Services, AI agents and automation
The agent that finds the problem cannot act on it. The one that acts cannot decide.
Two programs, two permissions. The analyst reads your ad account every day against a written checklist and holds no write access. The executor holds the write access and can only perform a change a named person has approved. The same split runs the lead routing, the monthly ledger, and the page builds, on our own accounts before yours.
Running on our managed accounts. 23-account sweep, ledger entry 006.
Two agents, two permissions, one change
5 steps
- 01
Analyst agent
Read only, no write tools
Reads the account every day against the written checklist. Holds no write access, so it cannot change anything.
- 02
Suggestion row
Written, not applied
Each finding lands as a row: what it saw, what it proposes, the evidence, and the expected effect.
- 03
Human approval
Named analyst
A person reads the row and approves, edits, or rejects it. Nothing moves without this step.
- 04
Executor agent
Write access, approved rows only
Performs exactly the approved change. It has no checklist and cannot originate a finding of its own.
- 05
Log
Every action
Who proposed it, who approved it, what changed, and when. The record is the audit trail.
Five kinds of agent, each with a job, an input, and an approver.
Each one is described the same way: what it reads, what it writes, and who checks it before anything takes effect.
- 01
Account QA analyst
- Reads
- The ad account, the conversion ledger, search terms, the landing pages
- Writes
- A daily suggestion row per finding: budget, negatives, paused assets, broken tracking
- Who approves
- A named analyst, before the executor runs anything
- 02
Lead qualification and routing
- Reads
- Form fills, call summaries, the CRM, the calendar
- Writes
- A scored lead record on the pipeline board, with the reason for the score
- Who approves
- The owner sets the routing rules; the agent applies them and flags the exceptions
- 03
Reporting agent
- Reads
- Ad spend, verified conversions, call logs, closed deals from the CRM
- Writes
- The monthly ledger: what was spent, what it bought, what changed, what was wrong
- Who approves
- A senior analyst signs it before the client sees it
- 04
Site and landing page builders
- Reads
- The onboarding brief, converting search terms, verbatim reviews, the brand system
- Writes
- Planned pages and landing pages, built to the written frameworks, staged for review
- Who approves
- Design review by a human, then the owner, before anything goes live
- 05
Integration workers
- Reads
- Whichever systems the task touches: accounting, project management, email, calendar
- Writes
- Records in the system of record, never a message to a customer
- Who approves
- Scoped in writing per task; anything outside the scope is refused and logged
The account QA analyst and its executor are the pair running today on our managed accounts. The fleet sweep they ran in July 2026 covered 23 accounts and found live defects in 9, including ours. Entry 006 on the ledger.
Four rules the structure enforces.
An instruction in a prompt is a request, and an agent can misread it or drift from it. Missing access cannot be misread. The analyst has no write tools, so a bad finding cannot touch the account. The executor has no checklist and no judgment of its own, so it cannot invent a change. The person in between is the only thing that connects them.
Every change on a managed account therefore carries a proposer, an approver, and a timestamp, which is what lets us put the work on a public ledger.
Rule 01
Read-only by default
An agent gets write access only when a task requires it, and then only to the one system and the one action the task names.
Rule 02
An approval gate on every change
Findings are proposals. A named person approves each one before it takes effect. The proposing agent and the executing agent are separate programs.
Rule 03
Every action logged
Proposal, approver, change, timestamp. When something goes wrong, the log shows the step where it went wrong.
Rule 04
No agent talks to a client
An agent can draft a message. Only a person sends one. No email, text, or call to a client or a customer originates from an agent.
About 12 business systems, one written procedure per job.
The agents are not a chat window beside your tools. They are connected to the tools, through integrations we maintain, and they follow the same named frameworks our analysts follow.
The integration fleet
12 systems
- 01Ad platforms
- 02Analytics
- 03Search console
- 04Tag management
- 05Call tracking
- 06CRM
- 07Project management
- 08Accounting
- 09Email
- 10Calendar
- 11Website
- 12Work-management board
An integration is added when an engagement needs one, and the agent that uses it inherits the same access rules. When a vendor changes something, the procedure stays the same and the integration is repaired.
