Two operators, one execution owner, fourteen days, no unlogged edits: that is the constraint for this Meta agency takeover. Daily management means daily observation and justified execution, not a daily obligation to change settings. An agency does not earn its fee by finding something to toggle before lunch.
One person on your side retains write access. The agency names exactly one human who executes changes. Retaining access is not permission to run a parallel optimization program: every intervention goes through the same shared change queue, with an owner, a hypothesis, and a rollback condition.
The reason is less theoretical than it sounds. A DTC toy seller spending about $500 a day at 1.7–1.9 ROAS described handing its account to a Meta agency, then finding more than 1,000 changes nineteen days later. The log included 20 campaigns, 50 ad sets, and 102 ads built during that window. Of those ads, 74 were created and switched off again, with a median lifetime around 31 hours. Two operators occasionally undid each other: one enabled a mature ad set at 11 p.m.; another disabled it at 1:23 a.m.
That is a self-report, not proof that every edit was wrong. A change count alone cannot tell me whether fewer changes would have produced better results. What it does expose is a governance problem: the account became the place where operators argued with each other.
The first takeover deliverable is a change queue, not a rebuild. Here is the fourteen-day plan, including what each stage costs, what it buys, and what gets cut.
1. Days 1–2: Establish decision rights and record the baseline
Before anyone touches delivery, I want three things from the incoming agency:
- One named execution owner. Not a rotating bench. Ask who runs the account on an ordinary Tuesday and who resolves conflicting instructions.
- An accessible action log. It must preserve the reasoning behind interventions and remain available after the relationship ends.
- A weekly explanation of decisions. Ask for a redacted sample report connecting spend, conversions, conversion value, and cost per result to actual business decisions.
Opascope’s hiring checklist makes a useful distinction: a diagnostic agency names the change, its mechanism, and the expected metric movement. A rehearsed pitch skips those details. A sample report is revealing because it shows the finished output, not the sales deck; PixelRush’s hiring checklist makes the same point.
Then build the queue. A shared document is enough. Every proposed intervention gets one row with five minimum fields:
| Field | What belongs in it |
|---|---|
| Owner | One name accountable for the intervention |
| Hypothesis | What should happen and why, in one plain-English sentence |
| Affected assets | Campaign, ad set, and ad IDs—not “the prospecting stuff” |
| Observation window | How long the intervention runs before a decision |
| Rollback condition | The trigger for reversal and the prior state to restore |
Add approval decisions, execution timestamps, and status alongside those fields. The agency’s named owner executes approved changes. Your named approver controls exceptions to the agreed budget limits. If the two sides disagree, they settle it in the queue before either touches the account.
Meta’s Activity History records who changed what and when. It does not record why. The queue is the durable why layer, not another dashboard someone has to admire.
Before the first intervention, record the baseline:
- Active tests and their hypotheses, including anything abandoned mid-flight.
- Conversion definitions and discrepancies between Meta and your CRM.
- Each ad set’s attribution setting, so comparisons use the same basis. Meta’s attribution-setting documentation belongs beside that record.
- Current budgets, hard daily and weekly limits, and who may approve an increase.
- The commercial KPI that decides whether the account is working.
That last item matters more than the prettiness of the queue. One commenter in the toy seller’s thread recommends new-customer CAC from Shopify rather than Meta-reported ROAS. Whatever number fits your business, agree on it before execution and put it in the contract. The platform should not be the only party grading its homework.
Cost: two of fourteen days, plus the discomfort of paying a team to observe before intervening. Buys: a baseline and decision rights you can use later. Cut: the ceremonial day-one rebuild. A genuinely broken account still gets fixed, but its fix gets a queue row first.
2. Days 3–4: Separate emergency fixes from experiments
The queue needs two lanes because a broken checkout destination and a fresh creative idea do not deserve the same clock.
The emergency lane handles faults that cannot wait:
- Broken landing pages or destinations.
- Conversion events firing incorrectly or not at all.
- Unauthorized spend.
- Policy violations that put the account at risk.
Those fixes do not wait for a normal observation window. The execution owner logs the fault, the intervention, and the exact time it lands. Mark the affected data so nobody later mistakes a tracking failure for a sudden collapse in demand.
The experiment lane handles everything else: creative concepts, audience changes, bidding strategy, and budget allocation. Each intervention gets an observation window before it launches, not a deadline invented after someone dislikes the first afternoon’s numbers.
The distinction is not “never touch the account.” The Digital Ring’s learning-phase write-up puts the useful line between unnecessary disruption and changes justified by broken tracking, the wrong optimization goal, or performance evidence. That is the standard I want the queue to enforce.
For converting ad sets, practitioners in the toy seller’s thread recommend waiting at least a week before a kill-or-scale decision, with an exception for spending multiples of target CPA without results. Treat that as a planning rule, not permission to ignore your agreed spend limits. When average order value varies, a rolling seven-day view also gives a less twitchy basis for judgment than yesterday’s ROAS.
The other boundary is significant edits. Meta’s documentation on significant changes identifies changes that can send an ad set back into learning, including targeting, creative, optimization events, bid strategy, and adding ads. Pausing an ad set for seven or more days also matters. Budget changes depend on magnitude; not every adjustment carries the same disruption.
Allow one significant intervention per comparison window. If the agency wants to change creative and targeting independently, sequence them. Otherwise, even if performance improves, the queue cannot tell you which change deserved the credit.
A necessary combination is different. If a creative concept only makes sense with a matching audience, define it as one combined intervention. You can evaluate the combination; you cannot pretend you isolated its parts.
Cost: two days to classify pending work and agree on observation windows. Buys: fast repairs without turning every optimization idea into an emergency. Cut: daily kill-and-scale decisions made because somebody opened Ads Manager and felt obliged to contribute.
