Over the past six months we have shipped more than fifty releases of Gydr. Looking back across them, the same handful of beliefs about customer-facing AI kept deciding the hard calls: what to build, what to refuse to build, and what to break on purpose. We think they are worth writing down, both so our customers know what to expect from us and so we can be held to them.
1. The business is the authority
An AI assistant that answers for a business speaks with that business's voice. It should say what the business knows, in the business's terms, and nothing else. Your knowledge base, your catalogue and your policies are the source of truth. The model's job is to find the right part, understand the question, and present the answer well. It may summarise, translate and explain. It may not add.
This one belief explains most of our work on answer quality: why an empty product description produces a name and a price rather than a paragraph, why prices are checked before they are quoted, and why "I don't have that" is a correct answer.
2. Honesty beats helpfulness when they conflict
A chatbot that confidently tells a customer something untrue does more damage than one that admits it does not know. We tune for the second, deliberately, even when it makes a demo less impressive. We wrote about this in detail.
3. Some sentences should never be improvised
Language models are remarkable at conversation and unreliable at promises. So we draw a line. Where an answer needs judgement, the model writes it. Where a sentence states a fact about the system, we write it ahead of time and translate it into fifty languages. "Someone from the team is joining." "This message could not be delivered." "Please sign in to see your orders." Those sentences have to be exactly true every time, so they are never left to chance.
4. A person, when it is promised, must be real
The worst moment in customer service is waiting for someone who is not coming. We would rather not offer a person than offer one who cannot be reached. When a handover cannot happen, the customer is told and given another way to get in touch. The full story is here.
5. Your customers' data stays with you
Every business on Gydr is its own world. No visitor, conversation, memory or lead ever crosses from one account to another, whatever the circumstances. That holds even when the same person talks to two businesses from the same browser, and even when both businesses are run by the same partner. Identity is decided by the sign-in a business already trusts, never by what someone types into a chat, and never by an ID a web page chose for itself.
6. Platform rules protect everyone, so we follow them strictly
When a business connects WhatsApp, Messenger, Instagram or Telegram to Gydr, it shares our connection with every other business on that platform. A policy violation would not only affect the business that caused it. It could affect all of them. So we read platform rules in their strictest sense. We decline capabilities that are technically allowed when we cannot guarantee they will be used correctly at scale, by every customer, for replies we did not write.
7. You should see what you pay for
Usage-based pricing only works if usage is legible. Every feature that uses credits says so where you switch it on. Every conversation shows what it cost. A turn that a broken channel could never deliver is not billed. And when an account runs out of credits, the chatbot says it is unavailable instead of pretending to try.
8. One assistant, everywhere
A business should build its assistant once. Website, WhatsApp, Instagram, Telegram and the API are different doors to the same knowledge, the same rules and the same standard of service. Every channel deserves the same care, and none of them deserves a lesser copy.
9. Software should be operable by AI, safely
People increasingly work through AI assistants instead of through screens. The software that will last is software those assistants can operate, under exactly the same permissions, confirmations and audit trail as a person. That is why we built Console MCP, and why it asks before it changes anything your customers will see.
Why this matters to us
Most of the world's businesses are small. They answer customers in more than one language, on whichever app the customer prefers, often outside working hours, often with a team of one or two. They cannot afford a chatbot that embarrasses them, and they have no engineers to watch it.
Those are the businesses we build Gydr for. AI can give them the reach of a much larger company, but only if it behaves like a trustworthy member of their team: one that knows the business, speaks the customer's language, admits what it does not know, and hands over gracefully when it should.
That is the standard we hold ourselves to. When we fall short of it, we want to hear about it.
Frequently asked questions
What principles guide Gydr's customer-facing AI?
The business is the authority; honesty beats helpfulness when they conflict; fixed sentences are never improvised; a promised person must be real; customer data stays with its business; platform rules are read strictly; costs are visible; one assistant serves every channel; and software should be safely operable by AI.
Does customer data ever cross between Gydr accounts?
No. No visitor, conversation, memory or lead crosses from one account to another, even when the same person talks to two businesses from one browser.