Choose a use case for your first agent
There is a lot an agent can do for your team, and the best first one is usually closer than it looks: a task people already repeat every week, with a clear idea of what a good result looks like. Creatio AI Studio makes an agent quick to build, so the interesting question is which one to build first.
This article is the thinking you do before the build. Work through it once and you can describe your agent in a paragraph — which is exactly what the Prompt Agent Designer and AI Twin both ask you for.
What makes a good first use case
A good candidate is narrow, repetitive, and low-blast-radius. Judge a candidate against the following.
Signal | What to look for |
|---|---|
Repetitive | The task happens many times a week in roughly the same shape. One-off work gives an agent nothing to learn from and no payback. |
Bounded | You can say what is in scope in one sentence, and what the agent should refuse in another. |
Checkable | A person can look at the output and say whether it is right. If correctness is a matter of opinion, you cannot evaluate the agent later. |
Reversible | A wrong answer is embarrassing, not expensive. Save the irreversible work — payments, cancellations, external emails — until you have a governance layer around it. |
Owned | Somebody is accountable for whether it works and can decide the grey areas. |
Resist starting with the highest-value process. The first agent is where you learn the platform, so optimize for a short feedback loop rather than for impact. A drafting assistant that ten people use daily teaches you more in a week than a revenue-critical workflow that takes a quarter to approve.
Describe the agent in six parts
Before building, write down the following. Each part maps to something you configure later, so vagueness here becomes rework.
Part | Question to answer | Where it lands |
|---|---|---|
Purpose | What problem does the agent solve, and for whom? | The agent name, description, and the opening line of the system prompt. |
Inputs | What does a user bring or ask? | The welcome message, and the examples you test with. |
Outputs | What does "done" look like — an answer, a logged record, a routed request? | The response rules in the system prompt, and later your eval cases. |
Knowledge it needs | What must it be grounded on to be accurate? | Knowledge sources. Learn more: Knowledge in Creatio AI Studio. |
Actions it takes | What must it actually do, beyond answering? | Tools and skills. Learn more: Choose between a skill, a tool, and a knowledge source. |
Risks to govern | What must a human check, or the system mask or gate? | Policies. Learn more: Set up protection of personally identifiable information. |
Define success before you build
"It seems to work" is not a result you can defend or repeat. Settle the question now, because the answer shapes what you build.
Write three to five statements of the form when the user says X, the reply must Y. Make them specific enough that two reviewers would agree on a pass or a fail. These statements become your eval cases almost word for word, which is how "it worked when I tried it" turns into a score you can re-run after every change. Learn more: Evals in Creatio AI Studio.
Include at least one statement about what the agent must not do. Agents fail more often by over-reaching than by under-performing.
Define the boundaries
Boundaries are the difference between an agent you can put in front of a colleague and one you have to supervise. Decide each of the following.
Boundary | What to decide |
|---|---|
Topic scope | What the agent politely declines. Write the refusal behavior into the system prompt, not just the documentation. |
Read or write | Whether the agent may change anything at all. A read-only first agent removes a whole class of risk. Tools carry a "Read," "Write," or "Admin" risk level. Learn more: Tools in Creatio AI Studio. |
Sensitive data | Which values must never reach the model. A PII protection policy masks them before the model runs, not after. |
Human checkpoints | Which actions need sign-off before they run. Learn more: Set up tool confirmation. |
Audience | Who reaches the agent, and through which channel. An internal assistant and a public website widget are different risk profiles. Learn more: Channels. |
Work out the required data
An agent is only as good as what it can see. For each question you expect it to answer, trace where the answer comes from — and be honest when the answer is "nowhere yet."
If the answer lives in... | Then you need... |
|---|---|
Documents, policies, FAQs — content that is stable | A knowledge source. Indexed content is retrieved on demand with citations. Learn more: Add a knowledge source. |
CRM records that change — prices, cases, accounts | A tool or an integration. Indexed content goes stale; a tool reads the current value. Learn more: MCP integration. |
Somebody's head | To write it down first. This is the most common blocker, and no amount of prompt engineering substitutes for it. |
A system Creatio cannot reach | To narrow the scope, or to solve the access problem before the build. |
Quality of retrieval depends far more on how well a source is described than on how much content it holds. A source with a vague description is one the agent never reaches for.
Quality also beats volume. Ten accurate, current pages outperform a thousand-page archive holding three contradictory versions of the same policy, so prune before you index.
See also
Build, release, and monitor your first agent