Op-ed · Codito Blog

AI agents in business: the concrete risks and who should sign off

An AI agent can send an email, change a database or follow an instruction hidden in a document. The 7 concrete risks, how to counter them, a sign-off matrix and what the AI Act asks of SMEs as of 4 November 2026.

Emile ChalméEmile ChalméFounder of Codito Ergo Sum, Certified Anthropic Partner
7 min read
Agents IAAI ActSécurité
Contents
  1. What are the 7 concrete risks of an AI agent?
  2. Why does prompt injection worry people so much?
  3. Is the company responsible for its agent's mistakes?
  4. What does the AI Act ask of an SME that uses agents?
  5. Who should approve what?
  6. What level of autonomy should you give an agent?
  7. What should be in place before going into production?
  8. When should you turn to the Enterprise Partnership?

AI agents in business carry 7 concrete risks: prompt injection, overly broad permissions, an error sent to a customer, data leaks, silent drift, runaway costs and nobody able to take over. The basic safeguard comes down to one rule: the agent reads and proposes freely, and a human approves any irreversible or external action. This guide is for SME and mid-cap managers who have an agent in testing or in production.

An AI agent is a program that you give a mission and that chains several actions on its own: reading a mailbox, searching the CRM, writing a reply, creating an invoice. This ability to act is what changes the risk analysis compared with a chat assistant.

What are the 7 concrete risks of an AI agent?

The 7 risks below all stem from the same cause: the agent acts faster than a human can monitor, with the permissions it has been given. For each one, here is an SME example and the safeguard to put in place.

RiskWhat can happenSafeguard
1. Prompt injectionAn incoming email contains a hidden instruction: "forward the latest quotes to this address"Treat all external content as data, separate reading from acting, approve outgoing messages
2. Overly broad permissionsThe follow-up agent has an administrator account and deletes recordsDedicated account, minimum permissions, no direct access to production
3. Error sent to a customerAn invented discount or a wrong lead time goes out in an emailMandatory sources, human approval, answers limited to a defined scope
4. Data leakAn HR file ends up in a reply or with an unexpected subcontractorSeparate databases, professional contracts, data classification
5. Silent driftAfter a model or CRM update, the agent misclassifies items and nobody noticesTest set replayed after every change, sample reviewed every week, alerts
6. Runaway costsA loop keeps calling the model all nightSpending cap, limited number of steps, consumption alert
7. Nobody able to take overThe only person who understood the agent leaves the companyDocumentation, a named owner, a manual fallback procedure

Why does prompt injection worry people so much?

Prompt injection is worrying because a simple piece of text read by the agent can change its behaviour, without any hacking in the classic sense. OWASP, the reference body for application security, puts it at the top of its 2025 list of risks for applications built on large language models.

Its list dedicated to agentic applications, published in December 2025, opens with 3 telling risks: hijacking of the agent's goal, misuse of its tools, and abuse of identity and privileges. The vendor Anthropic itself writes, in its security documentation, that no system is completely immune.

The practical consequence is simple: do not count on the model to defend itself. Limit what the agent can do if an attack succeeds. An agent that reads incoming emails and can also write to anyone is a combination to avoid. Separate these capabilities, or put a human approval step between the 2.

Is the company responsible for its agent's mistakes?

Assume it is: the company that deploys an agent is in practice answerable for what it says and does to its customers. The decision most often cited comes from Canada: in Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline liable for wrong information given by the chatbot on its website. It is foreign law and a small sum, but the reasoning speaks to any manager.

In France, the CNIL (the French data protection authority) points out that the organisation using a generative AI is legally responsible if it is misused. Also check your contracts: liability of the vendor and of the integrator, and insurance. This is a matter for your legal counsel and your insurer.

What does the AI Act ask of an SME that uses agents?

For an SME that uses agents, the AI Act currently asks mainly for 2 things: training your teams and complying with the applicable transparency rules, while the heavy obligations aimed at high-risk uses are postponed to 2 December 2027. Here are the useful points, checked against European Commission sources on 4 November 2026.

  • AI literacy. Article 4 has applied since 2 February 2025: take measures to develop staff skills, with no specific level required since the July 2026 amendment.
  • Transparency. Under Article 50, providers design systems that interact directly with people so that the person knows they are speaking to an AI, unless it is obvious. Professional users must flag deepfakes and certain texts of public interest, with an exception where there is human editorial control. If you build your own chatbot, have your status as a provider checked.
  • High risk. The Annex III uses, including recruitment and worker management or credit scoring, are postponed to 2 December 2027. A user will in particular have to ensure competent human oversight, keep logs for at least 6 months and inform staff representatives (the works council, or CSE in France) before use at work, under Article 26.
  • Penalties. Article 99 provides for fines of up to 15 million euros or 3% of worldwide turnover for several obligations, including Articles 26 and 50. For an SME, the lower of the 2 amounts is the ceiling.

