Stage 1
Free workflow audit
Half an hour on a call or screen share. You show the chore exactly as you do it today. We note every click, login, judgment call and failure point, and tell you honestly whether an agent is the right fix.
Process, guardrails, ownership
How a custom ad agent gets built, supervised and handed to you.
An agent with access to your ad accounts should be built like any other system that touches money: in stages, with written limits, a log, an off switch and a clear owner. Here is the whole process, including the parts most builders skip.
Six stages
The build process, step by step
Stage 2
Written blueprint and fixed quote
A short document: what the agent watches, what it does, what needs approval, what it never touches, which tools and permissions it needs, and the fixed price and dates.
Stage 3
Build and dry runs
We code the agent, connect it with least-privilege access, and test against test accounts, read-only data and dry-run requests until its outputs match what a careful person would do.
Stage 4
Supervised pilot
The agent runs on your real accounts, but each proposed action waits for your yes or no. We log what it caught, what it got wrong and the time it saved, and tune the rules.
Stage 5
Go live with limits
Approved jobs run on schedule. Anything you chose to keep under approval stays there. Caps, blackout hours and protected campaigns are set in the configuration.
Stage 6
Handoff and care
You receive the code, prompts, rules and a plain-English runbook. An optional care plan covers monitoring, platform updates and rule changes as campaigns evolve.
Guardrails
Seven guardrails built into every agent
Permission tiers
Every action is labeled automatic, needs approval or forbidden, and the code enforces the label.
Spend caps
Maximum change per action and per day, in dollars and percent, even for approved changes.
Protected lists
Campaigns, keywords, audiences and pages the agent may read but never change.
Dry runs
Changes are validated with the platform before they are applied, where the API supports it.
Audit log
Each run, draft, message and edit with a timestamp and the approver's name, stored where you can reach it.
Rate limits
It never floods an inbox, a customer or an API, even if something upstream goes wrong.
Kill switch
One control pauses every scheduled run and blocks all writes. A second step revokes its tokens.
Ownership
You own the code, the prompts and the keys
- The source code lives in a repository in your account, not ours.
- Prompts, rules and approval lists are plain files you can read and edit.
- API keys, model keys and hosting are in your name and billed to you at cost.
- Your ad, CRM and customer data stays in your systems; the agent processes it and keeps only the logs you choose.
- No lock-in: any competent developer can maintain it from the runbook.
Security
Locked down like a payroll system
- Separate service accounts and system users, never shared personal logins.
- Look-only permissions for any job that only needs to look.
- Secrets kept in a secrets manager, not in code or chat threads.
- Model providers and settings chosen so your data is not used for training, with the terms shown to you.
- Health information only once a BAA is in place with each vendor; finance and lending messages always go past a person.
Money
Build cost versus running cost
There are two numbers to understand. Build cost is a one-time project: discovery, code, connections, testing and the pilot. Running cost is monthly: hosting, AI model usage and any paid tools the agent calls, plus optional care.
The ranges in the chart are what agencies publish for 2026. They are not our prices; our own list is not out yet. Every Eye To Ad Media agent is quoted as a fixed price after a free 30-minute workflow audit, with no hourly meter. Most single-task agents for small businesses land at the low end of these ranges.
What pushes cost up or down
- Number of systems it connects to, and how clean their APIs are.
- Whether it only reads, or also writes to ad accounts.
- Platform access approvals that take time.
- How much judgment the task needs versus fixed rules.
Show the data as a table
| Build type | Published range |
|---|---|
| Simple single-task agent | $5k to $25k |
| Custom integrated agent | $15k to $75k |
| Production multi-agent system | $100k to $450k+ |
After launch
What the optional care plan covers
- Daily health checks: did every run finish, did any API call fail, did approvals pile up.
- Platform updates: new API versions, retired fields, policy changes.
- Rule tuning as your offers, seasons and campaigns change.
- A monthly note on time saved, approvals and anything we would change.
- Same-day response if the agent misbehaves, starting with the kill switch.
How we measure it
Hours saved, logged against your own before-pilot estimate; response time to new leads; how often you approve versus reject its suggestions; and errors caught before they cost money. Revenue claims are your call to make, not ours.
Know the limits
What we will not automate
- Final budget and strategy decisions.
- Prices, guarantees, health or legal claims in public copy.
- Replies to angry customers or anything that sounds like a dispute.
- Hiring, firing or anything about a specific employee.
- Payments, refunds and billing changes.
- Any task where a mistake cannot be caught before a customer sees it.
These agents exist to give people time back, not to replace them. If a task is better done by a person, we will say so on the audit call.
Questions
Process questions
How long does the supervised pilot last?
Long enough to see the agent handle your normal week and at least one unusual one. For a daily agent that is typically a few weeks. We end the pilot when its suggestions are consistently right and you are comfortable with the approval flow, not on a fixed date.
What happens if Google or Meta changes their API?
Both platforms release new API versions and retire old ones on a published schedule, and they change policies in between. On a care plan we track those notices, update the agent before a deadline, and test it against your accounts. Without a care plan, you get the documentation and code to have any developer do the same.
Who pays for the AI model usage?
You do, directly, at the provider's price. Keys sit in your accounts, so you can see the usage and the bill. Most single-job advertising agents use modest amounts because the heavy lifting (pulling data, math, rules) runs as ordinary code and the model is only called for language tasks.
Can we start small and add jobs later?
Yes, and we recommend it. Each new job is added as its own module with its own permissions and approval rules, so a reporting agent can later gain a search-term module without touching what already works.
Ready to map your first agent?
Start with the free workflow audit. You leave with a written view of what we would build, the approvals it would need and a fixed quote, or an honest "do not automate this".