Own the Pipes, Not the Keys: Who Should Own What Between an Agency and Its Client
- •
- 7 min read
An AI agent checks plugin updates across dozens of WordPress sites in minutes. This is a great time-saver, but checking is not the same as completing the maintenance work.
Someone still needs to review the results, investigate compatibility warnings, test staging, and confirm that critical functions such as checkout still work. The real question is how much work the agency has removed, not how quickly the agent ran.
Marketing and web development agencies have automated tasks like maintenance, uptime monitoring, and reporting for years. AI agents add something new: they work across tools, interpret results, and decide what to do next. With the right integrations, they can inspect sites, prepare changes, and trigger workflows without anyone directing each step.
That lets an agency handle more work without growing headcount at the same rate. It also brings costs that are easy to miss. Someone defines what the agent may do, checks its decisions, maintains its integrations, and answers when it gets something wrong.
Call it the cost of delegation. An agency that ignores it can automate itself into a more complicated business. One that counts on it, can build repeatable services and keep control of client sites.
Before you delegate work to an AI agent, you need three answers: what the agent may do, what evidence will prove the result is correct, and whether the workflow saves more than it costs to operate.
That holds whether you use a commercial platform, a custom agent, or a mix of existing tools.
A scheduled script can spot an outdated plugin and send an alert. An assistant can explain the risk and suggest a response. An agent can gather data from several tools, prepare an update, test it on staging, and ask for approval.
An agent may be useful when a task requires information from several systems and the next step depends on what it finds. A script is usually the better choice when the steps and conditions are already known. In that case, an agent can add complexity without adding much value.
Flexibility also raises the question of how much authority to hand over. OWASP lists excessive agency among the risks for LLM-based applications and names three root causes: excessive functionality, excessive permissions, and excessive autonomy. Its example fixes are modest: a read-only scope, and a person who reviews each draft before it goes out. Limit what an agent can do rather than relying on a prompt that says what it shouldn’t.
WordPress is building the plumbing. The Abilities API, part of WordPress 6.9+, lets core, plugins, and themes register what a site can do and expose it through REST endpoints. The MCP Adapter makes those abilities available to AI clients as MCP tools, resources, and prompts. Neither tells you whether a change is appropriate or a test is good enough. Your team still decides that.
The need for disciplined maintenance is obvious. Patchstack reported 11,334 new vulnerabilities in the WordPress ecosystem in 2025, 42% more than in 2024. Ninety-one per cent were found in plugins. These are reported vulnerabilities, not successful attacks, and their severity varies. The practical point is that you need a reliable way to identify issues, assess their risk and prioritise the right response.
An agent can pull data from dozens of sites and group similar issues, which saves gathering time. A developer still must investigate odd results, review risky changes and confirm key functions work. Somebody also must maintain the workflow. Measuring only the time the agent spends running a task overstates the savings, so count the whole process:
| Monthly measure | Illustrative example |
| Manual work avoided | 2 hours |
| Review and checking | −3 hours |
| Net time saved | −1 hour |
Suppose an agent saves two hours a month and reviewing its work takes three. You haven’t saved time. You’ve added a process. The fix may be narrower responsibilities, better tests, or a simpler design, and some tasks are better left to scripts or people. Judge the net benefit of the whole workflow, not the number of tasks the agent completes.
Track staff hours per site (review included), time from detection to resolution, failed updates and rollbacks, and time spent waiting for approvals. Compare the figures before and after you introduce the workflow.
Answer these before any workflow goes live. If your answer is “not sure,” the workflow isn’t ready.
Give every agent its own identity and never lend it a developer’s login. WordPress Application Passwords are revocable, per-application credentials, but requests still run with the permissions of the associated user.
Create a dedicated account with the narrowest role that works – keep staging and production credentials separate.
“Treat your AI API key the same way you’d treat your payment credentials.” Rocket.net
An Editor account can fix a typo, rewrite a product description, or change a price. Those carry very different consequences, so define authority by workflow, not just by role:
| Workflow | The agent can | A person decides |
| Site health checks | Collect data, flag anomalies | What needs attention |
| Content preparation | Draft, translate, suggest metadata | Facts, brand voice, legal claims |
| Plugin maintenance | Assess updates, prepare tests | Approval, based on risk |
| Code changes | Propose patches, run tests | Review and deployment |
| Payments and customer data | Flag issues within an approved scope | Every privileged operation |
A read-only health report can run unattended. A routine, reversible update can run under a policy. A database migration, a payment-related change, or a production deployment needs explicit approval and a tested way back.
