AI Agent Development

What is an AI agent, and how does it differ from a chatbot or automation?

An AI agent is software that uses a large language model (LLM) as its decision engine and is given a set of tools. You do not spell out every step for an agent; you give it a goal. The agent drafts a plan to reach that goal, decides which tool to use and in what order, updates its plan based on the result of each step and reports what it did once the work is finished. The term agentic AI, heard more and more often, describes exactly this way of working.

Three traits set an agent apart from a chatbot and from classic automation: it works from a goal, it uses tools (it searches a system, reads records, creates drafts) and it plans step by step. The table below sums up the difference:

The distinction matters in practice: building an agent for a task with clear rules means unnecessary cost and risk. For such tasks, AI automation offers a simpler, more predictable solution. If the main goal is to answer customer questions, our AI chatbot page is a better place to start. An agent makes sense where neither of these is enough and the steps change from case to case.

What are AI agents useful for?

Agents create the most value in work where an employee opens several systems, gathers information, compares it and finally produces an output. The use cases we see most often are:

• Lead research and qualification: enriches incoming enquiries or a target company list with public sources and CRM records, prioritises them against criteria you set and leaves the sales team a reasoned summary.
• Reporting and data collection: pulls data from ad platforms, analytics tools and spreadsheets, prepares period comparisons, flags unusual changes and leaves the interpretation to a person.
• Resolving support requests: classifies the incoming request, reads order or account details from the relevant system, prepares a proposed solution and a reply draft, and hands cases outside its remit to a member of your support team.
• Document preparation in internal processes: turns documents such as proposals, contract summaries, meeting notes or technical specifications into drafts using approved templates and company knowledge.
• Checks and audits: finds missing and inconsistent entries in a product catalogue, price list or content inventory and produces a list of corrections.

What these tasks have in common is that a person can check the result, and an agent's mistake does not cause irreversible harm. For a first agent project, we recommend starting with a process that meets both conditions.

When is an AI agent the wrong choice?

Agents are powerful, but they do not suit every task. A language model works probabilistically; it may not give exactly the same answer to the same input every time, and it can be wrong with confidence. For these reasons, in the situations below we recommend either not using an agent at all or using it only in a role that prepares suggestions:

• Irreversible actions such as payments, refunds, deleting records or approving contracts.
• Tasks that require exact calculation, such as pricing, tax, stock or progress payments; these are handled by classic code.
• Final legal, medical or financial decisions; the agent can only prepare a preliminary review and a draft.
• Processes where data is scattered, incomplete or contradictory; the data needs to be put in order first.
• Rare tasks that an employee finishes quickly; the effort of building and monitoring outweighs the gain.

If you are not sure which process suits an agent, it is wiser to start with AI consulting: we review your processes together and weigh agent, automation and classic software options side by side.

Tools and permissions: how is the principle of least privilege applied?

What an agent can do is defined by the tools you give it. A tool is a function with a clear definition and limits that searches a system, reads a record, creates a draft or updates a field. The most critical decision in agent design is which tool can be accessed with which permission. We apply the principle of least privilege here: the agent works with the narrowest permission its task requires, and it gets no write access where read access is enough.

A separate service account is created for the agent; personal accounts and broad admin keys are not used. Every tool call is logged together with its input and result. To connect tools to your systems, direct API integration or open connection standards such as the Model Context Protocol (MCP) are among the options; we choose together based on your existing infrastructure, security requirements and ease of maintenance.

At which points does human approval come in?

A well-designed agent does not try to do everything itself; it knows when to stop and ask a person. Approval points are not a security layer added later but are designed from the start as part of the workflow. Typical approval points are:

• Any content leaving the organisation: messages and documents going to customers, suppliers or the public.
• Actions that change records: especially updates that affect several records at once.
• Uncertainty: when the information at hand is contradictory or insufficient, the agent asks a question instead of guessing.
• Exceeding limits: when a task needs more steps, time or budget than expected, the work is stopped and reported.

