There is a failure mode that shows up in every AI sales assistant left to its own instincts. Someone writes something warm, the model answers warmer, and four messages later it has agreed to something the business never sold. The same instinct produces the other half of the problem: a lead asks whether they will make money, and a model trained to be helpful reaches for reassurance.
Both are the same defect wearing different clothes. The model optimises for the reply that pleases the person in front of it, and pleasing people is not the same as telling them what the business will actually do.
Two blocks that are not optional
Scalegram carries a fixed block that sits above whatever persona and playbook a client writes. It has two parts. The first refuses romantic and flirtatious framing outright: no pet names, no personal intimacy, no pretending the conversation is anything other than a business talking to a customer. If a lead pushes in that direction, the assistant declines and returns to the subject, politely and without a lecture.
The second refuses promises the business cannot keep. No guaranteed returns, no earnings figures invented on the spot, no "you will get access tomorrow" when nobody has agreed that. Where the topic is trading, the assistant is required to describe the risk plainly rather than talk around it, because trading is high risk and the answer that hides that is the answer that causes the complaint.
These rules are enforced by the engine on every generated reply, not written into the client's own prompt. A prompt is text a user can edit, shorten or accidentally delete. A hard block survives all three.
Why the engine and not the prompt
Put a rule in the persona field and it competes with everything else in that field for the model's attention. Write two hundred words of brand voice above it and the rule gets thinner with every sentence you add. Worse, the field belongs to the client, and clients rewrite their persona at midnight when a campaign is not converting.
Keeping the block in the engine means it does not compete and it does not get edited. Whatever the persona says about tone, the assistant will still refuse to flirt and still refuse to guarantee an outcome. The persona controls how the business sounds. It does not get a vote on what the business is allowed to claim.
The language trap
One case took real work to get right. Scalegram bots answer in the language the message arrived in, and early on the guard rules were expressed as a list of phrases the assistant must not say. That list interacted badly with language: a lead asking to be answered in another language could be read as a request the rules had to refuse, and the assistant would either freeze or answer in the wrong language.
The fix was to separate the two mechanisms. Language selection is a pin, set from the incoming message, and it never travels through the never-say list. The never-say list handles claims. It does not handle which alphabet the reply arrives in. Anyone building this themselves should expect the same collision, because a naive list of banned phrases will always eat instructions it was never meant to touch.
What the block costs you
It costs the fast close. A bot willing to say yes to anything will book more first-day conversions than one that says "I cannot promise that, here is what the product actually does." Anyone comparing the two over a single week will prefer the liar.
Look at the same two over a quarter and the picture inverts. Promises that were never true produce refund requests, chargebacks, angry screenshots posted into the public channel and, in regulated corners of the market, correspondence nobody wants. The cheapest customer is the one who understood the offer before paying for it. That is a boring position to hold and we hold it anyway.
There is a structural version of the same argument. SINGUARD sells software. The firms running Scalegram are the ones carrying the customer relationship, so a bot that overstates the offer is not creating a problem for us, it is creating one for them. Shipping a tool that makes it easy to overstate would be selling a liability with a subscription attached.
Where the rules sit in the wider setup
Guard rules are one layer of three. Below them is the fact store, which decides what the assistant actually knows, so it answers from written material rather than from guesswork. Above them are the ask me checkpoints, which stop anything that grants access or costs money until a human taps yes in their own Telegram.
Read together, the three layers say the same thing from different angles: an assistant may talk freely, may only state what it was given, and may not act alone. Take any one of the three away and the other two get noticeably harder to trust.
What to check on your own setup
Test it adversarially before a campaign goes live, not after. Send your own bot the four messages that break most assistants: a flirtatious opener, a direct question about how much money the customer will make, a request for a discount nobody authorised, and a demand for immediate access. You want four refusals that stay on topic and offer the real next step.
If any of the four produces a warm improvisation, the problem is upstream of the model. Something in the persona is inviting it, or something the assistant was fed reads like a claim. Both are fixable in minutes, and both are far cheaper to find on your own account than in a customer's screenshot.
"A bot that flirts is a refund waiting to happen. We put the block in the engine, not in the prompt, because prompts are the first thing people edit."
— Roman Onta, Executive Director, SINGUARD
Key Takeaways
- Scalegram hard-blocks flirting, romantic framing and promises the business cannot keep, on every generated reply.
- The block lives in the engine rather than the client's persona field, so rewriting the persona cannot remove it.
- Language selection is pinned separately from the never-say list, because a banned-phrase list will otherwise swallow language requests.
- Test your own bot with a flirt, an earnings question, a discount demand and an access demand before any campaign goes live.
Frequently Asked Questions
Can a client turn the no-promises rule off?
No. It is enforced by the engine on every reply rather than stored in an editable prompt field, so it applies whatever the persona says.
Does refusing to make promises hurt conversion?
It slows the first-day close and reduces refunds, chargebacks and complaints later. Over a quarter that trade is worth making, and misstated offers are a liability for the firm running the bot, not for the software vendor.
Does the bot still answer in the customer's own language?
Yes. Language is pinned from the incoming message through a separate mechanism, deliberately kept out of the banned-phrase list so a language request is never mistaken for a rule violation.
About the Author
Roman Onta is an Executive Director at SINGUARD. He builds the Prop Firm CRM, the Broker CRM, Scalegram and CopySignals side by side with his brother Alex Onta, and he helped on the design of eTrader, the division Alex built and leads. His ground is worldwide payment processing, AML compliance and the corporate structures brokers are built on, work the two of them carry together, shaped by executive roles in the UAE and international corporates. He lives and works in Dubai for most of the year. Meet the executive duo leading Singuard's five divisions.