The potential of AI-first business systems depends on something fundamental: the quality of the information available to them.
An AI system may be capable of sophisticated reasoning, interpreting complex documents and coordinating activities across multiple applications. But these capabilities are only useful when the system has a sufficiently accurate understanding of the organisation in which it operates.
Most established businesses do not possess a single, coherent representation of how they work.
Information is distributed across enterprise applications, documents, communications, employees, machinery and physical operations. Some of it is authoritative, some incomplete, some contradictory and some no longer relevant.
The challenge is not simply collecting this information. It is establishing what it means, how it relates to everything else and whether it can be relied upon.
AI-first organisations need a dependable shared understanding of their operations. AI could make establishing and continuously maintaining that understanding a normal part of how the business operates.
Where organisational knowledge actually exists
Traditional enterprise systems organise information around particular business functions.
Financial systems maintain transactions and accounts. Asset management systems record equipment and maintenance. Project applications maintain schedules, activities and resources. Document repositories preserve contracts, specifications and procedures.
These systems provide important authoritative records, but none necessarily describes the organisation as a whole.
A considerable amount of business knowledge exists outside formal applications.
Employees understand working practices that may never have been documented. Decisions and commitments are communicated through meetings and correspondence. Machinery and operational systems generate information about actual activity. Physical observations may reveal conditions that differ from those recorded in management systems.
An approved procedure describes how work should be performed. It does not necessarily establish how the work is actually performed.
A meaningful understanding of an organisation therefore requires information from multiple sources, together with the relationships between them.
Building the initial understanding
The first challenge is establishing a usable information foundation.
Attempting to ingest every available document and database into a single AI system would be expensive, difficult to govern and unlikely to produce a reliable understanding by itself.
A more practical approach begins with the business.
What are its fundamental entities, activities, processes, responsibilities and obligations? Which relationships matter? Where are authoritative records maintained? What evidence establishes that particular activities have taken place?
AI could assist with discovering the relevant information sources, interpreting existing records, identifying entities and relationships, recognising contradictions and highlighting missing information.
It could also help gather knowledge from employees through targeted questions, comparing their explanations with documented procedures and operational evidence.
The initial result would be an evidence-linked model of how the organisation appears to operate.
It would not need to be complete or perfectly accurate. Its limitations, uncertainties and unresolved questions would be part of the model itself.
Nor would it necessarily require replacing existing enterprise systems.
Financial transactions could remain in the finance system, contractual records in a controlled repository and operational information in specialised applications.
One coherent understanding does not require one physical database.
The shared information layer would maintain the relationships, context and evidence connecting these existing sources.
Continuous information gathering
Creating an initial understanding is only the beginning.
Organisations change continuously. New contracts are agreed, responsibilities change, equipment is installed, transactions occur, processes evolve and decisions are made.
A static information model would begin becoming outdated almost immediately.
An AI-first system should therefore be designed to gather information continuously, preferably as a normal consequence of business activity rather than through additional administrative processes.
Information could enter through several categories of sources:
Human contributions include decisions, progress updates, observations, explanations and professional assessments.
Formal records include contracts, policies, specifications, approved changes and documented procedures.
Business transactions include purchase orders, invoices, stock movements, financial records and updates to enterprise applications.
Process monitoring captures workflow events, task transitions, approvals and activities performed within business systems.
Machine-generated information includes sensor measurements, equipment status, industrial control signals and operational telemetry.
Physical observations include inspections, photographs, CCTV and computer vision analysis of activity in the real world.
Communications include emails, meeting discussions, supplier notifications and other exchanges through which information and commitments are communicated.
External information includes regulations, third-party data and other information relevant to the organisation's activities.
These sources provide different perspectives on business operations. Some describe intentions and obligations, while others provide evidence of what has actually happened.
AI-generated interpretations should be treated separately as derived information, not as independent primary evidence.
The objective is to maintain a continuously updated understanding by connecting information from these sources to the relevant business entities, processes and relationships.
Validation at the point of capture
Collecting more information does not necessarily produce a better understanding.
An organisation can accumulate enormous quantities of data while remaining uncertain about what is true, relevant or sufficiently reliable for a particular decision.
This suggests that information gathering and validation should be designed as a single integrated process.
Whenever information enters the system, appropriate checks could begin immediately.
Five dimensions are particularly important.
Authenticity: Is the information genuine, and can its origin be established?
Relevance: Does the information relate to the activity, process or decision under consideration?
Quality: Is it accurate, complete, current and sufficiently reliable?
Applicability: Can the information legitimately support the particular conclusion or decision for which it might be used?
Authority: Is the source or person entitled to establish or change the relevant business position?
Some checks can be automated using conventional rules, including identity verification, permissions, timestamps, required fields, valid values and process prerequisites.
AI could assist with more complex assessments, identifying contradictions, examining relationships, comparing information with existing evidence and recognising when a statement does not support the conclusion being drawn from it.
Importantly, validation is not necessarily a binary decision.
Information can be reported, observed, corroborated, verified, disputed or superseded. These are different characteristics and states of knowledge, not necessarily steps in one linear approval process.
An observation can be reliable without being authorised to establish formal completion. An approved decision can subsequently be superseded.
Unverified information may still be useful and should not simply disappear. It can remain available with its source, limitations and validation status clearly recorded.
The system should know not only what information it holds, but why it is considered reliable and what it can legitimately be used for.
From passive collection to active information assurance
Traditional information systems largely depend on people or connected applications supplying information.
An AI-first system could take a more active role.
