For the last 15 years, I have worked through one technology transformation after another.
The internet changed how businesses reached customers. Paper processes became digital. SaaS moved work into connected systems. Agile changed how products were developed.
Each wave felt completely different at the time.
But the deeper work was usually the same: the technology had made a new way of operating possible, and the organisation had to catch up.
AI is the latest version of that story. What makes this moment different is where the constraint now sits.
For most of the history of business technology, building the system was the expensive part. Strategy, coordination and design always mattered, but development cost so much that everything else looked secondary.
AI has changed that equation. The cost of producing software and carrying out knowledge work has fallen dramatically.
The bottleneck has moved.
It is no longer only whether the technology can do the work. It is whether the organisation can clearly define what the work should be.
That is why I believe businesses will become spec-driven organisations.
What is a spec-driven organisation?
A spec-driven organisation applies the idea of spec-driven development across the whole business.
Each important product, service, agent and repeated way of working has a living specification. It explains what that part of the business is for, how it should work, the rules it must follow and the decisions that have not yet been made.
People own the specifications and the outcomes. People, agents and systems use them to do the work.
When something falls outside the specification, it comes back to the owner rather than being guessed. When the work changes how the business operates, the specification changes with it.
The idea is simple:
Define what should be true. Use it to guide the work. Keep it current as reality changes.
The most practical unit of a spec-driven organisation is the AI loop specification. It turns an important repeated piece of work into a human-owned operating contract that people and agents can use, test and improve. Each important repeated piece of delegated work can have one, and master loops and sub-loops show how those units connect into an operating model.
AI is exposing disagreements that were already there
I have now sat with close to a hundred organisations to discuss AI transformation and building AI agents.
The technology is rarely the first real blocker.
The blocker appears when we ask people to describe the work clearly enough for an agent to do it.
Different people in the room disagree about what the workflow is. They disagree about what the workflow should be. They disagree about what a good result looks like and who owns it when something goes wrong.
AI did not create that ambiguity. It exposed it.
And the ambiguity is not always accidental.
Sometimes a workflow remains unclear because making it clear would force the organisation to decide who owns the result.
The calculator had the official rule. The approvals had the real one.
Quotations and proposals are a good example.
I worked with an organisation that had a pricing calculator shared across its sales team. Leadership had set the assumptions and numbers inside it. On paper, this was the agreed commercial process.
In practice, the sales team believed they could not win work using the calculator as designed. So they overrode it.
Different salespeople used different numbers. Custom approvals became normal. Scope was added to get deals over the line, and delivery inherited more work than it could reasonably complete for the price.
Every part of the organisation could still claim it was behaving rationally.
Sales was winning revenue. Leadership was protecting a valuable deal. Delivery was defending margin and feasibility. The calculator reflected the process leadership said it wanted.
The overall system was dysfunctional.
The sales process appeared to work because delivery absorbed the difference.
Now imagine giving that process to an AI agent.
Should it follow the calculator? Should it copy the exceptions that successful salespeople make? Can it reduce the price? Can it add scope? Who approves an exception? Is delivery required to honour anything a sales manager approves?
Those are not technical questions. They are commercial decisions.
Automating the existing process does not resolve the disagreement. It gives the disagreement speed and consistency.
The old process document sat beside the work
This is where many people become wary of the word “specification.”
They have lived through process-mapping exercises, governance programmes and documentation projects. A team spent weeks writing down how the business should operate. The documents were filed somewhere. Daily work continued much as it had before.
For the documented process to matter, someone had to know it existed, find it, read it, interpret it correctly and have enough time to follow it.
When the process did not fit reality, going around it was faster than changing it. The workaround became the real process and the document slowly became fiction.
People are right to remember that as bureaucracy.
A living specification is different because it participates in the work.
- Before the work, it gives the person or agent a warm start.
- During the work, it supplies the purpose, rules, language and boundaries.
- When a real decision appears, it identifies the person who owns it.
- After the work, it is updated if the work changed what is true.
The old process document asked people to follow it. A living specification gives people and agents something to work from.
What belongs in a living specification?
A living specification is closer to a concise, current brief than a procedure manual.
It should explain:
- what this part of the business is for;
- who it serves;
- what it is trying to achieve;
- the language people should use;
- the rules that must not be broken;
- where there is room for judgement;
- what has not yet been decided; and
- who owns the outcome.
It should not try to hold everything.
Today’s delivery status belongs with the work. Technical detail belongs with the technology. A temporary execution plan can disappear when the job is finished. Fast-moving customer and operational data belongs in the systems that produce it.
The specification holds the durable truth: the things a capable person or agent would be wrong not to know before beginning.
AI has made documentation cheap. It has not made clarity cheap.
One downside of generative AI is that it can produce extremely long reports.
The weight of the document feels like productivity. You asked for something and received 40 pages. It feels like you got a lot for your money.
But people cannot hold 40 pages of organisational instructions in their heads.
If the people responsible cannot understand the specification as a whole, they cannot recognise a contradiction, challenge a rule, explain an exception or remain meaningfully accountable for what agents do with it.
If people cannot hold the specification in mind, they cannot meaningfully own it, even if an agent can read every word.
A long specification often means the hard thinking has not happened yet. The AI included everything. Nobody felt confident enough to delete anything. Symptoms were listed instead of causes. Different opinions were all retained because the underlying disagreement was never resolved.
