verifiable

For certain work, preserve links between important claims and sources.

For example:

This does not mean publishing internal references on every commercial page. Traceability can remain in editorial production.

Relevant claims should be verifiable. AI-generated citations do not guarantee sources exist or support assertions. Important data require verification.

AI can organise knowledge. The company must retain criteria to distinguish confirmed information from matters still requiring decisions.

6. More context does not mean providing all information

Learning that context can improve AI work naturally suggests that more documents may improve answers. But not always.

Large collections can include irrelevant information, repetition, contradictory versions and data that should not be shared. Models also have information limits and do not always retrieve or use every long-context detail correctly.

OpenAI documentation distinguishes direct information provision from retrieval from a document base, and notes that long-context and retrieval systems need specific evaluation. [1][2]

Five conditions for useful context

Relevance. Information should actually contribute to the assignment.

Quality. Data and documents must be sufficiently reliable for their function.

Currency. Know which decisions remain valid.

Authorisation. The tool must legitimately use information in this task.

Separation. Data from other projects, clients or areas must not be improperly mixed.

These matter more than documentary volume alone.

Context is not training

Another common confusion needs clarifying: attaching documents to a conversation does not necessarily permanently train a model on them.

Business knowledge can be used through different mechanisms.

Direct context: information provided within an assignment or conversation.

Document retrieval: a system consults authorised sources and supplies relevant excerpts when responding.

Operational state: an application supplies current task, process or operation information.

Additional tuning or training: specific, distinct methods modify model behaviour.

These are not equivalent.

Document retrieval, commonly called Retrieval-Augmented Generation or RAG, retrieves external information and includes it in model context during a task.

It can help with a large document library, but small companies do not necessarily need RAG for contextualised responses. They can begin by manually providing relevant information.

Retrieving a document does not guarantee correct use

Imagine a system consulting every studio document retrieves an old rate when asked about service terms.

The problem could be classification, retrieval selecting the wrong source or subsequent interpretation.

Distinguish three questions:

Do we have correct, current information?

Does the system retrieve appropriate information?

Does AI use it correctly?

Different problems need different checks.

Confidential information needs specific limits

Context may contain customer, employee, project, pricing, contract or internal process information. We should not share it with any tool merely because answers might improve.

Personal data processing must respect GDPR principles such as purpose limitation, minimisation, security and accountability. [5]

Also consider confidentiality agreements, document rights and provider conditions.

Sometimes aggregate or anonymised data suffice. Other cases need restricted access or appropriately safeguarded configurations and services.

The best context enables work with necessary information within correct limits, rather than containing the most information.

7. How to check whether AI’s result is actually useful

Organised context improves working conditions but does not prove correctness. AI can misinterpret a document, omit requirements or make unsupported claims. Equivalent assignments can also produce different answers.

We therefore need work quality checks.

Do not evaluate only whether you like the answer

Imagine two commercial page proposals. The first is attractive, striking and persuasive. The second is restrained but represents the offering better and avoids unconfirmed promises.

Which is better? It depends on criteria.

If originality or sales force alone matters, the first may appeal. If a page must accurately reflect studio positioning, the second may suit better.

Evaluation must match work function. Our guide to what to measure in digital marketing applies the same principle to business results.

Six practical quality criteria

These adapt to different tasks. Internal summaries need not face contractual proposal standards, but we must know what good work means.

A small comparison exercise

Return to the architecture studio. To test whether contextual documentation improves common tasks, choose five representative assignments:

1. Write a service page.
2. Prepare frequently asked question answers.
3. Review a commercial proposal’s consistency.
4. Summarise technical documentation internally.
5. Develop a website content outline.

For each, prepare required-result criteria. Compare generic instructions with selected business context under reasonably equivalent conditions.

Repeat tests when needed: one answer does not demonstrate general improvement.

What we would record

A complex evaluation platform is unnecessary initially. A simple table can work:

The table shows possible criteria and outcomes, not experiment measurements.

OpenAI evaluation documentation recommends representative task sets and defined criteria to detect errors and check change effects. [2][3]

The outcome we seek

We want to see whether context reduces important errors and unsupported claims, improves offering consistency and brings work to review with fewer material problems.

We can also observe team correction effort. But an impression of speed should not automatically become an economic saving figure; that needs suitable measurement.

Human review needs clear responsibility

Asking AI to review its output can detect some contradictions, structural issues or apparently missing information. It is not an independent accuracy guarantee.

For public content, commercial decisions and sensitive work, establish who verifies relevant claims and authorises use. Review depth should match impact.

General management need not approve every internal summary, but a tool should not unilaterally publish unconfirmed commercial conditions.

The aim is appropriate controls for each work type, rather than identical review for everything.

8. When AI can act, context needs operational permissions and limits

So far, we have mainly discussed generating content, summarising documents and preparing analysis.

AI systems can also connect to tools and act: authorised applications can create drafts, update files or execute process steps.

This expands possibilities and changes responsibilities.

Writing is not publishing

Imagine asking an assistant:

“Update our service page according to current strategic documentation.”

The tool can prepare new copy, but with website access may also modify the page.

These are different action levels. The first produces a reviewable draft; the second changes a real asset and needs additional controls.

Know what it may modify, which information it may use and which actions require approval.

