When People and AI Systems Work Together, Who Maintains a Single Version of Truth?

After AI Enters the Workflow · Season One: “From Answering Questions to Participating in Work” · Article 9

A project team has five people and three AI assistants. One assistant summarises meetings, another conducts research, and a third drafts reports. The project manager changes the delivery date in a planning system. A meeting transcript still contains the old date. The research agent has stored an earlier summary, while a staff member tells the writing assistant about the change in a private chat.

Two days later, the draft report gives one date, the task board another, and an email to a supplier a third. None of the systems invented the dates. Each used information that had once been plausible.

The failure is not that AI forgot a fact. The team lacked a maintained relationship between conversation, record and authority. Adding more capable assistants will not solve that problem. It may allow inconsistent information to travel faster.

When multiple people and AI systems work together, a “single source of truth” cannot mean that every participant possesses the same complete memory. It means that for each consequential fact, the group can identify the governing record, its current status and the process by which it may be changed.

Collaboration does not require everyone to know everything

Complex work has always depended on distributed knowledge. Engineers, lawyers, finance staff and customers each see different parts of a project. No person holds the whole picture.

Coordination succeeds because institutions create records, roles and interfaces. A signed contract governs price even if many team members have not read every clause. A project plan records the current schedule. A decision log identifies an approved change. People do not need identical knowledge; they need reliable ways to locate the knowledge relevant to their responsibility.

AI intensifies this old problem. Each assistant receives a different context. A research agent may see public sources but not internal decisions. A drafting agent may see the report template and a summary but not the raw meeting record. An automation agent may read structured fields without access to the discussion that explains an exception.

The objective should not be to copy every fact into every context. That increases privacy exposure, cost and confusion. The objective is to maintain explicit sources and controlled handovers.

A single source of truth is not simply one file

The phrase often produces an image of a master spreadsheet or central document. Sometimes that is useful. But complex work contains different types of truth with different owners.

The customer record may control identity and contact details. The signed agreement controls contractual commitments. The project system controls the approved delivery plan. The issue tracker records defects. A policy repository controls organisational rules. A single giant file would become both unmanageable and inaccurate.

“Single” should therefore describe authority within a defined domain, not physical storage. For any relevant fact, the system should be able to answer:

• Which record is authoritative for this purpose?
• Who owns and may change it?
• When did it become effective?
• What did it replace?
• Which derived reports or indexes depend on it?
• How will conflicts be detected and resolved?

The truth system is a network of governed records and relationships. AI needs access to that network in a form that preserves provenance, not merely a folder containing many documents.

Multi-agent systems amplify handover failures

Multi-agent architectures can divide a large problem among specialised agents. Anthropic has described a research system in which a lead agent delegates parallel research to subagents and then synthesises their results. This can expand coverage, but the designers also have to manage coordination, context and evaluation. Anthropic, “How we built our multi-agent research system”

Google has similarly discussed production architectures for context-aware multi-agent systems, including shared context, orchestration and efficient communication. Google Developers Blog, “Architecting efficient context-aware multi-agent framework for production”

The more handovers a workflow contains, the more opportunities arise for meaning to change. One agent reports that a proposal is “likely”; another compresses this as “planned”; a third writes it as an announced decision. Source dates disappear. Qualifications are shortened to fit context. Separate agents perform internally reasonable work while the chain drifts.

The solution is not to require agents to exchange unlimited natural-language histories. Handover formats should distinguish facts, inferences, decisions, unresolved questions and requested actions. High-value claims should carry source references. A receiving agent should know whether it may rely on an item or must verify it.

Shared facts need provenance relationships

Provenance records where information came from and how it was transformed. The W3C PROV Ontology provides a standard vocabulary for describing entities, activities and agents and the relationships among them. W3C, “PROV-O: The PROV Ontology”

An organisation does not need to expose that vocabulary to every user, but it needs the underlying distinctions. A report value should be traceable to a database snapshot; a summary to a document version; a decision to an authorised meeting or approval; a generated recommendation to the inputs and model process that produced it.

