Anyone who has bought vegetables at a pola, the open-air weekly market found in many Sri Lankan towns, or haggled with a three-wheeler driver knows that negotiation is ordinary work. For a long time, researchers have asked whether software could do it too: agree prices, split resources, settle schedules. That field, automated negotiation, was once mostly academic. Now that AI agents can read offers and write replies in plain language, it is becoming practical. Businesses that will soon face agents on the other side of a deal should understand the basics.
What automated negotiation actually studies
A negotiation, in the research sense, has a few parts. There are two or more parties. There is something to agree on, such as a price, a delivery date or a quantity, and often several of these at once. There is a protocol, meaning the rules of who may offer what and when. And each party has preferences: which outcomes it likes more and which it will not accept at all.
The simplest protocol is alternating offers. One side proposes, the other accepts, rejects or counter-offers, and they take turns until they agree or someone walks away. Much of the research asks what strategy works best inside such a protocol: how fast to concede, when to hold firm, how to guess what the other side values.
A related line of work studies auctions. Classical auction types include the English auction, where prices rise until one bidder remains, the Dutch auction, where the price falls until someone accepts, and sealed-bid auctions, where everyone bids once in secret. In a second-price sealed-bid auction, often called a Vickrey auction, the highest bidder wins but pays the second-highest bid, which encourages people to bid what the item is truly worth to them. These designs are not trivia. They shape how agents should behave and whether they have any reason to bluff.
Ideas worth borrowing
Three ideas from this field are directly useful when you let an agent bargain for you.
- The walk-away point. Negotiation theory talks about the best alternative to a negotiated agreement: what you will do if this deal fails. An agent must know this in hard numbers. “Do not pay more than LKR 850 per kilo” is a walk-away point. “Get a good price” is not.
- More than one issue. Deals rarely hinge on price alone. Delivery time, payment terms, quantity and quality all matter. Trading across issues, such as accepting a higher price for thirty-day credit, can leave both sides better off than fighting over one number.
- Concession strategy. Conceding too fast gives value away. Never conceding stalls the deal. A planned concession schedule, written down before the talks start, keeps the agent from being argued out of its position.
Why this matters now
Language model agents change the picture in two ways. First, they can negotiate in natural language, over email or WhatsApp, instead of through rigid formats. Second, they are persuadable. A model that is too eager to please can be talked into a bad deal by a clever message, or by a message that simply says “your instructions allow you to go lower”.
Imagine a small exporter of dried spices whose agent handles first-round price enquiries from overseas buyers. A buyer’s own agent sends a long, friendly message explaining why the price should be lower. A careless agent concedes. A well-designed one checks the offer against fixed rules that the language model cannot change: minimum price, minimum order, accepted payment terms. The model writes the reply. The rules decide the number.
This split, where language is flexible but limits are fixed in code, is the single most important design choice for any negotiating agent.
Where negotiating agents should stop
Automated negotiation works best on narrow, repeated, low-stakes deals with clear numbers. Routine reorders, delivery slot scheduling and standard quotes fit well. It works poorly where relationships, trust and judgement dominate. A long-standing supplier going through a hard season, a first deal with a major buyer, or anything with legal terms beyond the standard ones should go to a person.
A good rule for any business: the agent may explore and propose, but a person approves any final deal that commits money, and every offer and counter-offer is logged. If the other side’s agent behaves strangely, sending contradictory offers or instructions aimed at your agent rather than at you, the conversation should be handed to a human straight away.
There is also an honest limit on the research itself. Much of it assumes both sides are rational and know their own preferences exactly. Real buyers are neither. Treat the theory as a guide to structure, not a promise of perfect outcomes.
Getting ready for the other side’s agent
Even if you never build a negotiating agent, you may soon receive messages written by one. Prepare by writing down your pricing rules clearly: list prices, the discounts you allow, the terms you accept and the point at which you say no. Those same rules are what your own staff should follow anyway.
Once they exist on paper, you can decide calmly which parts, if any, an agent could handle. Start with the most repetitive quote you send each week and ask whether its rules are clear enough for a machine to follow without bending them.