AI in the background: better negotiation options, clear human decisions

A customer asks for a substantial discount. There is also a tight deadline, while internal approval is still outstanding. AI may help in this situation by separating conditions, comparing possible responses and highlighting a question you have overlooked. Its usefulness must be tested in the actual setting. Responsibility for a commitment remains with the person authorised to make it.

Define the role before using the tool

A useful assistant needs a clear assignment. Should it organise meeting notes, examine counterarguments or prepare wording options? Anthropic distinguishes predefined workflows from agents that choose their path more dynamically. Starting with a manageable solution makes sense for many tasks; additional complexity needs a clear benefit. The conceptual background is described in Building effective agents.

For negotiations, a practical boundary follows: the assistant helps prepare possible responses and their consequences. That does not give it general permission to send messages, confirm prices or change records in other systems. Preparing an option and acting on it are separate decisions with different consequences.

Separate private thinking from approved statements

Consider a fictional example. A sales representative privately notes that she might be able to seek more flexibility for a longer contract. The customer says predictable costs are particularly important. AI could suggest exploring price and duration together. That must not become a confirmed discount. One possible formulation is: “A longer commitment could create a different framework. I first need to check our internal flexibility.”

Several reviewable options can help: a question about the required certainty, an indicative package for internal assessment or a reasoned refusal. Each option should identify its unresolved conditions. Private thoughts, commercial limits and internal comments must be handled according to the data and approval rules actually in place. An attractive chat interface alone does not demonstrate that these boundaries work.

Approval needs understandable information

A human-in-the-loop mechanism can pause a tool operation so that a person can approve or reject it. The OpenAI Agents SDK documents one such pattern. For a user, the important question is whether the information needed for that decision is visible.

Before an external action, at least the recipient, message, attachments and commitments should be clear. It must also be apparent who has authority to decide. This makes it possible to notice when “potentially” has become “definitely”. Approval should apply to the version that was actually reviewed. If that version changes materially, the earlier approval is no longer a sound basis for proceeding.

Test difficult situations

Try incomplete instructions, conflicting sources and withheld approvals. Check whether the assistant exposes uncertainty and stops when authority is missing. External documents may also contain unwanted instructions. OpenAI describes these prompt-injection risks; outside content must not acquire additional permissions.

Finally, compare the assistance with your existing approach. Are important conditions noticed more often? Are suggestions easier to check? How much rework remains? These observations are more informative than one impressive demonstration.

Vanellence is exploring sales and negotiation as an entry point for this approach. Its illustrative demo presents the idea; it does not demonstrate an approved Teams or CRM integration. The long-term Cognitive Gateway aims to coordinate AI, tools and human expertise according to the situation. Its effectiveness still needs to be demonstrated.