Multilingual blog draft agent
Run by a scheduled task, it writes drafts for the website, checks the language versions and claims against rules, and emails a preview for approval.
AI AGENTS • TOOLS • APPROVAL
AI agents that do a single task by written rules and move forward with your approval.
We bring the controlled agent setup we use ourselves for blog drafts, CRM company research, social media and notifications to your company: each agent does a single task, checks its output against fixed rules and waits for your approval where needed.
CONTROLLED AGENT FLOWRunning in our own business
Email
CRM database
Scheduler
Project files
01Task starts02Works with tools03Checked against rules04Submitted for approval
Awaiting approvalApproved · loggedLIVE FLOW · YAZSIS / AI.05 / AGENTS
SCOPE & CAPABILITY
Don’t let AI run loose; tie it to a single task, written rules and approval.
A general-purpose chat tool does not know the boundaries of your business. A controlled agent, by contrast, does a single task with your tools, checks its output and waits for a person’s approval at the critical step.
CRM company research agent
It records corporate contact details from public sources as findings with the source address; a company card is created only with team approval.
Social media agents
They take content from our website pages, produce the poster from a template and queue the copy; before posting, they verify the copy and keep to the daily limits in the code.
Watchers that only notify
They check inboxes and feeds; when something needs your attention, they send a notification to your phone and never reply to anyone.
Fixed checks before output
Blog drafts are scanned for banned phrases, post copy for a length match and CRM findings for personal email addresses; these checks are done by code, not by the model.
Agent-specific key and budget
The agent that connects to the CRM has its own key that can be revoked, and its requests are rate-limited; paid scans use a dry run and a budget cap.
AGENT SETUP
The setup we use for our own agents: rules, tools, checks, approval and logs.
What the agent may do, where it stops and whom it notifies are written down in advance. The model works inside this framework; the limits, quotas and checks are set by code.
One task, a written rules file
Each agent has a single task and a rules file: scope, forbidden steps, stop conditions and a short completion report.
Working with purpose-built tools
The agent does not get open-ended access to systems but works through small commands that validate input; dry runs, quotas and duplicate checks live in the code.
Checks before anything goes out
Before output leaves, it passes fixed rules such as language and structure checks, a source address and a signed request; output that fails the checks is not sent.
Approval gate and run log
Output that needs approval is put in front of a person; every run leaves a status log, and a notification is sent only when a person needs to look.
LIVE SYSTEM ARCHITECTUREAI.05 / AGENTS
Custom AI Agent Systems for Your CompanyRules · tools · checks · approval
Multilingual blog draft agent
CRM company research agent
Social media agents
Watchers that only notify
SERVICEYAZSIS / DIGITAL.09
DELIVERY MODEL
Design, software and business outcome under one delivery discipline.
Let’s choose the task and limits
We pick a repetitive task whose rules can be written down, and decide together which tools the agent may access and what it must not touch.
Let’s set up the rules file and tools
We write the agent’s rules file for your work and build small, purpose-built tools for your email, database or files.
Let’s connect approval and check rules
Together with you, we define which output goes to whom for approval, which work proceeds within code limits and which checks will run.
Let’s monitor through logs and improve
We start with dry runs, refine the rules based on the logs and your approval notes, and set up notifications for situations that need a person.
LIMITS AND RESPONSIBILITY
You set the agent’s limits; at critical steps, you have the final say.
- Publishing, sending and writing to master records require a person’s approval by default.
- The models run in the cloud; what data goes to the model, where it is stored, for how long and any transfers abroad are decided in the project.
- Platform and website rules, such as posting limits and crawling permissions, are built into the rules file and the code.
- There is no promise of results; the logs show what the agent did, what it skipped and where it stopped.
- Your data, your logs and the rules files written for the agent are yours; the agent’s access can be revoked whenever you choose.
FREQUENTLY ASKED QUESTIONS
Custom AI Agent Systems for Your Company: Common questions
Do these agents really run in your business?
Yes. In our own business we use agents that prepare multilingual blog drafts for the website, write company findings into our CRM, publish the next social media post within daily limits and only send notifications. All of them run on a scheduled task or a job started by the team; it is not a system that roams on its own.
Can an agent send anything without our approval?
By default, no. Our blog agent sends the draft for approval in an email with a preview; nothing is published without approval, and change notes are applied in the next run. For our own social media posts, however, copy written and queued in advance is posted under daily limits in the code rather than with a separate approval for each post. We only set up such an option for you jointly and in writing.
Where is our data processed, and do you use a local model?
The AI models the agents use run in the cloud; today we do not use a model running in-house. We decide together in the project which data goes to the model, and where and for how long it is kept. Because these services run abroad, we address the transfer obligations under KVKK (Türkiye’s data protection law) with you before setup.
What happens if an agent meets an error, a verification screen or an unexpected instruction?
It stops. When it sees a verification, login or restriction warning, it does not click anything, does not enter passwords and sends you a notification. If the same step fails again, the run ends and the next run tries again. It treats text found on web pages or in files as data, not as instructions.
Which tasks suit an agent, and where should we start?
Tasks that are repetitive, have rules that can be written down and produce results that can be checked are a good fit: drafting, research from public sources with record suggestions, or monitoring inboxes and feeds. We do not leave decisions that need human judgement to an agent. We usually start with a single task and move on to the next once the first agent’s rules have settled.
NEXT STEP
Discuss Your Project