Decide in advance what happens on an unfamiliar error: the agent stops, reports, and follows the recovery plan instead of hunting for another way in.
A successful API response shows only that a request was accepted. Check that the homepage and key pages load, forms submit, a WooCommerce checkout completes in a controlled test, integrations still work, and error logs are clean. Keep the results so a developer can see what passed and what remains uncertain.
“The automated output is a first draft, not a deliverable.” Rocket.net
Name one person per workflow. They review its performance, watch for failures, maintain its controls, and can switch it off. The AI vendor, the host, and the integration developer each control part of the system, but you still answer to the client.
Every automation creates a maintenance obligation. Someone must keep its integrations working, review its permissions, update its tests, and understand what to do when it fails.
We call the accumulated burden agency debt. It is the work an agency inherits after an automation goes live, even when the original builder has moved on. Include that work in the business case, or the apparent saving may be hiding a new support cost.
Caching shows the pattern without any AI involved. Jezweb, an agency with more than 1,000 sites on Rocket.net, said in its case study that every caching setup it tried eventually broke something, such as forms or plugins added later, and that one blank white page nearly cost it a client. The tool worked until something changed, and then someone paid for it.
Treat each agent workflow as something you operate, not a one-time build. Document its purpose, permissions, dependencies, tests, owner, and recovery procedure, and review them whenever something changes. Don’t automate a broken process either. Simplify it first: cut the pointless steps, define success, then automate and measure.
“Make Peter Parker proud and be careful to make changes against staging rather than production to make sure you don’t affect your live site in ways you don’t intend.” Rocket.net
Hosting can’t make an agent decide well. It can make a bad change easier to spot and undo. Rocket.net provides several pieces of that:
Test the recovery process before relying on it. Restore a backup in a safe environment, confirm that the activity log captures the changes you need to investigate, and check that staging tests cover the site’s critical functions.
Having these features available is not the same as being able to recover from a failed change.
Clients don’t buy maintenance because they care which AI tools you use. They want working sites, protected data, and problems handled well. Once agents are involved, expect questions like these:
Answer with evidence: the permissions, approval process, logging, testing and recovery arrangements that exist. Log the identity, the action, the approval, the result, and the verification evidence, and protect those logs from tampering.
For EU clients, data protection needs attention, too. Where you act as a processor of personal data, GDPR Article 28 says you must not engage another processor without the controller’s prior specific or general written authorization. An AI provider can count as one, so have your contracts and data processing agreements reviewed by a qualified professional.
Don’t promise that AI will never make a mistake. Show how you limit the damage, detect problems, and respond.
Pricing follows from this. If agents cut routine hours, charge for the service and its outcomes (fewer incidents, faster detection, tested changes, clear reporting) rather than for minutes. Be upfront about what automated monitoring can and can’t catch. For more on recurring revenue, see Rocket.net’s guide to the Top 6 Ways for Agencies to Build MRR.
Don’t connect an agent to every client site. Pick one low-risk workflow, such as a monthly read-only site-health report, and write down:
Measure the current process, run the agent alongside it for an agreed period, compare the results, and review the errors. Widen its authority only when the numbers support it.
An agent can check a hundred sites before your coffee goes cold. Whether you saved anything depends on what comes after: the review, the tests, the maintenance, and the recovery plan.
Before scaling an agent workflow, measure the work it removes and the work it creates. If review, maintenance, and recovery outweigh the savings, simplify the workflow or stop using it.
Capability is not the same as permission. An agent may be able to change a product price, deploy code, or modify a production database, but the higher the consequences of a mistake, the stronger the approval, testing, and recovery controls must be. The same rule applies to delegating work to a trainee.
Autonomy should be earned through evidence of reliability, not granted because integration makes an action possible.
Delegate the task. Verify the outcome. Keep ownership of the service.
Grow your business with lightning-fast, secure, and optimized websites that are easy to set up & manage. Top-tier agencies and online businesses choose Rocket.net as their trusted managed WordPress hosting provider – why shouldn’t you, too?