Enterprise engagement
From a read-only pilot to approved execution, in four bands
A PaidSync enterprise engagement moves through four autonomy bands, observe, propose, act with guardrails, and full auto. The agents execute in live ad accounts, not just recommendations, across 460+ tools and 14 platforms behind one endpoint, https://mcp.paidsync.ai/mcp, with OAuth 2.0 sign-in. Every write is approval-gated until you widen the band, and some actions never run without a human tap. You start read-only, set the success criteria, and decide per agent how far to go. Founder direct. Pricing scoped after discovery.
The four bands
Each band widens what the agents may do. You decide when.
The bands are not a sales phase. They are a setting on each agent in your workspace, enforced on our servers, and a client cannot widen them. The default is propose. Audit and report agents default to observe. Only an admin or owner can move an agent into a band that executes on its own.
Observe
The read-only pilot. Agents read and report. Nothing is written to any account.
- OAuth sign-in to the accounts in scope. You connect each platform yourself, and the tokens are encrypted at rest with AES-256-GCM.
- A seat for each person, viewer or editor.
- The questions you want answered first, and the dashboards you will check us against.
- Account audits and wasted spend reports per account.
- Cross-platform performance in one query, and scheduled checks that land as digests.
- A written baseline. The success criteria you set are measured against it.
Propose
The default band. Every change the agent wants to make becomes an approval card. Nothing runs until someone on your team taps it.
- A brand brief with monthly caps, in total and per channel.
- The campaigns that are off limits, so no card can touch them.
- Who may approve. Any editor, admin or owner can tap a card, and the executor is recorded on it.
- Cards with a dry-run preview, claimed atomically so two people cannot execute the same one.
- One-tap undo. The inverse is built at apply time and can be re-run safely. A new campaign can be undone for seven days. Some platforms return no safe inverse, and the card says so.
- A 7-day outcome check by a daily job, nine days for conversion tracking changes. A regression of $50 or 20% writes a rollback card.
Act with guardrails
Agents apply changes inside ceilings you set. Anything outside a ceiling becomes a card.
- The monthly caps from your brand brief. We divide them by 30.4 into a daily ceiling.
- Action limits per brand, up to 20 per run and 50 per day, with a default of 10 per run. By default a single run cannot move more than one channel-day of budget.
- The short list of tools the waste guard may use on its own, up to 10. The default list holds one entry, pause campaign.
- An execution trail per run, and a queue when the daily ceiling is reached. An alert fires at 85% of cap.
- Protected campaigns and the budget limit enforced in every acting band, not only this one.
- A brand-level kill switch. Turn it off and every automatic action becomes a card. Blocked actions are never dropped.
Full auto
Only where you choose. Set per agent by an admin or owner, for the work you have already watched run well in band 3.
- The agents and brands that qualify. Everything else stays where it is.
- Whether a regressing change may revert on its own. That needs three things on at once, the guardrails switch, auto-revert, and coexistence mode set to auto.
- Automatic rollback, only for changes the agent itself made. Anything your team changed by hand is left alone.
- The same caps, the same trail, the same kill switch. The never-auto list below still applies.
- A production acting switch that fails closed. If it is off, missing, frozen or unreadable, every run drops back to cards. Human taps are exempt.
What never runs on its own
In every band, these need a human tap. A dynamic tool the system cannot classify is denied by default.
- Campaign launches (every apply_ tool and apply_bundle)
- Image and video uploads (create_image_asset, upload_fb_ad_image, upload_fb_ad_video)
- Applying or subscribing to platform recommendations
- Device and location bid adjustments
- Adding audience targeting
- Removing campaigns or product partitions, excluding product partitions
- Raw mutate calls and paidsync_exec
- Saving, running or deleting workflows
- Creating accounts, including under a manager account
- Anything named delete_, remove_ or destroy_
Success criteria
You set the number each band has to hit.
We write the criteria down in discovery, before the first agent runs. Each band reports against them, and no agent moves up until your number says so. These are the ones teams usually pick.
Findings your team agrees with
Wasted spend found as a share of monthly budget, and how many audit findings your own strategists would have made.
Cards approved without edits
The share of cards tapped as proposed, the time from card to tap, and the undo rate.
Zero ceiling breaches
No protected campaign touched, no cap exceeded, every run with a complete trail.
Regressions caught, not found
How many of the 7-day checks flagged a change before your team noticed, and how many rollbacks were needed.
