EU AI Act deadlines: what they mean for Agentforce

The EU AI Act deadlines that Regulation (EU) 2026/1744 postponed matter less for most Agentforce projects in manufacturing than the one that has already passed. Since 2 August 2026, Art. 50 requires that people chatting with an AI agent know it is an AI, and in my reading a dealer service agent falls under that. The new high-risk dates, 2 December 2027 for Annex III and 2 August 2028 for Annex I, hit you only if an agent helps decide about applicants or employees, or if AI sits in your machines. This is my reading as an architect, not legal advice.

What did Regulation (EU) 2026/1744 change?

According to White & Case1, the regulation, known as the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July. It defers the high-risk obligations in Chapter III, where most of the work sits. The consolidated Art. 1132 on the Commission’s AI Act Service Desk gives the dates below; the 2 December 2026 date for machine-readable marking comes from the Commission’s FAQ on Art. 503:

1. 2 Aug 2026Transparency under Art. 50: chatbots, deepfakes, published text
2. 2 Dec 2026 · next dateMachine-readable marking under Art. 50(2) for generative systems on the market before 2 August 2026
3. 2 Dec 2027High-risk systems under Annex III, stand-alone uses such as recruitment and workforce management
4. 2 Aug 2028High-risk systems under Annex I, AI in regulated products

White & Case stresses that the deferral is limited to Chapter III and does not cover transparency. The AI literacy duty in Art. 4 got softer: providers and deployers now take measures that support their staff’s AI literacy instead of ensuring a sufficient level.

Who is provider and who is deployer in an Agentforce project?

The AI Act assigns duties by role. Under Art. 34, a provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name. A deployer uses an AI system under its authority in a professional context. By default, Salesforce provides the platform and you as the manufacturer deploy the agents you configure in your org.

Art. 25(1)5 shifts that split: a deployer becomes a provider if it puts its own name on a high-risk system, modifies one substantially, or changes the intended purpose of a system so that it becomes high-risk. Agentforce is a general platform; if you configure an agent that ranks job applicants, you arguably create that high-risk purpose yourself, and with it the provider duties of Art. 16. Salesforce’s AI literacy guide6 from May 2025 urges customers to consult their own legal counsel. I would take that literally: the classification of your agents belongs with your legal team.

A dealer chat agent is a transparency case

Most Agentforce agents at manufacturers answer dealers, customers or service partners about orders, delivery dates, spare parts or appointments. None of that is on the Annex III list. These agents fall under Art. 50(1), applicable since 2 August 2026.

According to the Commission’s FAQ, the duty applies when the system is AI, is built for a two-way exchange, communicates directly, and the counterpart is a natural person. A purchasing clerk at a dealer is a natural person, so as I read the FAQ, B2B does not take you out. An agent that only classifies cases in the background is outside; one that answers the dealer is inside. The notice has to come at the start of the first interaction, clearly and accessibly. The exception for “obvious” AI is to be read restrictively, so I would not rely on a robot icon. The FAQ names fines of up to 15 million euros or 3 percent of worldwide annual turnover.

Formally, Art. 50(1) is a provider duty. Whether that means Salesforce or you, the FAQ does not settle for this setup; in practice the agent’s welcome message lives in your org. My recommendation: put the disclosure there, in every portal language. Salesforce’s Trailhead module on guardrails and trust patterns7 adds clear rules for handing off to a human and lists the Einstein Trust Layer audit trail among its implementation best practices.

The labelling of published text under Art. 50(4) concerns information to the public on matters of public interest, not a delivery date for a dealer.

When does an Agentforce agent become high-risk?

Annex III lists stand-alone high-risk uses. In a manufacturer’s Salesforce org, point 4 on employment is the one to watch. The Commission’s Service Desk page on employment8 separates the cases:

• In scope, 4(a) recruitment: ranking candidates, scoring applicant answers, targeted job ads.
• In scope, 4(b) work relationships: assigning shifts or on-call windows based on worker performance, evaluating performance.
• Exempt under Art. 6(3): scheduling interviews, acknowledgement emails, organising CV data, as narrow procedural or preparatory tasks.
• Never exempt: systems that profile natural persons.

The field service angle is easy to miss. Assigning jobs by skill, territory and availability is far from 4(b). Ranking technicians by first-time fix rate and assigning work accordingly comes close to allocating tasks based on individual behaviour, although whether a rule-based optimiser is an AI system at all is a separate question. How the engine decides is in scheduling automation in Field Service.

