Legal teams deploying B2B AI agents in contract negotiations will need purpose-built playbooks designed to give machines executable, unambiguous instructions.
Most existing contracting playbooks are written for experienced human lawyers who can interpret vague guidance and apply professional judgment accumulated over years of practice.
The problem, as legal technology expert Olga V. Mack explains, is that “use judgment” is simply not an executable instruction for an AI agent operating in a live negotiation.
Companies will need playbooks designed for machines as well as people, which means making contracting logic explicit enough that an agent can follow it, recognise its limits, and know when to stop.
A conventional playbook might identify a preferred liability cap, offer one or two fallback positions, and instruct the negotiator to escalate anything materially less favourable — but what counts as material is rarely defined precisely.
The challenge deepens when considering the relationships between contract terms, since provisions rarely operate in isolation from one another within a single agreement.
A company might accept a different indemnity position if the liability cap changes, or permit broader data use if the data is sufficiently deidentified, creating compounding risks an agent could miss.
An agent following clause-by-clause instructions could produce an agreement in which every individual term appears acceptable while the combined risk is not, a significant and potentially costly danger.
Perhaps the hardest challenge of all is capturing what Mack describes as tacit exceptions — unwritten rules that experienced legal teams apply consistently but have never formally documented anywhere.
Rules such as “we normally accept that language, except for strategic vendors” work only because the same experienced people apply them repeatedly, a fragile arrangement that AI deployment will quickly expose.
Mack, who is CEO of TermScout and teaches at Berkeley Law, argues this is not necessarily bad news for legal departments wrestling with the challenge.
Preparing playbooks for agents may finally force companies to improve playbooks for everyone, surfacing vague standards, conflicting policies, missing approval paths, and undocumented exceptions that already create inconsistent negotiations.
Human lawyers currently compensate for those weaknesses through experience, memory, and internal relationships, but new team members and outside counsel may struggle just as much as any AI agent would.
The goal should not be to encode every possible negotiation outcome, since contracts are too contextual and genuine judgment cannot be reduced to a very large decision tree.
The better objective is to define clearly where the organisation has made a repeatable decision and where it has not, allowing agents to handle the former while humans retain authority over the latter.
Novel, consequential, or highly contextual decisions should remain visible as such and move to a person with the appropriate authority, preserving human oversight where it matters most.