3. Days 5–10: Execute one queue, not two competing strategies
This is where the constraint starts earning its keep. Two people can both be trying to help and still produce an account neither can explain.
Only the named execution owner applies changes. Your retained access is a safeguard, not a second steering wheel. If your in-house operator wants to pause an ad set the agency just scaled, the pause button is not the argument. The proposal goes into the queue, the disagreement gets resolved, and the owner records the decision before execution.
The same rule applies to budget. Changes above the agreed ceiling need approval from your side before they land. The queue should already contain:
- The daily cap.
- The weekly cap.
- The person authorized to raise either limit.
A ceiling is not distrust. It prevents two operators from assuming the other is watching spend.
Dependencies also need explicit handling. A combined intervention gets one row, one hypothesis, and one observation window. If a new creative concept requires a matching audience change, record both asset groups and the execution order. If the creative launches Tuesday but the audience change lands Thursday, the observation clock for the combination starts Thursday. Tuesday was not the start of the complete test.
That does not license bundling unrelated ideas into a row called “account optimization.” A row should describe a decision you can evaluate, not a folder for everything the agency wanted to do this week.
During these six days, daily management means checking delivery, spend, tracking, and the progress of active hypotheses. It can end with no change. That is a legitimate operational decision when the account is functioning and the evidence does not justify intervention.
I also want one honest sentence in every takeover: not every test will produce a clean read inside fourteen days. Some will finish. Others will still need observation. The queue preserves an interpretable comparison where budget and data allow; it does not manufacture certainty on a contract deadline.
Before each intervention, record the prior state. While it runs, resist adding another significant edit to the same comparison. If the evidence is still thin at review, “observation window extended” is a better outcome than a winner declared for presentation purposes.
Cost: six days of single-owner execution and patience with unfinished tests. Buys: changes that do not cancel each other out and comparisons that remain readable. Cut: parallel editing, undocumented budget increases, and the promise that every hypothesis will resolve in a fortnight.
4. Days 11–14: Review the money before the change count
The review starts with a lag check, not a victory slide. Purchases appearing in today’s reporting may relate to earlier activity. Reporting and attribution changes can also alter the numbers without the underlying delivery problem an agency thinks it sees. Funnelfox’s performance-drop guide warns that restructuring in response to reporting changes alone can create real performance problems.
Before blaming delivery, work through this order:
- Verify the pixel and Conversions API.
- Compare Meta-reported conversions with CRM and revenue data.
- Check frequency and CTR against the recorded baseline.
- Review the interventions against their actual observation windows.
The first question is whether the platform’s numbers match the money. Judge the agency against the business’s acquisition economics: new-customer CAC, qualified pipeline, or closed revenue, according to the KPI agreed at the start. Do not quietly substitute a more flattering metric at the review.
Now read the queue row by row. Give each hypothesis one status:
- Confirmed: the evidence supports it within the agreed window.
- Refuted: the evidence does not support it.
- Still observing: the window needs extending before a defensible decision.
Keep observation separate from interpretation. “CPA fell 12% inside the window” is an observation. “The new creative caused it” is an interpretation. Isolating an intervention makes that interpretation more useful; it does not make every other influence disappear.
Then choose the action for each row: continue, roll back, or extend observation. The prior state should already be recorded, so reversal is a decision rather than an archaeology project.
Rollback has limits. Meta treats pausing an ad set for seven or more days as a significant edit. If your plan depends on reactivating a paused mature ad set before that threshold, put the deadline in the queue. Restoring settings is not a promise that delivery will return to exactly the state you left.
This is also the moment to judge the agency’s operating discipline. Can it explain why each intervention happened? Did it respect the budget limits? Did it leave functioning tests alone long enough to learn anything? Can it distinguish monitoring from intervention without reaching for a change-count brag?
Ask the questions before hiring, too: What is your kill rule for a new ad? How long before you touch a fresh campaign? What would make your own team’s week-two edit volume a red flag? An answer about “proactive optimization” is not an answer.
Cost: the final four days of the takeover, including time spent admitting what remains unknown. Buys: a decision about the agency based on evidence and judgment. Cut: activity totals as proof of value. A thousand edits can be a thousand decisions, or one unresolved argument repeated a thousand times.
5. Day 15: Lift the approval bottleneck, not the discipline
The fourteen days are scaffolding. When the initial constraint lifts, keep the queue, the KPI definition, the budget ceilings, and the discipline around significant edits. What changes is the approval bottleneck—not the requirement to justify execution.
Once the record shows that interventions consistently carry an owner, a hypothesis, and a rollback condition, routine changes inside an approved hypothesis can proceed without waiting for another meeting. Keep one execution owner. Keep both operators reading the same queue. Widen permission only inside boundaries already agreed.
The mechanism is straightforward: when the decision rights and reversal conditions are settled in advance, less time disappears into asking who may do what. Observation can turn into action faster without turning the account back into a shared argument.
That is the operating-model distinction I care about in groas’s approach: continuous execution with named strategic ownership, rather than a bill justified by visible busyness. Its flat monthly fee and month-to-month terms reject percentage-of-spend billing. The standard here is just as blunt: explain the decision, not how busy the team looked making it.
Cost after day fourteen: maintaining the record and respecting its boundaries. Buys: faster justified action without parallel editing. Cut: meeting-by-meeting approval for routine work already covered by the agreed plan.
Back to the toy seller. The troubling part was not simply the change count. It was that nineteen days later, the log could not explain which change was supposed to do what. A thousand recorded decisions with reasons attached is a brutal fortnight of work. A thousand unexplained ones is noise wearing a retainer.
Take the five fields to kickoff: owner, hypothesis, affected assets, observation window, rollback condition. On day 15, remove the unnecessary waiting. Do not remove the reason for the change.

