One of the first ideas in an AI project is often to connect the agent to everything.
Give it access to the company wiki. Add the shared drive. Connect the CRM, project system and customer conversations. Include the prompt library everyone has been building.
Now the agent has context.
Except it may have several versions of the pricing policy, an old proposal template that people still use, a task describing a temporary exception and meeting notes that contradict the official process.
It has more information than any one employee could read.
It still does not know what the business wants it to do.
Access to knowledge and guidance for action are different things.
Retrieval is not understanding
A good knowledge base helps people find what the organisation knows.
That is valuable. AI makes it easier to search, summarise and combine that information.
But a search result does not arrive with organisational authority attached to it.
The agent may find:
- a strategy written last year;
- a process that describes the official approach;
- a task showing how the team handled one unusual case;
- a manager's message that changed the approach for a month;
- a proposal that won a large deal; and
- delivery notes showing that the deal was unprofitable.
Every item may be true in some sense.
Which one should guide the next proposal?
The problem is not that the AI needs more documents. The problem is that the organisation has not made the relationship between those documents clear.
More access gives AI a larger library. It does not give the library a point of view.
The same fact often lives in four places
Most businesses did not deliberately create an architecture for organisational truth.
Information accumulated wherever the work happened.
A product goal lives in a strategy deck. A slightly different version appears on the task board. The current wording sits in a prompt. The exception agreed in a meeting survives in someone's notes. A new employee asks a manager, who explains the version they remember.
Humans cope by interpreting context.
They know that the wiki is old, the task was a one-off and the sales director's message is what people actually follow. A long-serving employee can often reconstruct the real rule from experience.
An AI agent sees several plausible instructions.
It may choose the most recent, the most detailed or the one that looks most relevant. None of those is the same as choosing what the organisation has deliberately decided.
This is why connecting AI to every information source can make inconsistency more visible without resolving it.
A wiki, task board, prompt and specification do different jobs
The answer is not to put everything into one enormous master document.
Each kind of information needs a home suited to its purpose.
A wiki explains what people may need to know
A wiki is useful for reference material: background, policies, guides, research, history and explanations.
It can be broad because its reader is looking for information. Not every page needs to guide an immediate action.
The weakness appears when the wiki contains several versions of a decision without a clear owner or update rhythm. Search becomes easy, but authority remains ambiguous.
A task board explains what is happening now
A task holds a piece of work moving through the organisation.
It may include a request, status, acceptance criteria, assignee, discussion and delivery plan. Most of that information is temporary. Once the work is complete, its status should not become the permanent description of the business.
A task can change the durable truth. It should not quietly become its only home.
A prompt tells an AI how to behave in a particular interaction
Prompts are useful instructions for a particular tool, model or task.
But a prompt library often duplicates business context across many places. Pricing rules appear in the proposal prompt, the sales assistant prompt and the approval prompt. When the rule changes, somebody has to find and update all three.
The prompt becomes a hidden policy document maintained by whoever happens to edit the AI.
A living specification records what the organisation currently means
A specification holds the durable truth about one important product, service, agent or repeated way of working.
It explains:
- why it exists;
- the outcome it is responsible for;
- the language it uses;
- the rules that must hold;
- where judgement is allowed;
- what remains undecided; and
- who owns the result.
The specification does not replace the other systems.
It tells people and agents which durable commitments should guide their work and where to find the supporting detail.
A knowledge base preserves information. A specification creates an owned basis for action.
One home per durable truth
A simple principle makes this much easier:
Every durable truth should have one authoritative home.
That does not mean the information can only be used in one place.
It means other places refer back to the source instead of quietly creating their own version.
Suppose the organisation decides that standard proposals must protect a particular margin.
That rule should have one owned home. The proposal agent can use it. The approval workflow can check it. The sales guide can explain it. The delivery team can rely on it.
But those tools should not each become separate authorities for what the margin is.
When the rule changes, the organisation changes it once, through the person who owns the commercial outcome. The connected work then uses the new decision.
This is not mainly a tidiness exercise.
It is how the business prevents yesterday's prompt from overruling today's strategy.
A specification must be owned by a person
An AI-generated summary of company knowledge is not automatically a specification.
It may be an excellent draft. It may find contradictions, identify repeated concepts and suggest a concise statement of the current approach.
