Optmyzr Change Requests: Every Guardrail Has an Off Switch

Optmyzr change requests are the most carefully built human gate we have found in a PPC platform — and the vendor wrote the specification for them itself, in public, three weeks before our first look at the documentation. Which makes it a rare thing: a competitor claim we can check line by line against the vendor’s own manual.

Start with the article. On 15 September, MarTech published “3 ways to make AI safer in a live ad account.” It carries Optmyzr’s byline, and its footer carries the disclosure:

That tells you what you are reading: not a columnist’s verdict on the category, but Optmyzr’s own account of what safety in an ad account requires — published where a prospect can read it, and therefore checkable.

Three questions, and none of them is “do you trust the AI?”

The article is short, and it is better than most of what vendors publish on this subject. It frames agent safety the way you would frame a new hire, and reduces it to three questions:

• What can it see? An agent working off a thin data layer will give you a confident answer built on a third of your account.
• What is it allowed to do? Not what you told it to do. What it is structurally prevented from doing, no matter what anyone tells it.
• Who signs off? Not “we review the change history afterward.” Who has to say yes before?

Then it does something almost nobody does: it writes down the protocol it is asking you to expect. Every write becomes a draft change request. Policies evaluate it and carry their verdict. A human opens the review and sees the reasoning, the policy warnings, the timeline and the list of people eligible to approve it. You preview the exact deterministic change. You confirm. Only then does anything move.

And it closes with the sentence that turns the whole thing into a procurement document:

Five things. Not a vibe. A field list. Optmyzr wrote it down, and a field list is a thing you can check.

Two weeks later, they documented it

Most vendors who publish a framework like this stop there. Optmyzr didn’t, and that deserves stating plainly before anything else in this post. Between 24 and 29 September the company built out the paper trail:

• 24 September — the Sidekick FAQ is updated to answer “Can Sidekick make changes to my account without my approval?”
• 25 September — a Help Center article goes up: Change Requests — User Guide, complete with a policy-settings walkthrough and a sixty-row entity/action/platform support matrix.
• 28 September — the changelog announces Change Requests raised from Sidekick in Optmyzr or in Slack, with a /optmyzr command.
• 29 September — the MCP FAQ is updated with the same answer for the Claude and ChatGPT path.

That is a vendor documenting the governance layer of an AI write path inside a fortnight. Compare it to the rest of the field, where a write-side claim has sat off its own product page for three weeks, or where an audit-trail claim lives on a homepage and names nobody on the page engineers actually read. Optmyzr shipped the guide. Credit first, because it is structural:

It also handles the edge cases properly. A request that pairs a platform, entity and action the product doesn’t support is refused when it is created, not silently dropped later. Removal actions are flagged separately, on the vendor’s own reasoning that they can’t be rolled back. And the gate holds across channels: the guide’s own summary is that the Change Requests page shows every change request raised on the account, whoever raised it.

So the question is not whether Optmyzr has a gate. It does, and it is better than most. It is what happens to the other two layers of that framework when you read them against the documentation rather than the article.

The gate is not a setting. The guardrail is.

Question two in the article is the bracing one: what is the agent structurally prevented from doing, no matter what anyone tells it. The article’s own examples are concrete — no bid increase above 10% in a single move, no budget change beyond a set threshold — and it is emphatic that this layer is not a preference:

The documentation agrees about the architecture. The policy guardrails live in an organisation-wide policy sheet, grouped by category — Bids, Keywords, Account-wide — and every change request is checked against them before it can be applied. A blocked row shows up blocked, with the reason attached.

Then, in the same paragraph, the guide describes the lifecycle of those guardrails:

Any of them. On or off. And the changelog makes the point concrete rather than theoretical: on 25 September, Optmyzr shipped an update titled “Choose whether or not to apply budget limits,” which documents the article’s own worked example arriving with an off position — Select No limits to let budgets change by any amount needed to reach the target.

Read those two documents together and you have the cleanest statement of the category’s current condition we have seen from a primary source. A structural prevention and a switch are not the same object. A prevention holds under every path into the account, exactly as the article says. A switch holds only while nobody flips it. Optmyzr publishes both, in its own words, on its own domain, two weeks apart.

The record’s field list lives in an FAQ. Its lifetime lives nowhere.

The article’s audit-trail section is the part this pipeline has been circling for months, so it is worth being exact about what Optmyzr’s documentation claims and what it doesn’t.

What it claims is genuinely good. The Sidekick FAQ answers the approval question and then lands the field list, using precisely the nouns the article promised:

What was requested, who approved it, and when. That is three of the article’s five. The guide adds the other two — the rationale, via a Why? link next to each change, and the policy verdicts, via a summary line reading, in the vendor’s own example, All policy checks passed, no violations. Credit the whole field list. It is coherent, and it is documented.

