Transactions

There appears to be a fundamental conflict between vendors and users:  if Agentic AI or SI produces inherently and inevitably error-ridden results, should users be holding vendors responsible for such errors?  It would appear that vendors have the better side of this argument.  If the user understands that such errors are inevitable, then human control-side intervention would appear to be the only rational manner in addressing such error – which puts the entire onus on the user for the adverse effects of any such errors.  Nevertheless, there remain a range of AI and/or SI contractual provisions that should be uniquely considered.

  • Covenants

A sampling of sui generis AI/SI covenants are as follows:

  • Data use – clearly identifying which data will be accessible is key.
  • Data retention – clearly identifying when data will be used and retained (if at all) is often overlooked.
  • Model-use parameters – covenants associated with what the parameters of model use are can be critical.
  • Training parameters – how, when and to what degree training on data will occur as well as permute must be addressed.
  • Guardrail creation – what guardrails will be included and the measure of same must be set forth (and this mitigates for the user the liability associated with inherent error).
  • Ownership of IP/data – clear indication of the negotiated result here is important; generally speaking, users should make it understood that all derivations are owned by the user and there should be no licensed use other than to perform the task within the parameters of the contract.
  • Operational timeframes – given vagaries associated with energy demands and other extraneous attributes, sunset provisions are critical.
  • PII/Confidentiality mandates – external to guardrails will be extra mandates on PII and confidential data treatment.

 

  • Representations/Warranties

Related to some of the covenants are a small sampling of sui generis AI/SI representations and/or warranties:

  • Guardrail viability – while there will be covenants here, scope representations and warranties are critical as well.
  • No unauthorized model deployment/updated capabilities – once again, scope and status considerations must be set forth in writing.
  • No unlicensed IP or data usage – while imagining associated covenants, best knowledge provisions are important in this area.
  • Known error parameters – this probably cannot be addressed merely by covenants and while inevitable operational error will occur, some level of identifiable areas for directed human intervention should be discerned ahead of time.
  • Prior claims or lack thereof – it is helpful to know historical attributes as they will likely be repeated.

 

  • Indemnifications

There will be real action in this regarding in negotiations.  Hold harmless parameters within the uncertain environment of AI/SI requires intense examination.  In that respect:

  • Vendor will want to limit or only make indemnifications with limitations and not tied to underlying covenants, representations or warranties.
  • User will want indemnifications associated with a breach of any covenant, representation or warranty.
  • Query whether insurance provisions will back any indemnification.

 

  • Insurability
    • Determine whether there is any available coverage.
    • Additional insured coverage will be interesting as well.