Someone still needs to decide whether that statement is true.
If two policies conflict, the AI cannot determine which business trade-off the organisation intends to make. If an old task contains an exception, it cannot know whether that exception became a rule. If sales and delivery disagree, it cannot make the commercial conflict disappear by choosing the most confident paragraph.
A specification needs an accountable human who can:
- explain why it says what it says;
- resolve or name contradictions;
- decide what should change;
- distinguish exceptions from durable rules; and
- own the outcomes people and agents produce from it.
AI can help maintain the record. It should not become the unnamed authority behind it.
“We have not decided” is useful context
Knowledge systems often reward completeness.
An empty field looks like a problem. A missing answer invites someone, or an AI, to fill the gap.
But organisations have real open decisions.
Perhaps leadership has not decided whether an agent can approve a discount. Perhaps the new service has no settled cancellation policy. Perhaps the team is still testing which handoff should become standard.
The honest answer is not a plausible paragraph assembled from neighbouring documents.
It is: we have not decided, this person owns the question and the work must stop or route around it here.
Explicit uncertainty gives an agent a safe boundary.
False completeness gives it permission to invent the business.
Build a truth map before another AI integration
You do not need to reorganise every company document before using AI.
Start with one repeated piece of work and make a small truth map.
1. Name the work and its owner
Choose something concrete: preparing a proposal, onboarding a customer or approving a campaign.
Name the person who owns the outcome, not simply the person building the agent.
2. List the information the work needs
Include the goal, business rules, customer context, live operational data, templates, current work and genuine open decisions.
3. Identify the right home for each type
Ask:
- Is this durable organisational truth?
- Is it changing operational data?
- Is it temporary work status?
- Is it supporting knowledge or evidence?
- Is it an instruction specific to one tool?
The purpose is not to move everything. It is to know what each source can be trusted to answer.
4. Find the contradictions
Where do two sources describe the same decision differently?
Do not ask the agent to reconcile them silently. Bring the disagreement to the owner.
5. Create the smallest useful specification
Write the purpose, outcome, essential language, non-negotiable rules, room for judgement, open decisions and owner.
Keep links to the source systems for details that change quickly.
6. Connect it to one real piece of work
A specification proves its value when a person or agent uses it.
Run the next proposal, onboarding or campaign from it. Notice what still requires repeated explanation and what the organisation has not actually decided.
7. Return durable learning to its home
When the work changes what is true, update the specification deliberately.
Do not leave the new rule buried in the completed task, conversation or prompt.
What should the AI receive before it acts?
For a specific piece of work, an AI agent may need four layers of context:
- The specification: what the organisation intends, protects and has not decided.
- The live situation: the customer, request, record or event in front of it.
- Supporting knowledge: relevant policies, research, examples and historical evidence.
- The current task: what needs to be completed now and what done looks like.
These layers work together.
The customer record tells the agent what this customer said. The knowledge base explains the wider offer. The task says what to produce. The specification tells it which outcome and boundaries should govern the result.
Without the other layers, a specification can be too abstract.
Without the specification, the other layers are just a pile of context with no reliable way to resolve competing instructions.
How BusyWork Dispatch fits
BusyWork Dispatch is built around keeping durable context and current work connected without confusing them.
The Workbook holds the enduring truth about the important parts of the business. Dispatch holds the work moving now. Other systems can remain the source for customers, operations, code and supporting knowledge.
Before work begins, people and AI can start from the owned context instead of rebuilding it in every prompt. When completed work changes the business, the durable learning belongs back in the Workbook.
The practical goal is not to replace the company wiki.
It is to stop asking a wiki, a task board or a prompt collection to perform a job it was never designed to do.
Bottom line
A knowledge base tells AI what the organisation knows. A living specification tells people and AI what the organisation is trying to achieve, which rules should guide action, where judgement is allowed and who decides when the answer is unclear.
Connecting an agent to more information will not resolve conflicting versions of the business.
Give each durable truth one authoritative home. Keep current work, changing data and supporting knowledge in the systems suited to them. Then connect those sources through a specification a person genuinely owns.
That is how access to information becomes useful organisational context.
BusyWork Dispatch helps businesses create an owned Workbook for durable context and connect it to work being dispatched and shipped. See how BusyWork works or book a call with Ben.