Without provenance, correction becomes local and incomplete. A person fixes the report but not the stored summary that will generate next month’s report. A project date changes but the external communication agent continues to retrieve an old note. The work system has several copies but no dependency map.

Provenance turns correction from editing a sentence into maintaining a network. It permits the organisation to ask which downstream artefacts must be reviewed when a source changes.

People and AI need explicit write authority

Read access and write authority should not be confused. An assistant may need to consult an approved project plan without being allowed to alter the official delivery date. It may propose a correction to customer details but should not necessarily commit it.

For each source of truth, the organisation should define:

• who may create and approve records;
• which systems may update them automatically;
• what validation is required;
• whether changes are reversible and versioned;
• who receives notice of material changes;
• how disputes are escalated.

Agents should generally write proposals or structured change requests unless the domain is low risk and the validation is objective. Separation between “suggest” and “commit” prevents an inference from silently becoming a fact.

NIST’s AI RMF Core emphasises defined roles, responsibilities, documentation and monitoring. These requirements are especially important when outputs pass among models and teams, because responsibility can otherwise dissolve at each interface. NIST AI Resource Center, “AI RMF Core”

What meetings and chats should preserve

Conversation is useful for exploration, but not every sentence should become an official record. A meeting contains questions, options, disagreement, jokes, provisional statements and decisions. Saving a transcript preserves words without identifying their status.

A better process produces structured outcomes:

• decisions made and their authorised decision-makers;
• actions, owners and due dates;
• open questions and evidence still required;
• proposals explicitly not yet approved;
• changes to existing authoritative records;
• links to the underlying transcript where further context is needed.

AI can help extract these elements, but participants should confirm consequential outcomes. The confirmed decision should then update the proper system rather than remain only in a meeting summary.

The same applies to private chats. A message can alert someone to a change, but it should not become the only place where a project-wide fact exists. If the information matters to others or to an automated workflow, it must be promoted into the governed record.

Conflict is not failure; unmanaged conflict is

Two sources may disagree for legitimate reasons. A contract may state the original date while a later approved variation states a new one. Finance and operations may use different definitions of “completed”. A customer may dispute an internal record.

A trustworthy system should not force premature consistency by selecting whichever document was retrieved first. It should expose the conflict, show the authority and dates of each source and route the issue to a role capable of resolving it.

Context-engineering guidance for agents stresses selecting and maintaining relevant information rather than allowing context to grow without control. Anthropic, “Effective context engineering for AI agents” In organisational work, relevance includes epistemic status: whether an item is settled, contested, superseded or awaiting confirmation.

The ability to represent disagreement is part of maintaining shared truth. A clean but false single answer is less useful than an explicit unresolved conflict.

Conclusion: shared truth is ongoing maintenance work

No AI assistant, database or document becomes a single source of truth merely by being central. Authority comes from maintained ownership, status, versioning and change processes.

The guiding principle is:

In human–AI collaboration, every consequential fact should have an identifiable authoritative source, and every transformation of that fact should preserve a path back to it.

People need not know everything, and agents should not receive every record. They need reliable handovers and the ability to distinguish facts from proposals, decisions from summaries and current rules from history.

Maintaining this structure is not clerical overhead added after the “real” work. It is what allows distributed work to remain one project rather than many locally coherent but mutually inconsistent stories. AI can retrieve, compare and update records, but someone must govern the system in which one version becomes authoritative and another becomes history.

Primary sources and further reading

• Google Developers Blog: Architecting efficient context-aware multi-agent framework for production
• Anthropic: How we built our multi-agent research system
• Anthropic: Effective context engineering for AI agents
• W3C: PROV-O—The PROV Ontology
• NIST AI Resource Center: AI RMF Core

Continue reading: Explore the After AI Enters the Workflow series.