Designing an AI Sales Assistant That Indicates When Confirmation Is Needed

The most dangerous moment in a sales meeting is not a tough question. It is a tough question answered confidently with a number that does not exist.

Discount rates, delivery dates, implementation effort, SLA terms — if an AI assistant makes one of these up on the spot and the salesperson repeats it, the company has just made a commitment it never approved. In a hypothetical case, the discrepancy might not surface until delivery — when it is much harder to unwind.

Cuevane, our real-time AI support tool for live online sales meetings, is built around a simple rule: if an answer is not in the company's own registered materials, Cuevane indicates that confirmation is needed rather than inventing figures. This post describes how that rule shapes the design, and where it gets uncomfortable.

What Cuevane does

Cuevane runs alongside a live online sales meeting, in Japanese or English. It transcribes the conversation, retrieves from company materials that were registered in advance — sales decks, price lists, case documents — and shows the salesperson answer candidates with their basis within seconds.

Picture the scenario it is designed for: someone relatively new to sales, asked mid-meeting "so what would that cost for 50 seats?", with no time to dig through a shared drive. The tool shows answer candidates and their basis within seconds, so the salesperson can respond from the materials instead of guessing or stalling. Trust is an important design consideration for any tool used live in front of a customer: if an assistant invents a figure and it gets quoted, the company owns that figure.

Grounding starts with the corpus

One useful way to approach hallucination in this setting is to treat it as a corpus problem before treating it as a model problem.

Cuevane retrieves from pre-registered company materials. That constraint has a practical consequence worth stating plainly: the quality of the answers is capped by the quality of the materials. If the deck and the price list disagree, a retrieval-based assistant has no legitimate way to hide that — and maintaining a single, consistent set of registered materials may also help teams spot documentation problems before a customer does.

The confirmation-needed path is a feature, not a failure

For anything not covered by the materials — a discount outside the list, a feature that is not documented, a delivery date nobody has committed — Cuevane indicates that confirmation is needed rather than generating a plausible figure.

Designing this path matters more than it sounds:

• It should be fast. In our view, the confirmation-needed path should also be fast. A salesperson who learns promptly that the right move is "let me confirm that and get back to you" can say so naturally; one left waiting cannot.
• It has to be unambiguous. A hedged, half-answered response risks being quoted as if it were an answer. The designed behaviour is instead a clear indication that confirmation is needed.
• It has to be the default. When an answer is absent from the materials, the designed behaviour is to indicate that confirmation is needed — not to fill the gap with a plausible number.

This does mean Cuevane will sometimes indicate that confirmation is needed for questions that a general-purpose chatbot would answer fluently. That is the trade we chose. In a sales meeting, a fluent invention is a liability; an indicated gap can instead become a follow-up item.

Showing the basis, not just the answer

Cuevane shows answer candidates with their basis. This is not decoration. Displaying the basis lets the salesperson check whether it actually matches the customer's question before speaking, and if challenged — "where is that written?" — the provenance is already on screen. A confident answer with no visible source asks for blind trust. An answer with its basis attached allows the salesperson to check it.

What it costs to run

Real-time transcription and retrieval have variable costs, and for a support tool the per-meeting cost has to stay boring. In the usage dataset we prepared for our AppSumo submission pack (measured 2026-11-13), the median variable cost came out at approximately ¥23 per 60-minute meeting. Cuevane is priced at ¥9,800 per seat per month (excluding tax), with a 14-day free trial — one seat, no card required, no automatic conversion to paid. A trial matters for a tool like this because the grounding behaviour is hard to evaluate from a landing page: the informative test is watching what it does with your own materials, including how it handles information that is absent from the registered materials.

The same instinct, in a different product

Pentra, our AI SEO media operations platform, applies a related discipline in a different domain: it cross-checks citations and figures in AI-written articles against the original sources, and blocks publication if they do not pass. Different surface, same instinct — in our view, AI-generated citations and figures that do not pass source checks should not ship.

GlobeZenith also offers proofs of concept (PoCs): ¥600,000 fixed for 3–4 weeks, with the first two companies at ¥300,000 if we may publish their company name and industry in a case study. As a small, founder-led company whose representative is the sole employee, we prefer narrow, verifiable scopes, and we limit concurrent development projects to three.

What we would tell another team building this

1. Close the corpus. Decide exactly which documents the system may speak from, and treat anything outside them as unanswerable by design.
2. Make the confirmation-needed path fast and unambiguous. The fallback path is part of the UX, and a clear "confirmation needed" beats a hedged guess every time.
3. Show the source with the answer. Provenance is what turns an AI suggestion into something a person can responsibly say out loud.
4. Expect the tool to expose your documents. If your materials contradict each other, those contradictions may surface in retrieved answer candidates. Fix the materials first; that work pays off with or without AI.

An assistant grounded in your own materials will sometimes feel less impressive than one that answers everything. But when an answer is absent from the registered materials, its confirmation-needed behaviour avoids supplying an invented figure — and in a live sales meeting, that is the point.

GlobeZenith OÜ is an Estonian company (registry code 16979948, founded 3 May 2024), registered at Sakala tn 7-2, 10141 Tallinn. globezenith.net · Contact: [email protected] · Cuevane: cuevane.com