It is difficult not to be impressed the first time AI produces a serious-looking business document.
You ask for an operating model and receive 30 pages. It has principles, roles, risks, governance, a RACI matrix and a phased implementation plan.
It looks substantial.
The sheer weight of it feels like value. A piece of work that might once have taken weeks now exists before the meeting has finished.
Then you ask a simple question.
“What did we actually decide?”
The room goes quiet.
Everyone can see their concerns represented somewhere in the document. Nobody can explain which concern wins when two of them conflict.
The organisation has more documentation than it had an hour ago.
It may not have any more clarity.
Output is not understanding
Generative AI is extremely good at producing the shape of a complete answer.
It can give every section a heading. It can add the qualifications, edge cases and professional language that make a document feel considered.
This is useful when the thinking already exists and needs to be structured.
It is dangerous when the polished output conceals the absence of a decision.
A sales specification might say the business should “balance competitive pricing with sustainable delivery margins.”
Nobody disagrees with that sentence.
It also tells the salesperson or proposal agent almost nothing when a customer wants more scope for the same price.
Which margin must be protected? What can move? Who may approve an exception? Is the goal to win this customer, protect delivery capacity or preserve a standard commercial model?
The abstract sentence holds every preference at once. The decision remains unresolved.
AI can make an undecided organisation sound remarkably decisive.
Long documents can hide disagreement
People often assume more detail creates more precision.
Sometimes it does.
Sometimes the length is what allows everyone to avoid choosing.
One team’s priority appears in section three. Another team’s conflicting priority appears in section seven. A broad principle near the start promises both. A paragraph near the end says judgement should be applied according to the circumstances.
The document is comprehensive because it contains the whole disagreement.
This can be comfortable. Nobody’s view has been deleted. Nobody has had to say which outcome the organisation will own.
But when a person or AI agent reaches the real situation, the conflict returns.
The agent either asks a manager, makes up an answer or follows whichever sentence its reasoning happens to favour.
Length created the appearance of agreement without producing one.
A specification must remain human-ownable
An AI agent can read hundreds of pages without becoming tired.
That should not be our standard for a good specification.
The people responsible for the outcome need to understand it as a whole. They need to recognise when two rules contradict each other. They need to challenge an assumption, explain why an exception is justified and know what changed when a new decision is introduced.
If the operating model becomes too large for any person to comprehend, the organisation has handed understanding to the machine while leaving accountability with the human.
That is not a stable arrangement.
The accountable person may still approve the document, but approval is not ownership if they cannot explain the logic inside it.
If people cannot hold the specification in mind, they cannot meaningfully own it, even if an agent can read every word.
Human-ownable does not mean every specification must fit on one page.
It means its core logic is small and coherent enough for the people affected to reason about together.
Supporting evidence can be extensive. Customer records, legal sources, research, transcripts and technical documentation may sit behind it.
The specification is the concise statement of what the organisation currently believes and has decided because of that evidence.
The specification is not the whole filing cabinet
One reason specifications become bloated is that they are asked to contain every useful fact.
They become a wiki, task board, customer database, project history and operating manual at once.
That makes them difficult to maintain and impossible to trust.
A living specification should hold the durable truth about one important part of the business:
- why it exists;
- what outcome it owns;
- the language people and agents should use;
- the rules that must hold;
- where judgement is allowed;
- what has not been decided; and
- who owns the result.
Everything else should have an appropriate home.
The CRM holds the changing customer record. The work board holds current status. The technical system holds implementation detail. The research repository holds evidence. A temporary plan can disappear when the work is complete.
The specification can point to those sources without copying them.
This is not about having less information available.
It is about preventing the organisation’s current commitment from being buried inside all of the information that informed it.
A specification should carry the decision, not every document that led to it.
Short documents are earned
Writing a long document is often easier than writing a short one.
The first draft can include every observation, concern and possible exception. Nothing has to be discarded. The writer does not have to decide which part carries the most meaning.
Shortening it properly takes work.
You have to distinguish a cause from its symptoms. You have to turn five overlapping principles into one clear goal. You have to remove a rule that exists only because another rule was poorly expressed.
Most importantly, you have to resolve disagreement.
Suppose a proposal specification has 12 controls designed to prevent sales from committing delivery to unrealistic work.
It may be possible to keep refining those controls forever.
Or the organisation may decide that a proposal is not complete until delivery feasibility and a protected margin are both true. The agent can then find the most sensible route (adjust the scope, change the timing, use a standard product or escalate a commercial exception) while the two conditions remain clear.
