01Pre-flightTest run preview
Checked before launch
Budget, targeting, naming and policy checks, plus a test run, before anything goes live.
- Your rules, checked first
- A test run before any change
- Approval where you want it
Trusted execution layer for AI ad ops
Let agents run your ad accounts. Keep every decision.
Cesteral governs every AI-driven change across the campaign lifecycle: checked before launch, guarded while live, and proven after it ends.
run/q3-leadgen-001
- 01 Pre-flight40/40Checks run before anything goes live
- 02 In-flight95%A 12% search budget increase, watched against your spend limit
- 03 Post-flight10Errors prevented · budget shifted to search by 12%
13+ ad platform connectorsApache 2.0 connectorsImmutable audit trail
How it protects youThree phases, one approach
Secure at every phase
One set of controls covers the whole run, so your team can let agents work on live campaigns without losing control.
02In-flightSpend limits
Guarded while live
Spend limits stay on while campaigns run. A risky change is blocked the moment it would break one.
- Automatic spend limits
- Pause one agent, or all of them
- The right person alerted at once
03Post-flightAudit trail
Proven after it ends
Every decision is recorded against plan, with the checks that ran, who approved it and what happened next.
- A receipt for every decision
- Confidence checked against results
- A complete audit trail
01 — Pre-flightChecked before launch
Describe the outcome. The agent runs the ops.
Build audiences, launch campaigns, shift budgets or pause what isn't working, in plain language and across every ad platform. The agent plans the change, checks it against your rules, and waits for your approval where you've asked for it. Every step is logged.
Cesteral· YouTube AdsAsk a follow-up…
02 — In-flightWatching while live
Safeguards that stop risky changes
Set your limits once. When a change would break a spend limit, skip a required check or hit a broken connection, Cesteral stops it and sends it to a person to review.
01Spend risk
Spend limits
Stops a risky budget change before the money moves, so a fast overspend never turns into cleanup.
Daily spend95% of limit
Limit active02Rule checks
Change review
Catches naming, targeting, link and budget problems before an AI-proposed change reaches the ad platform.
Risk vs. your limitOver limit
Needs review03Connection health
Account access
Spots expired logins, unavailable tools and broken connections before an agent can act.
Connections
- Google AdsConnected
- MetaConnected
- LinkedInLogin expired
03 — Post-flightSample data · Period 12 · 418 decisions
The numbers, with the working shown.
Anyone can put 99% on a landing page. These figures come from our decision log — immutable, with each entry pinned to the policy revisions it was judged against — so every one traces back to the decisions behind it.
Sample dataIllustrative decisions that show the format, not live account data.
Rule-breaking changes caught before launch
98.6%
98.6 in 100 stopped before any spend1.4 not caught until after launch
How this is measured
- Definition
- Changes that broke your rules and were stopped before any spend, divided by all changes that broke your rules.
- Counts
- 1,186 of 1,203 rule-breaking changes, periods 1–12
- Source
- The decision log: one entry for every attempted change.
- Period
- Periods 1–12
Safe changes blocked by mistake
0.7%
safe changes cleared · 99.3%blocked in error · 0.7%
Guardrails that block everything are useless. These ones rarely stop a change that was safe.
How this is measured
- Definition
- Safe changes that were blocked, divided by all safe changes checked. A change counts as safe when a person later marked the block as a mistake.
- Counts
- 28 of 4,000 safe changes checked, periods 1–12
- Source
- The decision log, matched to the review of each block.
- Period
- Periods 1–12
Changes with a full record
100%
one line per action418 / 418 recorded
No sampling. Every change is recorded before it runs — no record, no execution.
How this is measured
- Definition
- Changes with a decision record written before they ran, divided by all changes.
- Counts
- 418 of 418 changes, period 12
- Source
- The decision log, compared with the ad platforms' own change logs.
- Period
- Period 12
Confidence you can trust
0.042
average gap between confidence and results
When Cesteral says it's 87% sure, it's right about 87% of the time.
That's why the cut-off can be trusted: a 0.62 waits for a person, and a 0.91 goes straight through.
How this is measured
- Definition
- The average gap between how sure Cesteral said it was and how often it was right, across ten confidence bands (expected calibration error). Lower is better; 0 is perfect.
- Counts
- 418 changes with a recorded outcome, in 10 bands
- Source
- The decision log and each change's measured outcome.
- Period
- Period 12
Decision receipts
Sample dataIllustrative decisions that show the format, not live account data.
Open any decision. Read the receipt.
Every line in the log opens to the evidence behind it: what changed, which policies ran, who approved it, how it turned out, and the policy revision it was judged against.
ClearedRaise Search budget +12%q4-brand · google ads#0410 · 14:02:11 · confidence 0.91
$4,200 / day$4,704 / day
- Daily budget cap · $4,704 ≤ $5,000
- Pacing · 78% of month
- Change size · +12% ≤ +20% limit
- Naming convention · q4-brand_srch_*
Approved automatically: confidence was above 0.70 and the change stayed inside the +20% limit.
