To deploy an enterprise AI agent, start with one narrow task, give the agent only the tools and permissions that task needs, test it against an evaluation set built from real cases, and keep a human approval step wherever it changes data or reaches a customer. Once it is live, monitor it like any production service, settle data protection questions under GDPR and KVKK before go-live, and cap the cost per task. The seven steps below follow that order.
Agent or chatbot?
A chatbot or RAG assistant answers questions from your organization's documents. An AI agent breaks a goal into steps and carries them out through tools: it looks up a record in the ERP, classifies an incoming email, fills in a request form and sends the result to an employee for approval. The difference is that an agent can act inside your systems, and that is where most of the risk comes from. If all you need is accurate answers with citations from internal documents, an enterprise chatbot with RAG may be a simpler and lower-risk place to start.
1. Choose the task
A good first task is frequent, follows rules that are written down or can be, has a measurable outcome and produces errors that can be reversed.
| Good first task | Risky first task |
|---|---|
| Classifying incoming supplier invoices and creating draft entries in the ERP | Issuing payment instructions without approval |
| Routing support tickets by topic and priority | Sending final replies to customers with no human check |
| Extracting specific clauses from contracts into a checklist | Making legal judgments and acting on them |
| Collecting order status from several systems into a summary | Changing records in several systems at once |
Once you have picked the task, write down how it is done today: who does it, which systems they check and when they escalate. That written process becomes the basis for both the agent's instructions and the evaluation set. Decide at this stage how you will measure success, for example handling time, error rate or the hours a person spends on the work each week.
2. Define tools and permissions
Every tool the agent can use, whether an API call, a database query or sending an email, should be defined individually and follow the principle of least privilege.
- Give the agent its own service account, separate from human users, with permissions limited to the task.
- Separate read tools from write tools, and at first limit write tools to creating drafts.
- Expose narrow API endpoints that validate inputs instead of giving direct database access.
- Log every tool call with who it ran for, what input it received and what it returned.
- Cap the number of steps and calls a single run can make.
3. Build an evaluation set
A few successful demos do not tell you whether an agent works; an evaluation set built from real cases does. Assemble it from past cases and record the expected outcome for each. Add edge cases next to the common ones, and include cases where the right answer is to hand over to a person. Track at least these measures:
- Task success rate: did the agent reach the expected outcome?
- Wrong action rate: did it do something it should not have?
- Handover rate: which cases did it pass to a person, and were they the right ones?
- Cost and time: how many model calls and how long per task?
Rerun the set whenever the model, the instructions or a tool changes. Base the go-live decision on thresholds you agreed in writing beforehand.
4. Design human approval
Human approval is not a weakness; it is a design element that lets the agent operate safely. Set the level of approval by the impact of the action.
| Action type | Suggested control |
|---|---|
| Reading, summarizing, classifying | No approval; regular checks on a sample |
| Creating drafts (records, emails, reports) | A user reviews the draft and sends it |
| Writing to a system of record | Approval by an authorized person |
| Payments, customer communication, contract actions | Human approval in every case, plus a second check |
As monitoring data shows the agent performing reliably, some approvals can move to sample-based review. That decision should be written down and have a named owner.
5. Set up monitoring
An agent in production is a production service. Keep traces that show every run's steps, tool calls and results. Set alerts for sudden changes in error or handover rates. Give users a simple way to mark an output as right or wrong, and have a person review a sample every week. Because a model provider's version update can change behavior, run the evaluation set against the live system at regular intervals too. Before go-live, make sure you can stop the agent with a single switch.
6. GDPR, KVKK and data protection
If the agent touches personal data, answer the data protection questions at design time.
- Purpose and data minimization: the agent should see no more data than the task requires.
- Where the model runs: cloud, on premises or hybrid, and whether data leaves the jurisdiction.
- The contract with the model provider: whether inputs are used for training and how long they are retained.
- Logs: prompts and outputs can contain personal data, so access and retention must follow policy.
- Transparency: employees and customers should know that an AI system processes their data.
In the EU, GDPR sets these obligations. In Türkiye, the framework is the Personal Data Protection Law No. 6698, and the Personal Data Protection Authority published a guide on generative AI and personal data protection (in Turkish) in November 2025 for organizations that develop or use these systems. For a broader review of your compliance position, see our KVKK and ISO 27001 compliance service.
7. Control cost
Model calls are a cost that grows with usage. To keep it manageable from the start:
- Budget per task: cap the tokens and calls a single run can use.
- Model choice: use smaller, faster models for routine steps and more capable models only where needed.
- Caching: cache repeated queries and fixed instructions.
- Measurement: report cost per task and compare it with the cost of doing the same work manually.
- Alerts: notify the team when daily or monthly spending thresholds are crossed.
Go-live checklist
- Task, scope and success measure are written down.
- Service account and tool permissions follow least privilege.
- The evaluation set has been run and results are above the agreed thresholds.
- Human approval steps are live and have named owners.
- Logs, alerts and the kill switch have been tested.
- The data protection assessment and any required privacy notice are ready.
- Cost limits and reporting are in place.
- Users are trained and have a way to give feedback.
Running the first release with a limited user group, or in shadow mode where the agent proposes and a person acts, is a low-risk way to see how it behaves on real data.
Next step
ALGOFACT builds AI agents that connect to the ERP, CRM and document systems companies already use, and deploys them with an evaluation set, human approval and monitoring. Work starts with a 30-minute discovery call, followed by a founder presentation of the proposed scope. If you have candidate tasks for an agent, book a discovery call.