The approval screen should be simple: the agent shows at a glance what it wants to do, why, and the sources it relies on; the person approves, corrects or rejects. These decisions are logged as well and used to determine where the agent will need fewer approvals. The scope of approval is narrowed only after the logs show the agent is consistent in that task, and only by your decision.

How is an AI agent tested and monitored?

Agents cannot be tested with unit tests alone, as classic software can; the same task can be solved in different ways. That is why we evaluate with scenarios built from real work examples:

1. 1.A scenario set covering easy, typical and difficult cases is prepared from real examples in the process; the expected outcome is written down for each scenario.
2. 2.Scenarios that deliberately contain missing information, conflicting records and malicious content are added.
3. 3.The agent runs through the whole set in every version; results are scored both by automated checks and by human review.
4. 4.Every scenario that fails becomes a permanent part of the set; this way you can track that a fix does not break in later versions.

After launch, each task's steps, the tools used, approval decisions and cost are logged. The monitoring dashboard tracks completed tasks, the share of tasks that needed human intervention, error types, time and cost per task, and the time saved for your team, all together. We build this measurement layer with the same approach as our CRO and analytics work; the agent's impact on the business is judged from logs, not from estimates.

AI agent security: prompt injection and external content

Agents read content from the outside world: emails, web pages, uploaded documents, support requests. Trying to steer an agent with instructions hidden inside that content is called prompt injection. For example, a sentence planted in an email might ask the agent to ignore its previous instructions and send records outside. This risk cannot be fully removed today, so we base security on the architecture rather than on the model's good behaviour:

• External content is always treated as data, never as instructions; only the system instructions you define determine the agent's task.
• An agent that reads untrusted content is not given tools that change records or send data outside.
• Sensitive tool calls go through human approval regardless of the content.
• Personal data is masked where needed; the agent sees only the data its task requires.
• Scenarios containing malicious content are a permanent part of the test set.

For agents that process personal data, we clarify the data flow, retention period, the contract with the model provider and the conditions for cross-border transfer under data protection rules such as KVKK and GDPR before setup. For the technical details of model choice, hosting and data privacy, see our AI integration and LLM solutions page.

How are an agent's costs kept under control?

An agent's cost grows with the number and length of the model calls it makes in each task. An agent that plans and checks its own result makes more calls than a bot that answers a single question. We use the following methods to keep costs under control from the start:

• A limit on steps, time and budget for each task; when the limit is reached, the agent stops and reports.
• Smaller, more economical models for simple subtasks, and stronger models for difficult reasoning.
• Caching for repeated information and short instructions that carry no unnecessary context.
• Cost per task reported on the monitoring dashboard side by side with the time saved.

Model options include companies offering closed models, such as OpenAI, Anthropic or Google, and open models running on your own infrastructure. We build the architecture with the model as a replaceable component; when a better option appears, there is no need to rewrite the agent from scratch.

A single agent or a multi-agent system?

In multi-agent systems, tasks are split into separate roles, such as agents that research, write and review. This structure can bring order to complex work, but every additional agent increases cost, latency and the difficulty of debugging. Information passed between agents also creates a new surface for errors and security issues.

That is why, in most projects, we recommend starting with a single agent that works with well-defined tools. If the task truly requires independent areas of expertise, or if having one agent review another's work measurably improves quality, we move to a multi-agent structure by comparing both with scenario tests. The decision is based on evaluation results, not on how appealing the architecture looks.

Why choose UNIT İstanbul for AI agent development?

UNIT İstanbul is an Istanbul-based agency that has been running performance marketing and growth work since 2009; it delivers software and AI projects from the same team under its Unit Software brand. Because of our marketing roots, we start every agent project with the same question: which work will this agent speed up, and how will we measure it? We do not just write the agent; we deliver its tools, permissions, approval points, test set and monitoring dashboard together.

We already use the CRM, advertising, analytics and e-commerce systems an agent connects to in our daily work, which makes it easier to design the process in the same language as your team. If you have a process that might suit an agent, write to us through our contact page for a free preliminary assessment; we will look together at whether an agent fits and whether a simpler solution would be enough.