+3.4%
7-day window, ad platform reporting
- Decision
- #0410
- Rule version
- budget.auto.v4
- Request fingerprint
- a17f…9c2e
ClearedShift spend to Advantage+retarget · meta#0411 · 14:09:40 · confidence 0.84
Manual retarget · 60%Advantage+ · 60%
- Audience overlap · 2.1%
- Budget neutral · $0 net
- Creative approvals · 6 / 6
Approved automatically: no change to net budget, and every creative was already approved.
+1.9%
7-day window, ad platform reporting
- Decision
- #0411
- Rule version
- budget.neutral.v2
- Request fingerprint
- b042…71aa
Held for a personPause underperforming ad setprospecting · tiktok#0412 · 14:21:03 · confidence 0.62
Active · $310 / dayPaused
- Confidence · 0.62 < 0.70
- Learning phase · exits in 2 days
- Spend vs plan · 38%
Held and sent to the prospecting owner for approval. No change has gone to TikTok.
Pending review
No change has gone to TikTok
- Decision
- #0412
- Rule version
- confidence.floor.v3
- Request fingerprint
- c9e1…3f50
Blocked by your rulesExpand creative to new geolaunch · linkedin#0414 · 14:52:47 · confidence 0.41
Markets · US, UKMarkets · US, UK, DE, AT
- Approved markets · DE, AT not listed
- Localized legal copy · missing · DE
- Budget cap · within $12k
Blocked by your rule geo.markets.v3. The campaign owner was told and nothing was sent.
Not run
Nothing reached the platform
- Decision
- #0414
- Rule version
- geo.markets.v3
- Request fingerprint
- e5a9…12bd
Select a decision to open its receipt
Method · period 12 · 418 decisions across Google Ads, Meta, TikTok and LinkedIn · Illustrative: shows the format, not live account data.
04 — DeploymentThree ways in, one set of rules
A layer, not a destination.
Whether your agents reach the ad platforms through our app, your own servers or the AI assistant your team already uses, the same rules, approvals and record go with them.
Open source · Apache 2.0
Our open-source connectors let any AI agent work with the major ad platforms. Read the code, change it or run it yourself. When your team needs approvals, secure account access and a full record, Cesteral Pro adds them.
Every connector comes with its safety rules built in, including a test run before any change.
cesteral/mcp-open-advertisingTwo changes run back to back. The first is cleared and executes. The second fails a check, is held for your team, and never reaches the platform.
- 03AgentsPropose changes.
- 02Governance layerChecks each change against your policy, then records it.
- 01Ad platformsRun only what is cleared.
01Open source
Run it yourself
agentyour agent
cesteralrules · approvals · record
runs onyour own servers
Run the open-source connectors on your own servers. You can inspect every request, and you keep the keys and the data.
View on GitHub02Managed
We run it for you
agentcesteral app
cesteralrules · approvals · record
runs oncesteral cloud
We run the connections for you. Link your ad accounts in minutes and manage everything from one place, with nothing to host.
Explore integrations03Claude, ChatGPT and more
Use your own AI assistant
agentclaude · chatgpt · cursor
cesteralrules · approvals · record
runs onyour Cesteral workspace
Connect Claude, ChatGPT or your own agent to Cesteral. Everything it does follows your rules and approvals and is recorded, with no new tool to learn.
Setup guide for AI assistants05 — Questions6 of 11
Asked before every rollout.
What teams want to know first when they put Cesteral on production campaigns.
All questionsCesteral is a control layer for AI-driven campaign execution. It lets agents plan and run work across ad platforms while your team reviews, approves, and audits what happens before spend is affected.
No. Teams can require review and approval before high-impact actions run. The point is to get the speed of AI execution without giving up control over budgets, targeting, or launch decisions.
Each official server covers one platform, and their safety settings don't match: new items start paused on one, while another allows full edit access. They're also run by the company selling you the media. Google's official server is read-only, so there is no official way to make approved changes on Google at all. Cesteral gives you one set of rules, one approval flow and one audit trail across every platform — owned by you, not by the platform you're buying from.
Yes. Cesteral works with any AI assistant that supports MCP — Claude, ChatGPT, Cursor or a custom agent — with secure sign-in and separate access keys for each connection. Every action your agent takes follows your organization's rules, approval limits and audit trail. Your team keeps working in the tools they already use; the controls travel with them.
The open-source connectors are free under Apache 2.0. The paid product is Cesteral Intelligence — the control layer that adds approval workflows, secure account access, audit trails, cross-platform workflows and review tools for your team. Hosting is included, but the value is the control, not the servers.
Every change an agent makes produces an immutable audit trail — the rule checked, the confidence score, the approval decision and the outcome. This is the kind of verifiable record that frameworks like the EU AI Act (high-risk system rules effective August 2026) and emerging US state AI laws increasingly expect for automated decisions that affect ad spend. Cesteral does not provide legal compliance certification, but it is built to produce the records compliance teams need.
Put your agents on the record.
Start free with the open connectors. Talk to us once real spend is on the line.
Apache 2.0 connectorsWorks with any AI assistantImmutable audit trail