The constraint is the coordination layer

An integration point can become a constraint

In The Goal, Goldratt uses factory machines and process steps to show how a constraint limits completed output.

In knowledge work, the constraint may be a person, a decision, or a review step. A chief marketing officer’s brief, a senior analyst’s report, or a success lead’s renewal view is a candidate only when its recorded delay holds up downstream work.

Block proposes replacing part of its former management layer with an intelligence layer that moves context across teams. The memo does not establish that this layer is the company’s constraint or that the reorganization has improved end-to-end output.

A recorded integration point may receive several inputs and delay several downstream tasks. Treat it as the constraint only when queue time and “waiting on” records show that its output limits end-to-end work.

Integration points can delay several downstream tasks

Goldratt's classic claim is arithmetic: throughput equals bottleneck throughput. Add capacity to the bottleneck and the whole system goes faster. That part is taught in operations programs.

In an interconnected system, the constraint is the point limiting end-to-end output. A chief marketing officer can become that point when brand, market, customer, channel, and team decisions all wait for one person. Improving the integration point can increase throughput until another step becomes the limit. Measure the end-to-end result instead of assuming a multiplier.

Block proposes a system that moves context between teams continuously rather than waiting for periodic status meetings. That is a design claim, not an observed company result. Done when: Block publishes a baseline, end-to-end measure, review period, and result that can distinguish the intelligence layer from other changes.

Why AI is shaped for integration

Large language models can summarize and combine approved sources. Treat that capability as a fit only when the source set, output, reviewer, and acceptance check are explicit.

Source completeness affects synthesis quality. Test the selected source set against an accepted baseline work product before the Fidelic agent handles live constraint work. Done when: the review records missing sources, accepted output, correction burden, and change in downstream wait time.

Treat the integration point as a senior role. Record its approved Slack channels and tools, written authority limits, review owner, acceptance check, and auditable action record. A vendor demo uses generic instructions; a hired role uses the team’s approved source record, brand guide, third-quarter objectives and key results, and campaign log.

Block describes internal and external information feeding a system that combines them into financial products. A role-level test is smaller: one delayed work product, approved sources, one reviewer, and a measured downstream result.

The five focusing steps, re-read

Goldratt's five steps are usually presented as a linear playbook. They are better understood as a discipline you cycle through indefinitely. AI labor lives mostly in step four, but the first three steps determine whether step four works.

1. identify the constraint

The constraint is the work product whose delay holds up the most downstream work. Find it with recorded “waiting on” notes, queue time, and the number of owners blocked by the same output. Manager demand alone is not evidence of a constraint.

2. exploit the constraint

Before adding capacity, record the time the role owner spends on source collection, routine triage, and first-pass synthesis. Move only a published work product with approved sources, destination, reviewer, and authority limit. Done when: the repeated preparation moves from the role owner to the Fidelic agent while unfamiliar judgment and approval remain human.

3. subordinate the rest of the system

Set upstream intake to the recorded capacity of the constraint. Done when: the queue limit, intake owner, and rule for holding or releasing work are written, and downstream wait time is measured after the change.

4. elevate the constraint

Add a Fidelic agent only when the constraint has a scheduled or monitored work product, approved source access, written authority limits, visible work, and an auditable action record. Compare end-to-end throughput with the baseline instead of inferring a multiplier from local activity.

5. repeat

Repeat the diagnosis when downstream wait time stops falling or a different queue becomes the longest. Block’s proposal does not establish whether its intelligence layer is a continuous diagnostic or a one-time reorganization. Done when: each new constraint record names the delayed work product, baseline wait time, accountable owner, and review date.

Four conditions before hiring at the constraint

Four conditions are required before a Fidelic agent works at the constraint: approved source access, written authority limits, a visible review surface, and an auditable action record. If any condition is absent, keep the output in draft or review mode until the accountable owner records the missing condition.

The approved source record includes every source required for the named work product and excludes private material outside the role’s authority. Done when: the reviewer can trace each claim in the work product to an approved source and identify every missing source.

• A written authority record maps each action to autonomous, review-required, escalate, or refuse. Done when: every action has one tier, every review or escalation names its owner, and every refusal names the rule that blocks the work.

The Fidelic agent places the work in the approved destination with evidence and review status attached. Done when: the assigned reviewer records a correction or approval in that destination.

The action record identifies the approved sources, work product, reviewer, corrections, and final status. Done when: a reviewer can reconstruct the result and schedule a recheck without relying on memory.

Do not assign the work to a Fidelic agent until all four conditions are recorded. Keep the accountable person at the constraint, remove unrelated triage where possible, and make the source and review records inspectable before testing added capacity.

Keep unfamiliar judgment and representation human

Three constraint types should not receive autonomous agent authority: judgment under unfamiliar uncertainty, work that requires personal accountability, and work whose required context cannot be made inspectable.

Judgment under unfamiliar uncertainty includes acquisitions, executive hiring, and board strategy. Available records may not contain a stable precedent or acceptance rule. Keep that decision with the accountable person.

• Bottlenecks of relational accountability. The customer call, the partner negotiation, the regulatory sign-off, the meeting where someone has to take the responsibility. The work is representation, not synthesis.
• Bottlenecks where the setup cannot be built. The role's context lives in private conversations, off-platform tooling, or the synthesizer's head. The agent cannot read what is not legible. Go back to step two.

A composite scene

A thirty-person SaaS company. The CMO has been the bottleneck for nine months: every brief ten days late, every campaign starting on the back foot. The team hires SCOUT in October. The setup includes the brand guide, the customer-interview record, the objectives and key results document, the campaign log, the analytics dashboards, the Slack channels where the marketing team coordinates. The December target is a sourced brief every Monday morning. Done when: the brief links approved sources, names unresolved questions, and reaches the chief marketing officer before the scheduled review.

By February, the constraint has moved. Revenue operations is now the integration point. The team has hired ALEK to do the chief-of-staff synthesis at the executive layer. By June, the constraint is in product strategy: somewhere else again.

Each later hire requires a new baseline, accountable owner, and measured downstream result. Done when: each expansion decision names the delayed work product, approved sources, review owner, and change in downstream wait time.