Salesforce AI governance has a weekly ceiling, and it is the approvers

Salesforce AI governance has a capacity limit, and it sits with the people who approve, not with the agent that drafts. In the constructed team below, three reviewers can approve about 32 AI-drafted changes a week. My position: once drafted volume runs above approval capacity for four straight weeks, review quality decays before any throughput dashboard shows it. Withdrawal conditions follow the arithmetic.

The capacity arithmetic for Salesforce AI governance

Every number in this section is constructed for illustration, not measured.

Take an admin team with three approvers, none reviewing full time. Each protects four hours a week, so the team has 3 x 240 = 720 review minutes a week.

Assume a median of 20 minutes to read one AI-drafted change: the diff, the sandbox result, the dependency picture. Assume a quarter of drafts bounce back once and cost another 10 minutes. The expected cost per change is 20 + 0.25 x 10 = 22.5 minutes.

Capacity is 720 / 22.5 = 32 approved changes a week. Re-check: 32 x 22.5 = 720.

That is effective capacity, not headcount. Three approvers are three effective heads only in a week with no leave and no incident.

The assumption the number is most sensitive to

Every figure from here on extends the constructed model.

The model is far more sensitive to review minutes per change than to the number of reviewers.

Drop the median to 15 minutes and the cost becomes 15 + 2.5 = 17.5, so capacity rises to 720 / 17.5, about 41. Raise it to 30 and the cost is 32.5, so capacity falls to about 22.

In the model, review minutes are the variable most likely to drift unannounced: under queue pressure the assumed 20 minutes could quietly become 8, throughput would look healthy, and Salesforce AI governance would have changed underneath the metric.

Where the claim fails

Two plausible changes break the 32.

The first is attrition. One reviewer leaves for a quarter: 480 / 22.5 = 21.3 changes a week. If a replacement ramps at half capacity for eight weeks, the team runs at 600 / 22.5 = 26.7.

The second is separation of duties. If whoever prompted the agent may not approve the result, and one reviewer is also the main prompt author, that reviewer is excluded from their own drafts. Capacity on those drafts drops to two reviewers while headcount still reads three.

If the agent drafts 45 changes a week against a capacity of 32, the queue grows by 13 a week: 52 waiting after four weeks. Either they age, or review minutes shrink until they fit.

How the tools place the approval seat

Every Salesforce AI governance design puts the human checkpoint somewhere, and that placement decides where the queue forms.

Gearset describes its agent this way, according to the vendor's page, read 1 Oct 2026: "Assign a ticket, and you get validated changes ready for your team to review." The Cam page, read the same day, also claims automated testing and an audit trail in the Gearset pipeline. Copado frames its standard this way, according to the vendor's Agentia page, read 1 Oct 2026: "Humans verify outcomes before any release reaches production."

FlowSprite AI separates the drafter from the system of record. The customer's own AI drafts a Salesforce metadata change; FlowSprite AI lands it as a GitHub pull request in the customer's private repo. The change is tested in a sandbox first, and a human approves before anything deploys. Production is never auto-touched, so the approval queue is the throttle by design. Approval workflows with separation of duties sit on the Enterprise tier at EUR 199 a month for 25 users; Professional at EUR 99 a month carries deploy preview and dependency analysis but not that workflow. A team can test whether those reviews take 20 minutes during the 14-day trial on FlowSprite AI, no card required.

FlowSprite AI does not make a reviewer faster; it makes the queue visible in a repo the customer owns. It fits a small admin team that brings its own AI. It does not fit an enterprise that wants the agent and release management from one supplier. According to the vendors' pages, read 1 Oct 2026, Copado sells agents across the delivery lifecycle in Agentia, while Flosum and AutoRABIT sell broader enterprise suites that pair DevOps with backup, security and governance. According to their pricing pages, read 1 Oct 2026, Gearset and Hutte offer 30-day trials, against FlowSprite AI's 14 days.

What would have to be observed to abandon it

The position is held while three signals stay inside their bands.

First, review minutes under load. If drafted volume runs above 32 for four weeks and median review minutes stay within 10% of their first-month value, review is not thinning, and the claim is withdrawn.

Second, queue age. If over-capacity drafting persists for four weeks while median queue age stays flat, saturation sits elsewhere, and the claim is withdrawn.

Third, defects after approval. If post-deploy fixes per approved change do not rise during a sustained over-capacity month, thin review is not decaying quality, and the claim's core fails.

Consequence for a planner this quarter

Salesforce AI governance starts with the agent's drafting budget, set from approval capacity. With the constructed numbers, that is a weekly ceiling of about 30 drafts, a buffer under 32, and about 20 in any week with a reviewer out.

Write the review-minute median into the change policy as a floor. A Salesforce AI governance policy without one measures throughput and calls it control.

Model separation of duties as capacity, not paperwork: count which approvers are excluded from which drafts before counting heads.

Held as of October 2026: in the constructed model, three part-time reviewers absorb about 32 AI-drafted changes a week, and drafting above that for four weeks degrades approval before throughput metrics show it. I will revisit it on 15 January 2027 against the three signals above, and say whether it is held or withdrawn.