Data & AI platform by Data Analytics

AI agents with business data: oversight before execution

An AI agent can break down a question, select available tools and use intermediate results for a next step. That does not make it an independently reliable employee. You determine which tasks, data and actions fall within its remit. ENABLE, the manageable Data & AI platform from Data Analytics, provides a data foundation for analysis, apps and AI. An agent application also requires its own execution rules, controls and management. Performance and value remain dependent on your setup, data quality and specific process; absolute security or autonomous task completion is not guaranteed.

The examples on this page are hypothetical situations, not customer cases.

Meet our team

At Data Analytics

Trusted by data-driven organisations

Define a task with an identifiable endpoint

Choose a process where a proposal or analysis has value without immediate execution. Specify the input the agent receives, the expected outcome and when the process stops. Avoid instructions such as “optimise the entire business”. A bounded comparison of anomalous purchase invoices is easier to test than an unrestricted improvement assignment.

Determine in advance who reviews the result and which source serves as the reference. An agent can carry incorrect assumptions into later steps and amplify errors. Require visible source references, the selections used and unresolved uncertainties. A convincingly worded explanation is not proof that the reasoning followed or the underlying calculation is correct.

Separate planning from permitted tool execution

Jordy · Data Analytics

The model can propose a plan and tool calls; the executing application determines whether those proposals are permitted. Validate every call against fixed parameters, user permissions and task context. MCP can expose tools to a suitable client, but does not itself grant unrestricted authority. Give the agent only the functions the task actually needs.

Set limits on the number of steps, tool calls, runtime, result size and cost. Stop execution when limits are exceeded, errors recur or permission is missing. Budgets must be enforced technically, not merely written into instructions. Also determine who may resume a stopped task and how you prevent the same step from being unintentionally repeated.

Limit data permissions and decide on model processing

Have the server and underlying data sources enforce permitted business entities, rows and fields. An agent must not retrieve through a second tool information that the first tool restricts. Treat retrieved documents and tool results as untrusted input: hidden instructions in them must not replace system rules or access restrictions.

Explicitly decide whether your selected data may be processed by an external LLM. Assess the chosen provider, contracts, settings and retention periods. An internal tool does not automatically keep model processing internal. For a private option, also document where the model runs and which administrators have access. Send only the context the task requires.

Work with proposals and explicit human oversight

Begum · Data Analytics

Hypothetical example: an agent compares purchase invoices with purchase orders and flags possible differences. The agent gathers permitted data and creates a draft list with source references. An employee reviews whether the differences are genuinely relevant. The agent does not approve payments or change accounting entries; those actions are outside this initial, read-only remit.

If you later want to allow changes, design separate write tools and scopes. Require explicit human approval for the specific action and parameters, with server policies that verify that approval. Agree recovery procedures, rollback where possible and idempotency. Permission for an analysis is not automatically permission to execute every proposed follow-up step.

Measure process quality before expanding

Test representative tasks alongside edge cases, missing data, malicious instructions and tool outages. Measure not just speed, but also source accuracy, missed exceptions and the number of corrections required. Compare results with a predefined reference. Track how much work human oversight actually requires; a fast draft may still demand substantial review.

Assign an owner for instructions, tools, model selection and incident handling. Reassess changes when a data source, tool description or model version changes. Data Analytics can explore alignment with ENABLE and the necessary integrations with you. Expand only when agreed quality thresholds are met and a workable stop and recovery procedure is available.

Frequently asked questions

Is an agent the same as a chatbot?

Not necessarily. An agent can plan multiple steps and call tools within a task. How much is executed depends on the application, available functions and configured controls.

Can an agent work without supervision?

There is no general guarantee of that. Match oversight to risk and have important outcomes reviewed. Specific write actions require separate permissions, explicit human approval and server checks.

How do you prevent unlimited costs?

Set technically enforced budgets for time, steps, tool use and cost. Specify when the task stops and who may approve resumption. Simply asking it to work economically is insufficient.

What data suits an initial trial?

Choose a limited, permitted dataset with clear definitions and verifiable outcomes. Where possible, start with non-sensitive data and reading. Test missing information and exceptions before expanding the scope.

Technical documentation

Discuss your question with Data Analytics

Which sources, users and checks does your application need? Discuss your question in an ENABLE demo.