The agency contract on your desk was written for a business that paid humans by the hour, and it no longer describes the work being done. Software drafts the keyword plan, fires the outreach emails, and stages the content calendar overnight. The partner you hired still exists, but much of what you're paying for runs on an autonomous system you can't see. Standard agency paper doesn't name that system, doesn't allocate the risk it creates, and doesn't tell you who signs for its mistakes.
That silence is where the legal exposure lives. The questions worth asking before you countersign aren't about deliverables or deadlines. They're about authority, oversight, and who eats the loss when an agent does something nobody asked it to do.
The Work Changed Under the Same Cover Page
Agencies have spent most of the past two years rebuilding their delivery models around AI agents. The pitch is familiar: faster SEO audits, content produced at a volume no human team could match, outreach sequences running without a coordinator in the loop. The November 21 press release titled Digital.Marketing featured on Adweek for its AI-driven approach to modern marketing agencies is one of the more public examples, and it captures where the industry is heading. Pricing is moving from billable hours toward deliverables and business outcomes, with humans kept on the judgment calls.
The contract templates most agencies still use predate the shift. They name a scope of services, a monthly retainer, a termination clause, and a boilerplate limitation of liability. They assume a person is pressing send on the outreach email. When the sender is an agent operating inside the agency's stack, acting on your data, inside your ad accounts, or against your brand guidelines, that assumption stops holding.
The Obvious Fix Is to Bolt On an AI Clause. It Won't Hold
The reflex, when a procurement team notices the change, is to add a short AI addendum. Something that says the agency may use AI tools, that the client consents, and that the agency remains responsible for the output.
It feels tidy. It signs cleanly. It solves almost none of the real risks.
A one-paragraph consent doesn't say which decisions the agent is allowed to make on its own. It doesn't name the threshold at which a human has to approve an action before it goes live. It doesn't require the agency to keep logs you can read when something goes wrong. And it leaves the liability cap at the usual fees-paid number, which was calibrated for a world where the worst-case screw-up was a missed deadline, not a mass-email campaign sent to the wrong list or a bidding agent that burns through a quarter's ad budget in an afternoon.
Law firms watching these deals have been direct about the mismatch. A Clifford Chance analysis of agentic AI contracts earlier in 2026 flagged that standard supplier paper often disclaims the agent's acts entirely and caps recovery at fees already paid. In a bad month, that number won't begin to cover what an autonomous system can do to a customer base or a budget.
The Liability Cap Is Where the Negotiation Actually Lives
Everything above is table stakes. The number that decides whether any of it matters is the limitation of liability. Fees-paid caps made sense when the worst thing a consultant could do in a month was produce a bad deck. They don't hold up when an autonomous system is acting at machine speed inside your accounts.
A recent Mayer Brown analysis of agentic AI deals walks through why the standard allocation needs to be reopened: liability caps, indemnities, and IP assignment clauses were all written assuming human authorship and human judgment at the point of execution. Push for an enhanced cap for AI-driven incidents specifically. Push for carve-outs from the consequential-damages exclusion where regulatory fines or data loss are foreseeable. And insist that the agency's IP indemnity covers output generated by its agents, not only work drafted by its staff.
Sign the Contract That Matches the Work Being Done
The agencies doing this well are not hiding the agents. They'll tell you which workflows are automated, show you the oversight model, and sit through the contract edits that go with it. The ones to worry about swapped the staff for software and left the paperwork alone.
Before you renew, read the agreement as if the next outreach email were going to be sent by a system acting on its own. If the contract can't tell you who authorized it, who approved it, and who pays when it goes wrong, it isn't ready for the work.