The next task must exist before the current one closes

Today we received another completed research package from one of the people helping us test a distributed commercial operating model.

The package contained two potential law firms, publicly verified senior recipients, evidence of international work, commercial hypotheses, unknowns, risks and draft messages. No external contact had been made.

At first glance, the next step looked simple: review the work, approve or reject the candidates and send feedback.

But that would have left a familiar and surprisingly expensive gap.

The research would have been finished. The researcher would have been waiting. The next activity would have existed in our heads, but not yet in the system.

So we changed the transition.

We closed the research assignment and registered the next executable assignment in the same decision: two approved role-clarification messages, a named owner, a specific date, clear reporting requirements and explicit stop rules.

The most important learning was not about the firms.

It was about the space between finished work and authorised work.

Most delays hide between tasks

When organisations try to improve productivity, they often focus on how quickly each task is performed.

Research should be faster. Analysis should be faster. Writing should be faster. AI should reduce the time required for all three.

That matters, but it is not always where the largest delay sits.

A task can be completed in an hour and then wait two days for someone to review it. A decision can be made in ten minutes and then wait another day before it becomes a concrete assignment. The next person may understand the general direction but still lack the authority, date or boundaries required to act.

None of this waiting appears as difficult work.

It appears as silence between states.

In a human–AI operating model, this problem can become worse. Production capacity expands quickly. Researchers and AI workspaces can investigate, compare and draft more than the central decision-maker can absorb. If the architecture does not convert completed output into authorised next work at the same speed, production does not become the advantage.

The decision queue becomes the bottleneck.

Architecture and production are different jobs

Our emerging model has two distinct roles.

Production gathers evidence, identifies relevant people, develops hypotheses, marks what remains unknown and prepares possible action.

Architecture decides what the work means, what may happen next and what must not happen without another decision.

The distinction is useful because it lets colleagues and their AI systems work with real autonomy inside a bounded assignment. They do not need approval for every search, comparison or draft. But they also do not gradually acquire authority to redefine the market, change the offer, expand the research or contact people on their own initiative.

The architecture does not need to repeat the production work.

It needs to answer a smaller set of consequential questions:

• Is the evidence sufficient for the next decision?
• Is the candidate relevant enough to test?
• Is the proposed recipient sufficiently decision-near?
• What is the smallest useful external action?
• What result must return before further action?
• Where must the work stop?

That is a much more efficient use of human attention than supervising every production step.

But it only works when the answer becomes an executable task immediately.

A decision without a registered next action is still unfinished architecture.

Stop researching when the next answer belongs to the market

The package also exposed another source of waste: trying to research information that is not publicly available.

We could verify senior roles, current titles, contact channels and evidence of cross-border work. We could form a reasonable hypothesis about decision proximity.

We could not verify who actually owns CRM, commercial tools, client development or pipeline processes inside each firm.

The natural response is to keep searching.

Sometimes that is necessary. More often, it is an attempt to create certainty from sources that cannot provide it.

Our new stopping rule is becoming clearer:

When a current, decision-near recipient is sufficiently verified, and the remaining question can only be answered by the organisation, stop the research and use the first contact to clarify the role.

This does not lower the standard.

It changes the purpose of the first contact.

The message is not yet a full sales approach. It does not assume a need, claim a gap or force a meeting. It asks a precise organisational question: Is this your responsibility, or should we speak with someone else?

A reply, redirection or even a bounce gives us information that another hour of public research might not produce.

The market becomes part of the research process.

One contact can test several assumptions

We deliberately approved two different organisational cases.

One is founder-led, with a managing partner who may have a relatively short path between strategic relevance and decision.

The other is part of a large international network, where local, regional and global responsibilities may be divided.

The messages are similar, but the comparison is useful.

We can observe:

• whether a decision-near leader responds to a narrow role question;
• whether the person owns the area or redirects us;
• how long the internal route appears to be;
• whether a founder-led structure produces a shorter decision path;
• whether network complexity creates friction before a commercial conversation can begin; and
• how much pre-contact research was actually necessary.

This is more valuable than simply recording whether an email received a reply.

The action is small, but it produces evidence about our operating model.

A handover should create a state transition

The transfer package is not an archive of completed work. It is an input to a decision.

For the process to move without unnecessary waiting, every completed package should trigger one of three outcomes:

• GO: approve a specific next action;
• HOLD: preserve the candidate but define what must become true before it moves;
• STOP: close the candidate and record why.

If the outcome is GO, the same review should create the next assignment with:

• a unique identity;
• an owner;
• an execution date or trigger;
• the exact approved action;
• the evidence or wording to use;
• the result that must be reported; and
• explicit stop rules.

The production assignment can then close cleanly. The next assignment already exists. There is no informal gap where someone waits, guesses or continues working beyond the authorised boundary.

This seems like administration.

It is actually throughput architecture.

Stop rules increase speed

Clear constraints can look like a brake on initiative.

In this process, they do the opposite.

The colleague does not need to wonder whether to do more research, find another recipient, send a follow-up or expand the sequence. The AI does not need to infer authority from momentum. Both can complete the approved action quickly and return the result.

The architecture then makes the next decision with new evidence.

This creates a tight loop:

research → decision → registered action → market response → learning → next decision

The loop remains controlled without becoming slow.

That balance matters. If every small action requires a meeting, the system cannot scale. If every production unit is allowed to interpret the plan freely, the shared commercial memory fragments.

The answer is not more supervision.

It is better task boundaries and faster state transitions.

What changed

We stopped treating a completed transfer package as the end of an assignment followed by a separate planning moment.

The architecture review now closes the completed work and creates the next authorised work in the same operation.

We also stopped assuming that research should continue until every internal role is known. When the remaining question belongs inside the target organisation, a precise role-clarification message becomes the next research instrument.

What I think matters

The speed of an AI-enabled organisation will increasingly depend on how quickly it converts completed work into the next authorised action.

Production capacity is becoming abundant. Coherent judgment, clear authority and fast queue management are not.

The organisation that learns to move work across those boundaries without losing control will gain more than task efficiency. It will learn faster from the market.

What remains uncertain

We still need enough real cases to learn how often decision-near recipients respond, redirect or remain silent.

We do not yet know whether founder-led firms consistently produce shorter routes, whether international networks create meaningful delay or how much the wording of the role question affects response quality.

Those questions should not be answered by assumption.

They are now attached to observable actions, registered outcomes and a shared commercial memory.

That is the point.

The next learning should already have somewhere to land.