Now the part that isn’t. We ran a character-exact scan across the Change Requests — User Guide — the one document that tells you where the record lives and how to work it, at 14,299 characters of stripped text — and across the MCP FAQ that answers the same approval question at 5,867 characters:

Here is the precision that keeps this honest: the FAQ does claim who approved it, and when, and we are not going to argue with a vendor’s own published claim. The finding is where the claim lives. The field list is in a support answer. The specification — the guide that documents the object, its tabs, its actions and its permission model — does not restate it, and attaches no lifetime to it.

Compounding it: Optmyzr does publish a retention number. Just not for this object. Its MCP page answers what does Optmyzr log when I use the MCP with Logs are retained indefinitely, describing the queries made and related metadata — the query log, not the change record. Its Slack page says personal data is kept for as long as needed to provide the service. Two published lifetime answers, both for different objects, neither one for the record the article told you to rely on. That is the same shape of gap we found in four vendors’ retention pages in September: the number exists, it is just attached to something else.

Filed under Utilities

There is one more signal, and it is cheap to check. Where does a vendor file the record? The guide’s answer is the first instruction on the page: Go to Utilities → Change Requests. The policy sheet sits under Change Review → Policy settings. The guardrails are configured on an organisation-wide policy sheet, alongside billing and team seats.

And across the surfaces a buyer actually reads — the homepage at 26,456 characters, the pricing page at 71,763, the MCP page, the AI page, the LLM-facing product definition page, the Slack page, the Claude connector page — the vocabulary scan returns zero for change request, change review, approver, audit trail, audit log and jurisdiction. Optmyzr’s homepage does carry a trust row, twice: No changes without your approval, ever. Its only “compliance” mentions are data encryption and security compliance — vendor security, not your regulatory position. Its fourteen mentions of “audit” on the pricing page are all account and feed audits.

None of this is a product failing. It is a positioning choice, and it is the whole finding: the article called the record a credibility argument for agencies. The product files it under Utilities.

What a regulator’s version of question three asks for

“Who signs off?” is the right question, and Optmyzr asks it well. A regulator’s version of the same question has five parts, and it is worth walking them because the gap is not in the gate — it is in the record of the gate:

• Who signed off, as a named person. Optmyzr documents a person on your team and the Apply permission that person needs. A role is how you control access. A name is how you answer a question about one decision.
• When, as a field. Claimed in the FAQ; not specified in the guide.
• For which client engagement. The record is scoped per account. The guide says nothing about clients, and for a multi-client agency that is the axis a compliance officer actually asks about.
• Under which jurisdiction. Nothing documents a tag, and nothing in the product’s vocabulary references financial regulation at all.
• For how long, and exportable on demand. Neither is documented, and the one lifetime Optmyzr publishes governs query telemetry.

Two of those five are answered thoroughly. Two are unanswered. One is answered in an FAQ and unspecified in the manual. That is a fair read of the documentation as it stands today, and it is a better read than “they have no audit trail” would be — which is exactly why the count matters.

Where Ott sits, and what we won’t claim

Ott’s Activity Log records the operation rather than the approval dialog: the actor, the value before and after, the timestamp, and the linked object — stored independently of the platform that accepted the change, so the record survives the account restriction that wipes the platform’s own history.

Two things we have to say against ourselves before anyone else does.

• Optmyzr covers more platforms than we do. Their Change Requests matrix spans six — Google, Amazon, Microsoft, Yahoo, Meta, LinkedIn. Ott’s Activity Log covers Meta today, with Google Ads on the roadmap. If your problem is Google Ads and you want a gate on it this quarter, Optmyzr is the shipped answer and we are not.
• We publish no retention number either. That is the same gap, and it is not closed by us having a different product. What we will say is where ours differs in kind: our record is written to a store that is not the ad platform’s, and you reach it outside the platform rather than through a screen inside it.

On pricing, the same discipline applies. Ott is priced by brands and seats rather than by ad accounts or data sources — the live page says no charges per ad account, data source, or Telegram join — and the parts that page states less loudly are that pricing scales at 1% of ad spend above included amounts, and that Telegram conversion tracking is not included on Starter.

Five questions for any vendor selling you an AI change gate

Whether you evaluate Optmyzr, Ott or anything else, the article got the right three questions. Here is what we would add, and every one of them is answerable from a vendor’s own documentation this afternoon:

The half the article didn’t write

Optmyzr’s framework is good enough that we published a post measuring a competitor against it, which is a strange compliment. It got the hard part right. The gate is real, it is mandatory, it covers six platforms, and it exists in a category where most vendors are still publishing a preview modal and calling it a control.

But the article’s own closing argument was that the review queue turned out to be the best account documentation we’ve ever had, and that for an agency this is a credibility argument as much as a safety one. Credibility is the one claim in that framework that does not resolve on the vendor’s side of the fence. It resolves when your client, or their regulator, asks about one decision months later and you can hand over a row that names a person, a time, an engagement, a jurisdiction and a lifetime.

Optmyzr documented the gate and left the record’s lifetime unpublished — and then wrote you can turn any of them on or off in the guide for the layer above it. Both of those are fixable. One of them is a paragraph.