The next wave of AI features does not answer; it acts. Agents that research, negotiate, book, purchase and file. The legal system has centuries of experience with one entity acting for another; what it has never had is an actor with no legal personality, no intent, and a temperature setting. Until legislators catch up, agentic products are governed by contract law, agency analogies and product design. Here is how we structure them.
Contract formation: the easy part, surprisingly
Under Turkish law (TBK) and most European systems, a contract formed through an automated system binds the person who deployed the system; the machine is a tool transmitting a declaration of will. Your agent clicking “confirm” binds you or your user the way a standing order does. The real questions are one level deeper: whose agent was it in that transaction, what authority did it have, and who bears its mistakes?
The authority matrix
Agency law’s oldest tool translates directly into product architecture. For every agentic capability, define:
| Tier | Agent may… | Example |
|---|---|---|
| Autonomous | Act and inform after | Research, drafts, reversible config changes |
| Bounded | Act within pre-set limits | Purchases under a spend cap, standard reorders |
| Propose-only | Prepare, human executes | Contract terms, payments above cap, anything irreversible |
| Blocked | Never | Legal commitments, credentials, regulated filings |
Write the matrix down. It becomes three documents at once: your product spec, the user-facing terms (“what the assistant may do on your behalf”), and your defence exhibit when an agent does something unexpected; because it shows the boundary was designed, not improvised.
Mistakes, hallucinated purchases and error doctrine
When an agent buys the wrong thing, classical error and consumer-withdrawal rules do more work than any AI-specific theory. B2C flows retain statutory withdrawal rights regardless of who clicked. The hard cases are B2B and irreversible services, which is exactly why the matrix pushes those to propose-only. Between platform, agent vendor and user, allocate mistake risk contractually now; default rules were not written for a stochastic intermediary.
The compliance overlay
- Article 50: counterparties interacting with your agent must be able to know it is one (from August 2026, EU-facing); an agent that negotiates by email as “Mehmet from procurement” is a designed deception.
- Payments: agent-initiated transactions must fit authorisation frameworks; pre-authorised mandates and caps, not stored credentials improvisation.
- Logs as the agent’s memory of record: every action, input context, tool call and override; retention matched to your dispute windows (see our incident-response piece).
- Upstream terms: model providers’ usage policies restrict autonomous high-stakes use; your agent may be violating its own supply chain (contract clauses).
Five clauses for agentic products
- Authority matrix incorporated into terms, with user-configurable caps.
- Explicit mandate language: the user appoints the agent within defined limits, and can revoke.
- Mistake allocation and correction protocol (unwind first, argue later).
- Counterparty disclosure commitment (the agent identifies itself).
- Log retention and access rights, including for the user’s own disputes.
Agentic AI is where product ambition runs furthest ahead of doctrine, which makes it exactly the layer to design conservatively and document generously.
This article is for general information only and does not constitute legal advice. It reflects the position as of July 2026.
Author
If this is on your desk
Templates and checklists are free in the Founder Academy; for a specific situation, book a 30-minute intro call.
Founder AcademyBook an intro call