# Agentic AI security: independent consultant, agency or SaaS? How to choose

[Original page](https://danilolapegna.com/guides/agentic-ai-security-consultant-vs-agency-vs-saas.html)

Language: en. Content updated: 2026-10-08.

An honest comparison of independent consultant, agency and SaaS product for securing AI agents in production, and when each one is the right fit.

An AI agent can read data and take actions, such as drafting an order or sending a customer reply. Securing it requires technical controls and someone to maintain them. The choice between a product, agency and consultant depends on which work is missing.

Before requesting proposals, prepare one page: workflow, accessible data, allowed actions, observed problem and internal owner. Give every candidate the same case so you compare deliverables rather than unrelated sales presentations.

## What are you buying?

**A subscription product**, often called SaaS, may filter inputs, detect suspicious behaviour or collect events. Its controls need configuration, integration and testing. Someone still owns decisions the product does not make.

**An agency or specialist firm** may provide several skills and operational continuity. Verify the actual assigned team, technical decision-maker and contracted coverage.

**An independent consultant** may deliver a bounded intervention and transfer the method to your team. Check relevant competence, availability, absence cover and handover materials as you would for a firm.

These options can work together. A product may remain in operation while a specialist designs controls, and targeted external work can complement a capable internal team.

## Compare evidence, not categories

| Question | Evidence to request from every candidate |
|---|---|
| Who fixes the problem? | Named people or roles, required access and included responsibilities |
| How will we demonstrate success? | A legitimate request that succeeds and a forbidden action that is blocked |
| What remains with the team? | Configurations, tests, documentation and restoration instructions |
| Who responds afterwards? | Coverage, agreed response times, absences and escalation |
| What is the full cost? | Setup, integration, licences, usage, maintenance and internal work |
| How can we switch supplier? | Exports, transferable materials and a tested handover |

Costs and start times are not universal properties of a supplier category. A product may need substantial integration; a firm may offer a small engagement; a consultant may require other specialists. Compare the same scope and period.

## A concrete example

An agent prepares draft orders, and the company wants proof that it cannot confirm them alone. Ask the candidate to map permissions, restrict them in a test environment and run two cases: draft creation should succeed; final confirmation using the agent's identity should fail.

A suspicious-instruction detector may help, but it does not replace access enforcement. If the internal team can apply and maintain the correction, a targeted engagement may be sufficient. If the service needs continuous coverage and nobody is available outside office hours, that becomes a separate requirement. Agent count alone does not decide the delivery model.

The [agentic-security guide](https://danilolapegna.com/guides/agentic-ai-security.html) helps prepare these questions. For the permissions test, see [least authority](https://danilolapegna.com/guides/least-authority-ai-agents-italian-smb.html).

## When the internal team is enough

Work internally when you have relevant skills, configuration access, time to test and a maintenance owner. A prescribed organisational chart is unnecessary; the ability to do and verify the work matters.

Buy a specific missing skill where needed. If nobody will maintain the control after delivery, assign that responsibility first. Otherwise a good intervention may lose effectiveness after the next update.

## Test the handover

My criterion is a handover exercise: your team reruns a test and changes a control without the supplier's intervention. If it cannot, identify the obstacle: missing access, incomplete instructions, a technical dependency or a skill still to learn.

This tests independence better than a “no lock-in” promise. Code and documentation help only if the remaining team can use them. Agree the exercise and expected team capability before signing.

## The next step

Prepare the initial page and request a proposal for the first verifiable problem. Resolve unclear critical requirements before granting production access or expanding the engagement.

The [AI agent control and security service](https://danilolapegna.com/services/hardening-controllo-agente-ai) can start from that case. Where continuous coverage or additional expertise is the main need, include it in supplier selection rather than discovering it after delivery.

## Frequently asked questions

### Independent consultant or agency for agentic AI security?

It depends on missing work, required continuity and internal skills. Ask every candidate for the same test, handover materials and total cost. A product and specialist support can be combined.

### How much does it cost to have a consultant secure an AI agent?

Cost depends on actions, data, integrations, access and coverage. Define an initial test and compare setup, usage, licences, maintenance and internal work over the same period.

### When do you not need any external help?

When the team has relevant skills, configuration access, testing time and a maintenance owner. Verify those capabilities on the actual case rather than assuming a separate role is needed for every function.