Every week brings a new reason to sign an AI agent contract quickly, and a new reason to regret signing carelessly. This guide collects what our reporting keeps finding in one place: the questions that decide, eighteen months from now, whether your agentic AI platform is an asset or a hostage situation.
Why lock-in is the central risk
Agents differ from classic software in one structural way: they accumulate. Memory, workflow data, integrations and rebuilt processes all deepen with use, which is why we have argued that lock-in is the real fight as agents leave the pilot phase. The pressure to commit is engineered: model providers are subsidizing adoption with free compute, defaults are being upgraded to agent-grade models as with Claude Sonnet 5, and the vendors themselves are consolidating, as ServiceNow’s move for ai.work shows.
The seven questions to ask before signing
1. Where does agent memory live, and in what format does it leave with you? 2. Can work route across model providers, or is the platform welded to one lab? 3. What is the data egress path, priced and tested, not promised? 4. How many systems does each agent touch? Every integration is a strand of lock-in; count them like costs. 5. What happens at credit or discount expiry? Model the renewal price as the real price. 6. Who owns improvements the agent learns from your workflows? 7. What is the exit clause, in days and dollars, if the vendor is acquired?
The margin test
One diligence shortcut from our margin analysis: ask the vendor how their gross margin works. A platform renting all its intelligence from a foundation model at thin margins has every incentive to deepen your lock-in, because switching costs are its only moat. A vendor with real economics can afford portability.
The bottom line
Adopt fast, sign carefully. The winners of this cycle will be the buyers who negotiated exit rights while vendors were still courting them. Leverage never improves after the signature. The stakes of getting this right are rising fast in regulated industries, where legal and compliance agents are now drawing institutional investment from their own future buyers. Identity is the piece most buyers still skip: 67 percent of nonhuman accounts sit unmanaged at the average enterprise, which means an agent can inherit access nobody mapped before you signed. Integration is where the vendors themselves now compete: TCS is staffing thousands of forward-deployed engineers on the bet that making AI work in production, not the model, is the real moat.
Frequently asked questions
What is agentic AI lock-in?
Agentic AI lock-in happens when an AI agent accumulates memory, workflow data and integrations over time, making it progressively more expensive to switch vendors even if a better platform becomes available.
Can I switch AI agent vendors after deployment?
Usually, but the cost rises the longer you wait. Ask about data portability and multi-model routing before signing, not after, since leverage is highest before the contract is signed.
Do free AI credits from providers come with strings attached?
Yes. Subsidized compute and token credits typically embed one provider’s models and data formats into your product, which is a deliberate customer-acquisition strategy, not a discount.
What should a good agentic AI contract include?
Clear data egress terms, the ability to route work across more than one model provider, a defined exit clause with timelines, and pricing that survives the end of any introductory discount.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