The doctrine they run
- 01Onboarding intelligence framework
- 02Site framework
- 03Landing page framework
- 04Ad-group decision engine
Each framework is a written procedure with pass and fail conditions. The agents execute it; a second analyst can verify it. When a case falls outside the procedure, the agent flags it instead of guessing.
Read how the doctrine worksWhat agents are bad at, from running them.
Most of the approval gate exists because of this list. A human stays in the loop for these reasons, and the loop is shaped around them.
- 01
Judgment on thin evidence
Three days of data is not a trend. An agent will happily extrapolate from it; a senior analyst will not. The approval gate exists for exactly this.
- 02
Anything that depends on a relationship
A client who is nervous about spend needs a phone call, not a suggestion row. Agents do not know who is nervous.
- 03
Situations outside the written procedure
The agents run doctrine. When a case does not fit the doctrine, the right output is a flag to a human, and that is what they are built to produce.
- 04
Knowing when they are wrong
An agent's confidence is not evidence. Every output on an account is treated as a claim to be checked, which is why the analyst has no write access in the first place.
- 05
Being cheap to build well
A working agent is mostly the checklist it runs, the tests around it, and the logging. The model is the small part of the budget.
Three engagement shapes. Scope in writing before any of them.
Start with the audit if you do not know where the hours go. Start with a build if you already know the one job you want off your desk. The monthly program is for owners who want the whole operation run.
Shape 01, Two weeks
Automation audit
We map where the hours go each week, across the owner, the office, and the agency. You get a written list of the tasks an agent can take, the ones it should not, and the order to build them in.
Priced per scope
Set per scope, in writing, before work starts.
Shape 02, Scope in writing
Scoped build
One agent, one job, one system of record. Built, tested against real data, logged, and handed over with the procedure it runs written down.
Priced per scope
Set per scope, in writing, before work starts.
Shape 03, Monthly program
Monthly operation
The fleet runs your account and your funnel under our doctrine, with a named analyst approving every change and a monthly ledger of what happened.
$3,500 to $7,500/mo
Monthly band, defined scope.
The monthly program is described in full on its own page, including the one engagement of this shape already on the record. AI Operations Partner. All prices are on the pricing page.
On the record
- What AI agents does MyWebReps build and run?
- Agents that QA ad accounts daily and write suggestions for a human to approve, agents that qualify and route inbound leads, reporting agents that write the monthly ledger, and build agents for sites and landing pages. All of them run against the client's live business systems.
- Can an agent change my ad account on its own?
- No. The analyst agent is structurally read-only: it holds no write access, so it cannot change anything. A separate executor performs only changes a human has approved, and it cannot originate a finding of its own.
- Which systems can the agents work with?
- About 12 so far: ad platforms, analytics, search console, tag management, call tracking, CRM, project management, accounting, email, calendar, and the website itself. New systems are added when an engagement needs one.
- Will an agent ever contact my customers?
- No. No agent sends anything to a client or a customer. Agents draft, humans send. That rule is written into the doctrine the agents run under, not left to judgment.
- What are AI agents bad at?
- Judgment calls with thin evidence, anything that depends on a relationship, novel situations outside their written procedure, and knowing when they are wrong. That is why every agent output on an account passes a human before it takes effect.
- What does an AI agent engagement cost?
- Three shapes. An automation audit and a scoped build are each a fixed fee for a defined scope, published before work starts. Monthly operation runs in the $3,500 to $7,500/mo band.
- Does MyWebReps use these agents on its own work?
- Yes. The same fleet that runs client accounts ran a 23-account QA sweep in July 2026 that found live defects in 9 accounts, including our own. The finding is filed on the proof ledger.
- How do I start?
- Book a 15-minute triage call. Bring the three tasks that eat the most hours each week; you leave with a view on which of them an agent can take.