If an agent is high-risk, Art. 269 gives you as deployer concrete duties from 2 December 2027: use it according to the provider’s instructions, assign human oversight to competent people with authority, check the input data you control, monitor operation, keep automatically generated logs for at least six months unless other law provides otherwise, and inform workers’ representatives and affected workers before use. In Germany the works council is involved much earlier anyway; that path is in AI coding tools, works council and GDPR.

Annex I concerns the machine, not the CRM agent

Annex I covers AI in products under EU harmonisation law. For machine builders, Cuatrecasas10 reports the key change: the Machinery Regulation (EU) 2023/1230 moves from section A to section B of Annex I. Only a few AI Act provisions then apply directly to AI in machinery. The substantive high-risk requirements are to be added to the Machinery Regulation through delegated acts, which the Commission must adopt by 2 August 2028 at the latest. Until those delegated acts or the corresponding harmonised standards exist, manufacturers can, according to Cuatrecasas, keep referring to the harmonised standards or common specifications adopted under the AI Act to show presumption of conformity.

This concerns AI in machine control or safety functions and belongs with product engineering and legal, not with your Agentforce agents. The new Art. 6(1a)11 also states that AI systems “solely used for non-safety related aspects of user assistance” are not safety components. Art. 6(1b) takes precedence: AI whose failure or malfunctioning would endanger health and safety is a safety component. An agent that suggests a spare part to a technician is user assistance in that sense.

Example: four AI ideas at a machine builder

Assume a machine builder with SAP, Service Cloud, Field Service with 60 technicians and a portal for 250 dealers. All circumstances are assumptions. The AI steering group asks which AI Act date applies to four ideas.

Dealer service agent in the portal. It answers order status and delivery dates. Transparency case under Art. 50(1), due since 2 August 2026. Action now: disclosure in the welcome message in all portal languages, a handoff on request, and a test case that checks the disclosure, as described in testing Agentforce agents before go-live.

Technician assignment by performance score. Field service wants complex jobs to go to the technicians with the best fix rate. Possibly Annex III 4(b) from 2 December 2027, and a works council topic today. My recommendation: assign by certified skills and availability, and keep performance data out of dispatch.

Applicant pre-screening for service technicians. HR wants an agent that ranks applications. Clearly Annex III 4(a), and under Art. 25 the company likely becomes the provider. The alternative is an agent that only schedules interviews and sends acknowledgements, both listed among the Art. 6(3) exceptions.

Condition monitoring in the next machine generation. Product engineering embeds a model that detects wear. Annex I applies only if it acts as a safety component; pure maintenance support could fall under Art. 6(1a), unless its failure would endanger health and safety under Art. 6(1b). Either way a workstream for product engineering, outside the CRM programme.

One idea is overdue, two get redesigned to stay out of high-risk, and one belongs to another team.

List every agent in the org

1. List every agent and AI feature in the org, with purpose and the people it talks to.
2. Classify each one by role, provider or deployer, and by category: transparency, Annex III, Annex I or none.
3. Fix Art. 50 now for every agent that talks to people.
4. Design HR and workforce use cases into the Art. 6(3) exceptions where possible, and plan for Art. 26 where not.
5. Check log retention of agent sessions and the Trust Layer audit trail against the six months in Art. 26.
6. Put AI Act roles into contracts with Salesforce and your implementation partner.
7. Get legal sign-off on the classification before go-live.

Where the classification shapes the architecture, for example which decision needs a human approval step, I support the client side as part of Salesforce architecture. Whether Agentforce is the right start at all is covered in Agentforce for manufacturers: start now or wait?, and chapter 9 of the AI masterclass puts the AI Act next to works council and GDPR.

Sources11

1. 1White & Case (opens in a new tab)whitecase.com
2. 2Art. 113 (opens in a new tab)ai-act-service-desk.ec.europa.eu
3. 3Commission’s FAQ on Art. 50 (opens in a new tab)digital-strategy.ec.europa.eu
4. 4Art. 3 (opens in a new tab)ai-act-service-desk.ec.europa.eu
5. 5Art. 25(1) (opens in a new tab)ai-act-service-desk.ec.europa.eu
6. 6AI literacy guide (opens in a new tab)salesforce.com
7. 7Trailhead module on guardrails and trust patterns (opens in a new tab)trailhead.salesforce.com
8. 8Service Desk page on employment (opens in a new tab)ai-act-service-desk.ec.europa.eu
9. 9Art. 26 (opens in a new tab)ai-act-service-desk.ec.europa.eu
10. 10Cuatrecasas (opens in a new tab)cuatrecasas.com
11. 11Art. 6(1a) (opens in a new tab)ai-act-service-desk.ec.europa.eu