Ask a business to explain how a piece of work happens and it will usually start listing steps.
First we receive the request. Then somebody checks the details. Then they open the system. Then they copy the information into a template. Then they send it to a manager. Then the manager sends it back.
By the time the process is documented, there are twenty steps, seven arrows and three diamonds asking whether an approval has been received.
That made sense when software had to be told exactly what to do next.
It makes less sense when the worker is a capable person, or an AI agent that can reason about the outcome and choose a route.
We learned to describe work as steps because our systems could not understand outcomes.
Now we have an opportunity to describe the work itself more honestly.
Following the process is not the same as doing good work
A step-by-step process offers a comforting promise.
If everyone follows the recipe, the output will be correct.
Sometimes that is true. Often the steps are only a rough attempt to protect the result.
Imagine a proposal process that says:
- Copy the customer details into the template.
- Select a service package.
- Add the standard introduction.
- Insert the pricing table.
- Send the draft to the sales manager.
- Make the requested changes.
- Export it as a PDF.
An agent can follow every step and still produce a poor proposal.
It might miss the customer’s actual priorities. It might include work delivery cannot fulfil. It might protect the template while damaging the margin. It might use the right sections and make the wrong commercial promise.
The process has validated movement.
It has not validated the result.
There's a name for this: Process Karaoke is what happens when an agent performs every visible step of an SOP but never checks whether the intended business result became true. It looks compliant and productive while quietly missing the point of the work. It's one of the six AI agent failure modes I keep seeing in production.
Start with the state you need to create
A better description begins at the end.
What must be true when this piece of work is finished?
For a proposal, the answer might be:
- it reflects what the customer said they need;
- the recommended scope can be delivered;
- the margin remains inside the agreed boundary;
- required legal and brand elements are present;
- any exception is explicit and owned;
- the customer can understand the value, price and next step; and
- the approved proposal and its source context are left in the right places.
That is a definition of done.
It does not dictate how every sentence is written or which system the agent opens first. It defines the state the work has to reach before it can move on.
Validate the quality of the result, not whether someone followed the recipe.
A workflow is a series of reliable handoffs
Most work crosses several people, teams and systems.
The sales conversation becomes a proposal. The proposal becomes a commitment. The commitment becomes delivery work. Delivery produces an outcome the customer has to accept.
The risk often sits at the handoff.
Sales believes it has transferred the customer context. Delivery receives a document with important assumptions missing. Marketing believes an asset is ready to publish. The channel owner discovers there is no approved call to action. A hiring manager says a role is ready to recruit. The recruiter finds that location, budget or decision authority has not been settled.
Adding more instructions inside each team does not necessarily fix this.
Instead, define the state that must arrive.
The work is ready to leave one stage when its definition of done is true. It is ready to enter the next stage when that stage’s entry requirements are satisfied.
Guarantee the handoff, not every movement that happens before it.
This gives each worker freedom inside the stage while making the promise between stages dependable.
What does an AI agent need instead of a long SOP?
For most knowledge work, five things create a useful specification.
1. A clear start
What event or request begins the work? What information must exist before the agent should pick it up?
“A salesperson asks for a proposal” is not enough if the agent also needs a customer transcript, an agreed service package and confirmed commercial owner.
Entry requirements stop incomplete work being pushed downstream and disguised as progress.
2. The goal
What is this work trying to achieve for the business and the person receiving it?
The goal should be more meaningful than “produce a document” or “complete the workflow.”
A proposal is not successful because a PDF exists. It is successful when a customer receives a clear, commercially viable offer the business can keep.
3. The definition of done
What must be true when the work is finished?
Describe the result, its quality and the state it needs to be in for the next person or agent. A good definition of done is specific enough to check without becoming a hidden twenty-step process.
4. The boundaries
Which rules must not be broken? Where can the agent use judgement? What requires an authorised decision?
Boundaries protect what matters while allowing the route to adapt to the situation.
5. What the work leaves behind
Which artefacts, records or updates must exist after completion?
This might include the approved proposal, the recorded exception, an updated customer record or a proposed change to the living specification.
Work should not disappear into a chat window. It should leave the organisation better prepared for what happens next.
These five map onto the six-question loop specification
Since writing this, I've formalised the same thinking into the six-question AI loop specification: what the loop is called and who owns it, when it starts, how it knows it is done, what is real when it is done, what it has available and the guardrails it works within.
The five things above are the middle of that specification. The six questions add the two anchors that make it an operating contract: a name for the work and one human owner accountable for the specification.
One clarification worth making: this is not an argument for deleting your SOPs. SOPs remain useful for people who need reproducible steps. Loop specifications describe delegated work for agents. An organisation may sensibly maintain both for the same business process. You can see the format applied across twelve worked examples.