The document gets shorter as the organisation’s understanding gets better.
Specify the outcome, not every movement
Traditional automation had to be told exactly what to do next. That trained businesses to describe work as a sequence of steps.
Agentic work does not always need that level of prescription.
Instead, define:
- what must be true before the work starts;
- the goal it is working towards;
- what must be true when it is finished;
- the quality checks at the handoff; and
- the boundaries it cannot cross.
Then allow a capable person or agent to find the best path inside those boundaries.
Exact steps still matter where order itself is part of correctness: a safety control, legal requirement or inflexible external system. But they should be the special case, not the default language of work.
Atul Gawande’s The Checklist Manifesto shows how a small set of critical checks can help experts handle complexity without attempting to replace their expertise.
That same principle applies here.
Guarantee the handoff, not every movement that happens before it.
An agent should know when to act, and when to ask
A specification does not need to contain every answer.
It needs to be honest about which answers exist and which do not.
“We have not decided” is valid operating guidance for an agent.
The agent can work around the decision, stop at the boundary, present options or return it to the person who owns it. What it should never do is quietly invent a policy because the organisation left a gap.
This is how specifications create room for autonomy rather than remove it.
Guardrails are not the opposite of autonomy. They are what allow autonomy to scale.
People should own fewer, more important decisions
In many organisations, the manager is the real specification.
The rules, priorities and trade-offs live in their head. Everyone else has to bring decisions back because nobody can reliably predict what the manager will choose.
The manager believes the answer should have been obvious. The team is told to show more initiative. Then the next situation is similar but not identical, and the team asks again because the consequences of guessing wrong are real.
Sometimes the team is not lacking initiative. It has no reliable way to know whether the manager will choose option A or option B.
If the manager remains the undocumented specification, every person and agent has to keep asking them what to do.
This is also why adding AI can make managers busier.
Agents generate more drafts, more options and more work. If the decision logic is still trapped in one person’s head, all of that output flows back to the same bottleneck for review.
AI can multiply the work that was already not working.
In a spec-driven organisation, the manager’s role does not become something futuristic. It finally becomes the role it was supposed to be.
They spend less time firefighting and interpreting the same decisions. They spend more time developing people, improving how the team works, setting direction and making the smaller number of decisions that genuinely need their judgement.
The felt benefit is not only speed. It is calm and control.
Agents should learn faster than the organisation changes
A living specification should improve, but it should not change every time an agent notices a pattern.
There are three different speeds of learning:
- Signals are observed. Agents gather data, find patterns and surface insights.
- Experiments are tested. The business tries a change without declaring it the permanent rule.
- Specifications are changed deliberately. An accountable person decides the organisation has learned something durable.
Agents surface what the organisation might learn. Accountable people decide what the organisation has learned.
Otherwise, leaders risk vibe-managing the business: repeatedly changing how it works according to AI recommendations and short-term signals without anyone meaningfully owning the decisions or their consequences.
Agents can keep up with a specification that changes a thousand times a day. The people responsible for the outcome cannot.
Every change also carries a tax. It must be understood, communicated and absorbed across systems, teams and connected work. AI can reduce that change tax. It cannot eliminate it.
The ability to change faster does not remove the need to change deliberately.
Start small and whole
A business does not need to specify the entire organisation before beginning.
The specification does not even need to contain every answer.
Two things need to be true:
- The people doing the work and receiving its output agree on the part being introduced.
- The agent can distinguish what has been decided from what remains open.
Then choose the smallest useful part of the work that can run properly from beginning to end.
For a proposal agent, the first version might collect customer context, draft a proposal using standard pricing, flag anything outside agreed margins and send it to a human for approval.
It may not yet negotiate, approve exceptions or learn from delivery outcomes. That is fine. Its beginning, output and human handoff are clear.
Start small and whole, not broad and rough.
The bigger vision
The goal is not to remove people from the work.
It is to increase the reach of their judgement.
One clear decision, captured in the specification, can guide every future person, agent and system that encounters the same situation. People no longer have to explain the business from scratch or personally supervise every action.
That is how hundreds of people and thousands of agents can work in the same direction without requiring thousands of additional coordination decisions.
The amplification works both ways. A clear decision can guide an enormous amount of good work. An unclear decision can create chaos at the same scale.
This is why the specification matters.
How BusyWork Dispatch fits
BusyWork Dispatch is being built around this model.
The Workbook holds the durable truth about the important parts of the business. Dispatch holds the work that is moving now. People and AI can begin with what the organisation already knows, shape a clear request and send it to the BusyWork team to build.
When completed work changes what is true, that learning belongs back in the Workbook so the next request begins better informed.
The product is the practical starting point. Begin with one product, one service or one repeated piece of work. Define what should be true. Run the next real request from it. Then improve the specification from what happens.
Bottom line
A spec-driven organisation gives each important product, service and repeated way of working a living specification. People own what it means and the outcomes it produces. People, agents and systems work from it, return unanswered decisions to their owner and keep it current as reality changes.
The idea is not to document the whole business.
It is to make one part of the business clear enough that people and AI can work from the same understanding, then allow every piece of completed work to make that understanding better.
That is how the organisation moves faster without becoming harder to direct.
BusyWork Dispatch helps businesses turn AI ideas into working software, automations and agents without starting from scratch each time. See how BusyWork works or book a call with Ben.
