AI in support is exactly as good as the context you gave it. The model doesn't know your product on its own — it only knows what you told it. So the single biggest factor in AI draft quality isn't the model itself but the product description you fill in once. A good description turns AI into a useful assistant; a bad one turns it into a source of confident fiction.
Why the product description decides everything
When the model lacks facts, it doesn't stay silent — it fills in the gaps. This is called hallucination, and in support it's especially dangerous: the customer gets a confident but wrong answer about billing, refunds, or what the product can do. An agent in a hurry may not notice the error before sending it along.
The only way to avoid it is to give the model enough accurate context that it has no need to invent. An AI draft is built from two sources: the product description (general facts) and the history of the specific conversation (this customer's context). If the first source is empty or vague, the model compensates with guesses — and the more confident the draft sounds, the harder it is for an agent to notice it isn't grounded in fact.
So the product description isn't a checkbox to tick in settings — it's effectively the instructions AI follows every day when it talks to your customers in your voice.
What to put in your product description
A good description for AI isn't marketing copy — it's a working reference, the kind you'd write for a new support hire on their first day. Include:
- What the product is in plain words: what it does and for whom. Avoid generic phrases like "innovative solution" — they carry no facts.
- Key scenarios: how to get started, how to pay, how to set up, what to do in common situations. This covers most day-to-day questions.
- Answers to frequent questions — the exact ones that recur every day. If agents answer the same question every week, it belongs in the description.
- Limits and exceptions: what the product can't do, what the constraints are — so the model says so honestly instead of inventing functionality that doesn't exist.
- Tone of voice: how you talk to customers — formal or casual, playful or reserved. On a consistent tone, see the article on support tone.
Details that are easy to forget
Beyond the basics, write down the smaller but frequent details: delivery or processing times, how to cancel a subscription, what to do if a customer forgot their password, which channels you support. These "small" details are usually what makes an AI draft read as vague instead of concrete.
If you already have a set of standard phrasing agents use by hand, that's a great source to draw from too — see the article on support reply templates.
Bad vs good description: an example
Bad: "We're a modern service that helps businesses grow. Our team is always happy to help." There's not a single fact here the model can build an answer from. If a customer asks about refunds, AI will either deflect with a vague phrase or invent terms — both bad outcomes.
Good: "We're a support service. Billing is monthly, by card, via the Pricing page. Free trial is 14 days, no card. Refunds within 7 days of payment. Integrations: Telegram, widget, API." From a description like this the model can compose an accurate answer to almost any common question, and the agent just needs to check it and hit Tab.
The difference between these two examples isn't style — it's the number of checkable facts. You can mentally test a good description by asking "where did the model get this from?" — and every fact should have an answer in the description text.
A simple starter template
If you're not sure where to start, copy the structure below and fill it in with your product's facts:
- What we do and for whom — one or two sentences.
- How to get started: signup, setup, first step.
- Billing: plans, payment method, free trial.
- Refunds and cancellations: terms, timeframes.
- What the product can't do: an honest list of limits.
- Tone: how we talk to customers.
That's enough for a first version of the description — after that, you only need to add to it.
Common mistakes when writing a description
Even with good intentions, it's easy to write a description that looks complete but doesn't work. A few common traps:
- Overly generic wording. "Flexible pricing" instead of an actual number or condition — the model can't give a concrete answer based on a vague word.
- Stale facts. A description written a year ago often contains terms that have since changed. AI will confidently repeat them until someone fixes the text.
- No stated limits. If you don't write down what the product can't do, the model sometimes confirms features that don't exist, because nothing in the description rules them out.
- English-only description when customers write in several languages. The model drafts in the customer's language, but it relies on the facts in the description — so make sure the key facts are clear regardless of language. More on this in the article on multilingual support with AI.
Keep the description alive
The description isn't a "write it and forget it" document. Every time you catch AI being inaccurate, or a new frequent question appears, add the relevant fact to the description. The simplest approach is to keep the description open nearby and add a line to it after every awkward conversation.
Over time it becomes a concentrate of knowledge about your product — and the drafts get steadily more accurate. This is especially noticeable in the first few weeks of use: the more real conversations pass through the system, the more obvious it becomes which facts are missing from the description.
Once a month or two, it's worth rereading the whole description: remove outdated points, check pricing and terms against the current state of things, and add new scenarios that appeared alongside new product features.
How it's set up in Duck Float
In Duck Float the product description is filled in once in settings, and it's exactly what the model leans on when it prepares a draft reply together with the history of the specific conversation. The more detailed and accurate the description, the better the drafts you only need to review and send: the draft appears right in the reply field, the agent accepts it with the Tab key, edits it as needed, or writes their own reply from scratch. Nothing is ever sent to the customer automatically.
How the drafts themselves work — in the article on AI drafts. Duck Float is free during the beta; want to try it — message us through the widget on this site.