By maintaining an understanding of business processes, obligations and dependencies, it could recognise when important information is missing, contradictory or no longer current.
An installation reported complete without inspection evidence could trigger a request for verification.
A supplier's revised delivery date could prompt a request for the reason behind the change and an assessment of its consequences.
An approved contractual change that has not been reflected in the delivery forecast could be identified for reconciliation.
An operational measurement inconsistent with expected equipment behaviour could initiate a permitted investigation.
Instead of merely recording that information is missing, the system could request clarification, assign verification activities or seek additional evidence from authorised sources.
This creates a continuous process:
Capture → Interpret → Validate → Identify gaps → Gather further evidence → Update understanding → Revalidate.
The process would need appropriate limits. Not every inconsistency should trigger an intervention, and not every information gap justifies additional work.
The objective is to gather the information necessary to maintain a dependable understanding while minimising unnecessary administrative activity.
From information administration to information assurance
This could produce an important shift in how professional effort is used.
Traditionally, a substantial proportion of skilled time is spent gathering information, organising it, reconciling inconsistencies and preparing it for decisions.
AI could perform much of this preparatory work.
Human judgement could then concentrate more directly on the authenticity, relevance, quality, applicability and authority of the information itself.
A programme manager might spend less time assembling progress reports and more time challenging the evidence supporting forecast completion dates.
An engineer could spend less time searching through technical documentation and more time determining whether available evidence demonstrates compliance.
A commercial manager could spend less time maintaining change registers and more time assessing contractual entitlement, obligations and consequences.
These activities require an understanding of context, professional responsibility and the significance of decisions.
AI can support that judgement, but the value of accountable human assessment remains.
The shift is from people maintaining information so they can exercise judgement, towards AI maintaining information while people concentrate their judgement on its reliability, meaning and consequences.
This would not necessarily eliminate information management or assurance roles. It could substantially change the nature of their work.
Establishing what has actually happened
Consider an equipment installation within a major infrastructure programme.
A supplier reports the installation as complete.
A traditional project reporting process might record that statement in a progress tracker, subsequently incorporating it into a schedule update and management report.
An AI-first information system could establish a more detailed picture.
CCTV or computer vision might provide timestamped evidence that equipment has been physically installed. The supplier's report would provide a separate statement of completion. Installation records could identify the relevant equipment and scope.
An inspector might then verify that the installation meets the required technical criteria. Testing and commissioning records would establish whether the equipment is operationally ready.
These observations and records relate to the same activity, but they do not establish the same facts.
Equipment being physically present does not demonstrate that it has been correctly installed, tested, commissioned or formally accepted.
The system could preserve these distinctions, associate each claim with its supporting evidence and recognise which verification activities remain outstanding.
Where evidence is missing, it could request the appropriate action rather than accepting a single completion percentage.
The Base-e AI Field Intelligence System Using Video venture explores one potential component of this approach: using video to generate structured observations of physical activity and progress.
Combined with contractual, technical and acceptance records, such observations could contribute to a more reliable understanding of actual delivery.
Measuring the quality of organisational understanding
An additional possibility is that organisations could begin measuring the reliability of their understanding rather than simply the amount of information they possess.
A system might identify which milestones have verified supporting evidence, which commitments depend on uncorroborated statements and which forecasts rely on assumptions that have not recently been reviewed.
It could highlight unresolved contradictions, outdated records and processes where evidence of compliance is insufficient.
This would provide a different perspective on management information.
The organisation would not simply know that a milestone is forecast for a particular date. It could understand the evidence supporting that forecast, its important assumptions and where uncertainty remains.
Numerical confidence measures might sometimes be useful, but they should not replace transparent evidence, explicit validation criteria or professional judgement.
The quality of an organisation's decisions depends partly on the quality of the information supporting them.
Making that quality visible could improve both the focus of management attention and the predictability of business outcomes.
Maintaining understanding without surrendering control
A continuously maintained organisational information model introduces significant security and governance responsibilities.
The system would potentially connect information across departments, applications and security boundaries.
Relationships between records may reveal sensitive information that is not apparent from the individual records themselves.
Information gathering therefore needs to respect source permissions, confidentiality requirements and information classification from the outset.
AI should not have unrestricted access to information simply because it has identified a potentially useful relationship.
The model must also distinguish authoritative records from interpretations and recommendations.
An AI-generated assessment should not silently change a contractual obligation, financial commitment or formally accepted business position.
Provenance, version history, validation rules and appropriate authorisation remain essential.
The purpose of the shared information model is to improve organisational understanding, not to transfer business authority to the AI maintaining it.
A permanently maintained understanding of the organisation
The initial development of an organisational information model should not become another large, one-off enterprise transformation exercise.
A practical implementation could begin with a bounded process or operational area, establish the relevant information sources and validation rules, and demonstrate value before expanding further.
As the system develops, additional activities, relationships and information sources could be incorporated.
New applications and specialised AI agents could use the established understanding rather than independently reconstructing the same organisational context.
This connects directly to the problem of information entropy.
Organisations continually generate new information, different interpretations and changing circumstances. Maintaining a consistent understanding has traditionally required substantial effort through reporting, meetings, reconciliation and governance.
An AI-first information system could address this at the point where new information enters the organisation, ensuring that it is interpreted, connected and appropriately validated as part of its creation.
The longer-term opportunity is not merely a more comprehensive enterprise database or a better way to search business information.
It is an organisation that continuously develops and maintains an increasingly reliable understanding of itself, with human effort concentrated on the judgement and assurance needed to make that understanding dependable.
That shared understanding could become one of the foundations of AI-first business operations.