Maxims for Talking Machines

The failures we have seen

Every way our AI can fail a customer, we have probably seen it. And because it answers on behalf of the business, before anyone there has even seen the message, every one of those failures is the business’s failure. It invents an appointment that was never offered. It quotes a price from nowhere. It reads a phone number back as one long number, “five billion, five hundred fifty-five million…”, instead of something the caller can dial.

That list feels endless when you are building the test suite for it. It is not. Every conversational failure we have seen comes down to a version of one of four rules a philosopher worked out in the 1960s.

One principle, four rules

In the 1960s, the philosopher Paul Grice, whose work on meaning and conversation shaped modern linguistics, set out to explain how people understand each other when so much goes unsaid. His answer was that conversation runs on an unspoken assumption that both sides are cooperating. From that shared assumption fall four rules a good contributor follows. He called them the maxims of quantity, quality, relation, and manner.

• Quantity: Provide as much information as needed, but no more.
• Quality: Say only what you have good reason to believe is true.
• Relation: Stay relevant to the topic at hand.
• Manner: Be clear, brief, and orderly; avoid obscurity and ambiguity.

Anyone who has iterated on chatbot prompts will recognize these in some form. When you spend enough time prompt-engineering conversational AI agents, you start to notice you are writing different versions of the same behavioral guidance over and over.

The tests we run are each special cases of one of the four. Naming them this way is not academic decoration. It tells us where a failure lives and what kind of fix it needs, and those turn out to be different for each maxim.

The taxonomy covers conversational behavior: what the agent says and how it says it. It does not cover infrastructure failures, a webhook that never fires, a CRM push that times out, a calendar sync that drops a slot. Those are system problems, not conversational ones, and they need different diagnostics. The maxims are for the part where the agent opens its mouth.

Quantity

Left alone, a model over-explains. It restates your question back to you, adds a caveat nobody asked for, and turns a one-line answer into a paragraph. Contractors do not text like that. People do not text like that. A customer feels the shift even if they cannot name it.

The data backs this up. Across tens of thousands of first messages in our system, the longest ones, averaging over 600 characters, get roughly half the reply rate of messages under 100 characters. When the agent over-explains, customers stop responding. Not because the information is wrong, but because it feels like talking to a machine.

In text this is mildly annoying. On a phone call it is a real problem. The caller cannot skim past it, and by the time the agent stops talking they have lost the thread. So a lot of our tuning is pulling the model back. Shorter, plainer, one thing at a time. Quantity is the maxim you fix with restraint.

But restraint has a wrinkle. The hardest case is not the model saying too much, it is the model saying something when the right answer is “I don’t know.” Models are trained to be helpful, which biases toward answering. When a customer asks something the agent has no grounds for, the cooperative move is to say so and defer to a human. But the training pressure to be useful pushes the other way, and that is where a quantity failure becomes a quality failure. The model fills silence with a guess.

Quality

This is the one with real money on it. A made-up arrival window is a promise the business now has to keep. A made-up price is a number the customer will hold them to. When a model offers a Tuesday 3 PM slot that was never on the table, it is not malfunctioning in some exotic way. It is saying something it has no grounds for.

The fix is not telling the model to “be truthful.” The fix is defining what counts as grounds for a claim, which for us means the business’s own instructions, the lead’s real details, a tool result such as a calendar lookup, or the knowledge base, and then forbidding the model to assert anything outside that. Provide the warrant, then hold the model to it. This is why our grounding tests catch prompt problems as often as model problems. If the instructions are a tangle, the model runs out of grounds and fills the gap on its own.

It also matters to tell the model that not knowing is allowed. When in doubt, use the grounding tools available, and if there is still a gap, say so and defer to a human. This is the same thread from the quantity section: the model’s bias toward saying something rather than nothing is what makes silence feel like failure to it, even when silence is the right answer.

Relation

Customers ask for things the business does not do. In our data, roughly one in twelve leads turns out to be mistargeted, the customer is asking for a service the business does not provide. A roofing company gets a call about plumbing. A plumber gets asked about gutters. The agent has to recognize the mismatch and redirect, not pivot into explaining what the business does do as if that were the question.

People also try to pull the agent off task, sometimes by accident and sometimes on purpose. When people pressure-test a phone agent, they try to get it to go off course, read back its system prompt, or say something inappropriate. Normal calls rarely go that far, but they hit more subtle versions of the same thing.

Relevance is not just staying on topic, it is staying on the point of the conversation. If a customer asks where the business is located, the answer is the address, not three facts about the city. If they ask whether you handle slab leaks, the answer is yes or no and a next step, not a tour of every service on offer. In a small but telling share of our first messages, the agent lists services the customer never asked about. It is a natural impulse for a model trained on marketing copy. But it is a relation failure. Relation is the maxim of knowing what is on the table and staying there.

Manner

Manner is about the way things are said, and Grice packed a few things into it: do not be obscure, do not be ambiguous, do not be convoluted, keep it orderly. In text this means things like avoiding weird punctuation or formatting artifacts, for example a phone number rendered with link markup instead of plain digits the customer can read. On a voice call the model has to say it out loud, and that is a whole new way to get it wrong. A phone number is ten digits with a shape. Read it as one number and the caller hears “five billion, five hundred fifty-five million…” trailing off into nonsense, instead of a number they can dial.

Phone numbers, prices, dates, email addresses, URLs, they all have their challenges. Voice adds a layer of choices that text never had, and every one is a place to be misheard. A dollar figure read as “one five zero zero” instead of “one thousand five hundred.” A date collapsed into something the caller cannot parse. Text and voice are different channels with different conventions, so the same reply can be clear on the page and a mess out loud. Each modality has its own failure modes, and manner has to be tuned to the one you are in.

Why the framing earns its keep

We could call these things “grounding,” “scope,” and “readback,” and often we do. Those are the words on some test cases. But the maxims give us something the flat list did not. They sort the failures by how to fix them, and the fixes are not the same kind of thing.

Quality cannot be solved with better wording. A model cannot tell a true statement from a good guess, so no amount of tone or instruction makes it more honest. It needs real grounds and a hard rule confining it to them. That is an architecture problem, not a prompting one. Relation is a scope boundary, and part of it is a safety concern, since the line between on-task and off-task is the same line a pressure-tester is trying to cross. Quantity and manner are the tunable ones, movable with the prompt or by switching models, and for voice, manner is a mix of the inference model choosing the words and the synthesizer deciding how they sound.

The maxims do more than organize our test cases. They give the team a shared language for what we are aiming at: agents that stay grounded, stay in scope, and hold a purposeful, human-like conversation. When a call goes wrong, “that is a quantity failure” or “that is a manner problem” gets everyone looking in the same place.

Underneath all four is Grice’s starting point: a good participant is one who is cooperating. That is what an agent answering for a business has to be. The maxims are just how we spell that out, and how we check it held.