Quality gates do not have to become approval gates
A quality gate checks whether the agreed conditions are true.
An approval gate asks a person to inspect the work and take responsibility for deciding whether it is good enough.
Those are not the same thing.
If a proposal must use the current terms, remain inside an agreed margin and include a defined set of customer information, an agent or system can often validate those conditions.
A person should become involved when the work reaches a genuine judgement or authority boundary, not because the process has reached step 12.
This is how a definition of done can reduce management load.
It lets the organisation agree what quality means once, then apply that agreement repeatedly.
Checklists protect judgement better than recipes
Some work is too complex to leave entirely open-ended.
Important conditions can be missed even by experienced people. The answer is not necessarily to control every action. It may be to protect the few things that must always be checked.
Atul Gawande’s The Checklist Manifesto explores how concise checklists help experts manage complexity across fields including medicine, aviation and construction.
The useful principle for agentic work is simple:
A checklist protects the few things that must be true. The agent is free to work out how to make them true.
A proposal checklist might confirm delivery feasibility, commercial boundaries, required terms and customer-specific evidence.
It does not need to dictate the order in which the agent assembles them.
When do exact steps still matter?
There are situations where sequence is part of correctness.
A safety procedure may require a control to be completed before equipment is touched. A legal or regulatory obligation may require a particular review and record. An external system may only accept actions in a fixed order. A financial transaction may require separation of duties.
In those cases, prescribe the sequence.
The important point is to know why the steps exist.
Do not preserve a 20-step process simply because that is how a person navigated the old software. Do preserve a step when changing the order would create legal, safety, security or operational risk.
Exact steps should be a deliberate control, not the default language for all work.
Use a spectrum, not an ideology
There are three useful levels of prescription.
Goal and outcome
Use this for open-ended work where capable judgement should choose the route.
Examples include developing a campaign concept, researching a customer problem or drafting an internal recommendation.
Goal, boundaries and quality checklist
Use this where the work is flexible but important conditions must not be missed.
Most proposals, hiring briefs, customer communications and operational reviews fit here.
Exact sequence
Use this where an external constraint makes order itself part of the correct result.
Many workflows will contain a mixture. An agent may choose how to research and draft a proposal, then follow an exact sequence when obtaining a legally required approval or updating an inflexible customer system.
The aim is not to eliminate steps.
It is to stop confusing the route with the result.
A simple way to rewrite an existing SOP
Choose one repeated piece of work and put the current procedure to one side.
Bring together the people who do the work and those who receive its output. Then ask:
What must be true before this begins?
List the minimum information, authority and conditions required for the work to start properly.
What outcome are we trying to create?
Write one sentence that explains why the work matters beyond completing the task.
What must be true when it is done?
Describe the quality of the result and the state required for a reliable handoff.
What must never be compromised?
Identify the commercial, legal, safety, brand or customer boundaries.
Which checks protect those conditions?
Keep the list short. If a check does not protect the outcome or a real boundary, ask whether it is necessary.
Where does judgement end?
Name the situations that require a person with particular authority or experience.
What should be left behind?
Define the artefacts and updates that make the completed work useful to the next person, agent or future run.
Only then return to the original procedure.
Keep the steps whose order genuinely matters. Remove the movements that merely describe how someone used to get there.
The result should be shorter, but that is not the main benefit.
It should be clearer about what good work actually is.
How this becomes a Loop
In BusyWork, a Loop is a repeated piece of work with a recognisable start, a definition of done, something it leaves behind, the resources it can use, the guardrails it must respect and an owner for the outcome.
That language is intentionally different from a traditional process map.
A Loop defines the commitment around the work. The person or AI agent carrying it out can reason about the best path. When the work reaches a real decision, it goes to the owner. When reality changes the durable understanding, that learning can return to the Workbook.
This is one of the practical building blocks of a spec-driven organisation.
The organisation becomes more dependable at the boundaries without becoming more rigid inside them.
Bottom line
Your AI agent usually needs a clear start, a meaningful goal, a definition of done, quality checks, boundaries and a reliable handoff, not a long list of movements copied from the old process.
Exact steps still matter when sequence itself protects safety, legality, security or operational correctness.
Everywhere else, define what must be true and let capable people and agents work out how to make it true.
That is how you protect quality without automating yesterday’s constraints.
BusyWork Dispatch helps organisations turn repeated work into clear, buildable Loops connected to durable business context. Start with one valuable piece of work, define its outcome and boundaries, and use the Workbook to keep what the organisation learns. See how BusyWork works or book a call with Ben.