Access and data
Approvals are shared. Data stays yours.
Viewer, editor, admin, owner
Viewers are read-only. Editors and above can execute and undo cards. Admins and owners set bands. Every execution records who tapped it.
Encrypted at rest, bound per session
Platform tokens and any model keys you bring are encrypted with AES-256-GCM. Account bindings live per session, so one chat cannot reach into another's accounts.
Read in chat, not yet exportable
Every live mutation is logged with user, account, tool, source and actor, readable through the change-log tool up to 100 rows at a time. CSV, SIEM and webhook export are not built. Deletion is self-serve, 30 days of grace, then a hard purge. See the privacy policy and the security page.
You talk to me. I run discovery, I set the bands with you, and I stay on the line while the agents run. My team and I have managed over $500M in ad spend, so the pilot starts read-only because that is how I would want it on my own accounts.
Pricing
Scoped after discovery. The call is free.
We price the engagement once we know your accounts, spend, platforms and which band you want each agent to reach. Self-serve plans stay available for individual seats, Free at $0 with 15 tasks, Pro from $99, Team from $249. Details on the pricing page.
Proof and reading
Agentic advertising is already here
The dated record of what ran, on which platform, and what changed.
20 things teams do with PaidSync
From a wasted spend sweep across a manager account to a weekly cross-platform report.
Not sure the enterprise engagement is the right shape? The enterprise overview covers what we build, and the ads agents page shows the specialist agents that run inside each band.
Get in touch
Start with a read-only pilot
Tell us what you're running. We reply within 24 hours, founder direct.
- You talk to the people who built it and ran the spend.
- We map your accounts, set the success criteria, and scope the bands on the call.
- Pricing is custom and quoted after we understand the work.
Frequently asked questions
What are the four autonomy bands?
Observe, propose, act with guardrails and full auto. Observe is read-only. Propose turns every change into an approval card, and it is the default band. Act with guardrails lets an agent apply changes inside the ceilings you set, with anything outside a ceiling becoming a card. Full auto is set per agent, by an admin or owner, only where you choose. Audit and report-digest agents default to observe.
Do we have to start read-only?
Yes. Every enterprise engagement starts in the observe band. It is how you and we agree on a baseline before any agent proposes a change, and it is how you check the numbers against your own dashboards. Moving to propose is your call, and the agents cannot widen their own band.
Who can approve a card?
Anyone in the workspace with the editor, admin or owner role. Viewers are read-only. Executing a card and undoing one are both shared actions, and the person who executed is recorded on the card. Only admins and owners can set an agent to a band that executes on its own.
What can the agent never do on its own?
Campaign launches, bundle applies, image and video uploads, applying platform recommendations, device and location bid adjustments, adding audience targeting, removing campaigns or product partitions, raw mutate calls, creating accounts, saving or running workflows, and anything whose name starts with delete, remove or destroy. These always need a human tap in every band. A dynamic tool the system cannot classify is denied by default.
What happens if a change hurts performance?
Every applied change is checked seven days later by a daily job, nine days for conversion tracking changes. If the metric it targeted regressed by $50 or 20%, a rollback card is written for your team. Automatic revert runs only when you have turned it on, with auto-revert enabled and coexistence mode set to auto, and only for changes the agent itself made.
How do you stop everything at once?
Turn off the brand-level guardrails switch and every automatic action becomes an approval card. Blocked actions are never dropped. Above that, a platform-level acting switch fails closed: if it is missing, frozen, corrupt or unreadable, no agent, scheduled run or auto-revert can execute. Human approval taps are exempt, so your team keeps working while the agents wait.
Where does the execution trail live, and can we export it?
Every live mutation is logged with the user, account, tool, an argument summary, its source and the actor. Dry runs are not logged. Your team reads the trail in chat through the change-log tool, up to 100 rows at a time, and each automation run keeps its own execution trail. CSV, SIEM and webhook export are not built yet. If you need them, say so on the call and we scope it.
How much does an enterprise engagement cost?
Pricing is scoped after discovery. We price the engagement once we understand your accounts, spend, platforms and which bands you want each agent to reach. The consultation is free. Self-serve plans are Free at $0 with 15 tasks, Pro from $99 and Team from $249.
Do you offer SSO and DPA on request?
Yes. SSO and DPA on request. Raise it on the discovery call and we scope it with the engagement.