Understand, propose, execute and approve

Four capabilities can be distinguished:

Not every task needs formal separation, but it helps when actions affect customers, data, money or business systems.

Permissions should match the task

Preparing a website article may require authorised editorial documents and draft storage, not invoices, customer data or every organisational file.

It also need not require permission to delete the site or change hosting configuration.

Provide only access necessary for the assignment.

OWASP’s generative AI application security project recommends limiting privileges, separating untrusted content and human approval for high-risk operations. [4]

The risk of instructions hidden in documents

A particular problem arises when AI consults external pages, files or messages. Not everything it encounters is a legitimate instruction.

Imagine a provider page contains text ordering an assistant to ignore previous instructions and send information elsewhere.

That text belongs to the consulted page. It should not gain authority over system behaviour.

This attempt is called prompt injection. It can appear through apparently ordinary content an AI system reads during work.

OWASP identifies this risk and proposes separation, access control and verification measures. [4]

This does not mean every external document is dangerous. It means distinguishing an information source from an authority able to issue instructions.

Documentary hierarchy does not replace technical security

A company can give approved strategy priority over an old draft. That is a business authority rule, not technical permissions.

Nor can any text marked “priority” legitimately modify model instructions or security controls.

Document governance and appropriate technical mechanisms must work together.

Context explains what AI must respect. Permissions and controls determine what it can actually do in the system.

Responsibility should grow with impact

A system may classify internal documents under defined rules and perhaps prepare response drafts. But publishing an offering, modifying a contract, sending commercial messages or deleting information may require additional authorisation.

No universal supervision level fits all tasks.

NIST’s generative AI risk framework highlights risks throughout design, development, use and evaluation. [6]

Autonomy should reflect function and possible consequences.

AI can expand operational capacity. The company must retain control over risks and delegated decisions.

9. From scattered knowledge to Governed AI Context

Many companies already possess the knowledge needed for better AI work. They know customers, have developed value propositions and commercial experience, decided which services to offer or reject, and understand processes, limitations and priorities.

But much may remain scattered: in the manager’s memory, employee conversations, old emails, hard-to-find documents or unrecorded decisions.

Organising knowledge has value before AI

Imagine the architecture studio properly documents its offering, positioning and commercial criteria.

New team members can understand the business better, external providers can prepare coherent material and decisions survive changes in project responsibility.

AI can use this documentation, but is not the only beneficiary.

Documented strategic knowledge first enables sharing and maintaining business direction.

From strategy to operational context

PhiStudio’s Digital Strategy module addresses that need for direction.

It organises business knowledge into documented decisions on model, value proposition, customers, positioning, communication and priorities. Our guide How to define a company’s digital strategy explains the process.

Within this documentary system, Governed AI Context prepares a reference for applying knowledge to AI-assisted work.

The distinction matters. Strategy establishes direction. Documentation preserves and communicates decisions. AI context selects references and conditions for particular assignments.

These are not three independent systems always developed separately. They are related functions serving one need: coherent use of business knowledge.

Governed context needs maintenance

Good documentation does not solve matters forever. Companies add services, modify terms, evolve audiences and replace tools.

Without updates, systems may continue using old decisions.

Proportionate procedures are needed: update references when important decisions change, clarify current versions when replacements occur, distinguish historical documents from current conditions.

Periodically check that AI systems still retrieve and use appropriate sources.

We need not rebuild all documentation weekly. We need to keep decisions affecting work current.

Preparing context is not implementing an AI platform

This commercial distinction deserves clarification.

A company can have Governed AI Context without developing a custom assistant, automated knowledge base or agent connected to applications.

Strategic documentation prepares knowledge for different uses. Tools, integrations or agents are additional implementation work requiring assessment of functions, costs, security and responsibilities.

Not every company needs that implementation. Some gain value by manually supplying organised information for specific tasks; others can justify advanced systems.

The decision depends on workload, processes and opportunities sought.

First understand which knowledge the business needs for coherent work. Then decide which tools can sensibly use it.

10. AI expands working capacity; direction establishes purpose

AI can help perform tasks previously requiring more time, resources or specialist knowledge: structuring information, writing content, comparing alternatives, preparing analysis and assisting processes.

But producing results does not mean automatically knowing every organisation’s reality.

An answer can be convincing yet misrepresent the business. A recommendation can seem reasonable yet contradict a current decision. Automation can correctly execute an action that should never have been authorised.

We therefore need more than technology: purpose, relevant information, documented decisions, clear instructions, suitable limits and verification criteria.

Not every task requires extensive strategy documentation. Simple work can use AI simply. But important processes increase the relevance of consistency and responsibility.

Context helps explain what to do and which reality to respect. Documentation preserves and shares it. Controls limit access and actions. Evaluation checks the expected function.

None eliminates errors entirely or guarantees security through organised documents alone. But it establishes better conditions for AI in real business processes.

Organised knowledge also retains value when technology changes. A clear proposition is independent of a model; documented commercial decisions guide different professionals; coherent information architecture aids people and new tools.

At PhiStudio, this is how we understand Digital Asset Direction. The aim is not AI in every task merely because it exists. It is understanding business objectives and deciding how digital capabilities serve that direction.

AI can produce more work. Appropriate context and clear direction help make that work meaningful to the company.