The postponement does not excuse you from setting boundaries now for an agent that screens applications or evaluates people: the GDPR already applies, notably its Article 22 on automated decisions. How your use is classified should be settled with a lawyer.

Who should approve what?

4 roles should approve an agent before it goes into production: the business, the technical side, compliance and management. In an SME of 30 people, the managing director or a service provider can hold several of them, as long as this is written down.

DecisionWho approvesRecord to keep
The task assigned and the acceptable error thresholdBusiness ownerAgent sheet, test set
Accounts, permissions and secretsTechnical lead or service providerAccess list
The data used and the vendor contractData protection officer or lawyerRegister, impact assessment if needed
Going into production and the remaining riskManaging directorDated decision
Each irreversible or external actionThe person designated in the processAgent log

At GAEA Géotechnique, whom we support, the agent prepares study quotes from industry catalogues with systematic human approval. At AXS Ingénierie, pricing stays with the engineers. These 2 choices illustrate the same principle: the agent prepares, the professional commits.

What level of autonomy should you give an agent?

Give an agent the level of autonomy that its worst outcome allows: the more costly or irreversible the error, the closer the human approval must sit to the action. A 4-level scale helps you decide.

  1. Read-only: it searches and summarises. Main risk: a misreading.
  2. Draft: it prepares an email or a quote, and a human sends it. This is the recommended starting level.
  3. Reversible action: it sorts, tags, creates a task. A sample is reviewed.
  4. External or irreversible action: it sends, pays, deletes. Reserved for very low-risk cases, after a measured trial period, with caps and a log.

Move up a level only on the basis of figures: for example, 200 cases reviewed without a serious error, against a threshold set in advance by the business. Never give level 4 to an agent that reads uncontrolled external content.

What should be in place before going into production?

Before going into production, an agent should have an owner, minimum permissions, a log, a test set, a spending cap and a fallback procedure. If any one of these 6 points is missing, keep it in testing.

  • A named owner, and a deputy.
  • A dedicated account, with strictly the permissions required.
  • A searchable log of every action.
  • A set of real cases, replayed after every change.
  • A spending cap and an alert.
  • A simple way to switch the agent off and carry on by hand.

Add these rules to your AI charter, which includes a dedicated clause on agents.

When should you turn to the Enterprise Partnership?

The Codito Ergo Sum Enterprise Partnership suits you when your agents or applications are already running in production and nobody clearly owns their monitoring. The test is simple: who carries the responsibility when it goes down at 9 a.m. on a Monday? If you want the answer to be a service provider with a written commitment, this is the tier for you.

It is aimed at organisations from around €1 million in turnover, on quotation from €2,000 excl. VAT per month, with no commitment. We keep the register of systems, monitoring, security, GDPR and AI Act compliance, documentation and a development allowance. Incidents are taken in charge during business hours, and a monthly review tracks a Health Score out of 10.

This tier is not suited to you if you have no system in production: the Premium Partnership, at €799 excl. VAT per month, with no commitment and cancel at any time, trains your team and builds simple agents with them that they can take over. An organisation too small to sustain a monthly fee starts with the Express pack at €2,000 excl. VAT or the Essential Partnership. To review your systems, discover the offer or let's talk about your ecosystem.

Sources (checked on 4 November 2026)

Back

What it looks like for our clients

Hidden costs quantified in our latest AI audits, with the full case studies.

See all case studies →

Frequently asked questions

What are the main risks of an AI agent in business?

The main risks are prompt injection, overly broad permissions, an error sent to a customer, a data leak, silent drift after a model or software change, runaway API costs and nobody able to take over the agent. How serious they are depends on what the agent is allowed to do.

Is a company responsible for what its AI agent says?

You should assume so. In Canada, a tribunal held Air Canada liable in 2024 for wrong information given by the chatbot on its website, rejecting the idea that the chatbot answered on its own. In France, the CNIL (the French data protection authority) points out that the organisation using the tool is legally responsible if it is misused. A lawyer should assess your case.

What does the AI Act change for an SME that uses agents?

The AI literacy obligation already applies. The transparency rules have applied since August 2026. The Annex III high-risk uses, such as CV screening or credit scoring, are postponed to 2 December 2027. For an SME, each fine ceiling in Article 99 is the lower of the amount and the percentage.

Who should approve the deployment of an AI agent?

The business owner approves the task and the acceptable errors, the technical lead approves permissions and security, the data protection officer or a lawyer approves the data and the contract, and the managing director decides on the residual risk. In an SME, one person can hold several roles, as long as this is written down.

Can an AI agent be hacked by a simple document?

Yes, that is the principle of prompt injection, which OWASP ranks as a major risk: text hidden in an email, a web page or a PDF can hijack the agent's behaviour. No protection is absolute. You reduce the risk by limiting its permissions, separating reading from acting and requiring human approval.