The price has dropped, but delivery has slipped by a week. The purchasing department\u2019s assistant considers the negotiation a success. Whoever is waiting for those components to start production might disagree. Imagine a negotiation handled by two artificial intelligence systems: one seeks savings, while the other protects the supplier\u2019s margin. Both could achieve their goals and still leave someone unhappy.<\/p>\n
This hypothetical scenario helps explain what changes when software is asked to negotiate on our behalf. Browsing a catalogue means choosing between predefined terms. Negotiation allows those terms to change in exchange for something else: a larger order, an advance payment or a less urgent delivery.<\/p>\n
Here, automation enters delicate territory. A request such as \u201cget the best terms\u201d sounds clear until someone has to decide what \u201cbest\u201d means. Price is only one possible answer.<\/p>\n
An AI assistant can connect to tools that check availability, verify budgets and send proposals. To negotiate, it also needs a mandate and room to manoeuvre: how much to spend, which terms to accept and when to end the exchange.<\/p>\n
The conversation can take place through structured messages, without mimicking a phone call between sales representatives. A proposal contains a quantity, a price and a date; the other party changes one condition; the system assesses whether to counter. A language model can help interpret requests and formulate alternatives, while separate software controls should check figures, availability and authorisations.<\/p>\n
The difficult part is translating preferences into decisions. How much is it worth to receive the goods on Friday rather than Monday? The answer depends on what is happening within the business, often beyond the system handling the negotiation.<\/p>\n
Let\u2019s return to the initial purchase. The supplier offers a lower price in exchange for a later delivery. If the buyer\u2019s agent knows only the budget, the offer may look excellent. If it knows the delay would bring a production line to a halt, the assessment changes.<\/p>\n
The problem starts before the negotiation, with the quality of the available information and the priorities set. An overly narrow objective rewards compromises that no manager would have approved.<\/p>\n
A useful mandate must therefore include non-negotiable conditions. \u201cReduce the cost, provided delivery takes place by Friday\u201d sets a boundary. If no offer meets that requirement, reporting the impasse may be the right outcome.<\/p>\n
Researchers have already studied agents capable of negotiation. In 2022, DeepMind presented experiments involving the game Diplomacy, in which communication and agreements help players coordinate their moves. The study also examined what happens when an agent can abandon its commitments and which mechanisms make doing so less rewarding.<\/p>\n
A game remains a controlled environment: those findings do not establish a system\u2019s reliability in business purchasing. They do, however, show why the ability to reach an agreement should be distinguished from any guarantee that it will be honoured.<\/p>\n
In an operational negotiation, a statement such as \u201cdelivery confirmed\u201d should correspond to verified availability and a confirmation recorded in the supplier\u2019s system. The confidence with which a model writes adds no value to an unsupported promise.<\/p>\n
For a business, it makes sense to separate the authority to explore offers from the authority to approve them. An agent could compare quotations and prepare a proposal, leaving a person to confirm more significant commitments.<\/p>\n
That separation needs to be built into permissions. Writing \u201calways ask for confirmation\u201d in a prompt provides a different level of protection from making it technically impossible to submit an order without approval.<\/p>\n
Confidentiality also enters the negotiation. The system may need to know the maximum budget without being authorised to disclose it to the other party. Accessing information and being allowed to share it are separate permissions and should be managed accordingly.<\/p>\n
The benefits need to be measured across the entire process. Automated negotiation may reduce repetitive exchanges, but it requires integrations, exception handling and checks. If every proposal has to be reviewed from scratch, some of the time saved returns as supervision.<\/p>\n
A sensible first application might involve recurring purchases, comparable terms and modest amounts. Evaluation should cover errors, delays and manual interventions, as well as the discount obtained. There should also be a readable record of offers and accepted terms, so events can be reconstructed without relying on the model\u2019s subsequent account.<\/p>\n
When two AI assistants negotiate, human work therefore begins long before the first offer: deciding what matters, which concessions are acceptable and who can commit the business. At the end of the negotiation, someone will still have to answer for that week\u2019s delay.<\/p>\n","protected":false},"excerpt":{"rendered":"
Possible problems<\/p>\n","protected":false},"author":8,"featured_media":13655,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[39],"tags":[],"class_list":["post-13654","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-39"],"translation":{"provider":"WPGlobus","version":"3.0.0","language":"en","enabled_languages":["it","en"],"languages":{"it":{"title":true,"content":true,"excerpt":true},"en":{"title":true,"content":true,"excerpt":true}}},"yoast_head":"\n