The shorter version is not less rigorous.
It contains a better decision.
The document gets shorter as the organisation’s understanding gets better.
Lean does not mean vague
There is a bad version of simplicity.
“Use good judgement.”
“Act in the customer’s best interests.”
“Protect the brand.”
“Escalate where appropriate.”
These phrases are short. They are not necessarily clear.
A lean specification removes unnecessary material without removing the distinctions a person or agent needs to act.
It should still answer practical questions:
- What are we trying to achieve?
- What must be true when this work is done?
- What must never happen?
- Which trade-offs are available?
- Which decision belongs to the person or agent doing the work?
- What requires someone else’s authority?
- Which questions have not yet been answered?
Leanness comes from precision, not omission.
A clear sentence may take several conversations to earn.
AI should help us compress, not inflate
The temptation is to use AI’s capacity to make the specification larger.
It is often more valuable to use AI in the opposite direction.
An agent can help find:
- two rules that contradict each other;
- three sections that repeat the same principle;
- a long list of symptoms that may share one cause;
- language that sounds specific but does not change a decision;
- a term used differently by sales and delivery;
- a rule with no named owner; and
- a paragraph that mixes an agreed policy with an open question.
It can ask what evidence supports a rule. It can show which other parts of the business would be affected if the rule changed. It can propose a shorter version and explain what meaning might be lost.
That is useful assistance.
The agent still cannot decide which commercial promise the organisation is willing to make. It cannot manufacture agreement between teams. It cannot take accountability for the choice.
AI can lower the cost of organising the thinking.
People still have to do the deciding.
A practical compression pass
Before approving a process, product or agent specification, put it through this six-question pass.
1. Can the owner explain the whole thing without reading it aloud?
They do not need to recite every word. They should be able to explain the purpose, outcome, non-negotiables, available judgement and open decisions.
If they cannot, the document may be approved but not owned.
2. Which sentences change what someone would do?
If removing a paragraph would not change a decision, a boundary or a handoff, it may be background rather than specification.
Move useful context to its proper source and link to it.
3. Where have we described symptoms instead of the outcome?
Several controls may be compensating for one unclear goal. Ask what they are collectively trying to protect.
4. Which parts only sound agreed?
Look for words such as “appropriate,” “reasonable,” “strategic” and “where possible.” They may be valid, but ask what happens when two people interpret them differently.
5. What belongs somewhere else?
Remove live status, temporary plans, duplicated reference material and implementation detail. Keep one agreed home for each kind of information.
6. What have we not decided?
Turn disguised uncertainty into an explicit open decision with an owner. “We have not decided” is clearer than a page of language designed to hide the gap.
The goal is not the fewest possible words.
It is the smallest amount of shared guidance that preserves the decisions people and agents need.
Clarity has a cost, and a return
AI can generate the first draft almost for free.
The valuable work begins after that.
People have to compare the draft with reality. Teams have to surface competing expectations. Leaders have to decide which trade-offs the organisation will make. Someone has to own the result strongly enough to delete what does not belong.
That takes time.
But once the thinking is captured, one decision can guide many future actions. People stop repeating the same explanation. Agents stop returning the same ambiguity for review. New starters do not need to reconstruct the organisation from scattered documents and conversations.
The cost is paid once in clarification instead of repeatedly in coordination.
That is the return on a lean specification.
How this fits a spec-driven organisation
A spec-driven organisation is not the organisation with the largest library of AI-generated process documents.
It is an organisation whose important products, services and repeated ways of working have concise, owned and living definitions.
People understand the intent. Agents use it while doing the work. Open decisions remain visible. Completed work supplies evidence that can improve the specification without turning it into an archive of everything that happened.
The point is not to make the business readable only to AI.
It is to make the business clear enough that people and AI can work from the same understanding.
Bottom line
AI has made it almost free to produce more documentation. Clarity still requires people to decide what matters, resolve contradictions and express the result simply enough to be understood and owned.
A long document may contain more information. It does not automatically contain a better decision.
Use AI to organise, challenge and compress the thinking. Keep the specification focused on durable truth. Let supporting information remain in the systems where it belongs.
The goal is not less thinking.
It is more thinking per sentence.
BusyWork Dispatch uses a Workbook to preserve the concise, durable context behind a product or repeated piece of work, while Dispatch holds the work that is moving now. See how BusyWork works or book a call with Ben.
