Why an AI assistant must refuse to guess
An AI chatbot that invents plausible numbers is worse than having no assistant. Why constraint by architecture matters more than instruction.
Danny Starr
Co-founder, Backline · 16 August 2026 · 2 min read
In short
- An AI that guesses is worse than no AI. The moment it invents a plausible number, you cannot trust any answer it gives.
- An AI that says 'I don't know' is more useful than one that guesses. A manager acts on known gaps; they get misled by confident fiction.
- Disagreement between surfaces is worse than missing a feature. If the dashboard says 214k and the assistant says 219k, the manager distrusts everything they see.
- Instructions alone will fail. 'Please be careful' and 'Only answer from connected data' are hopes, not guarantees. Architecture has to enforce it.
- The real cost of a guessing AI is destroyed credibility. Once a manager catches it inventing, every answer becomes suspect.
An AI assistant that can guess is dangerous. It will confidently state a monthly listener count that sounds plausible, someone will email it to a label, and the moment it's wrong the label stops trusting any number from your company.
A useful assistant refuses to guess. When it does not have an answer, it says so. This is not a nice-to-have. It is the difference between a tool and a liability.
Why guessing destroys credibility
The moment an AI invents a number, it has broken something that is hard to repair. A manager will test it on something they already know the answer to, catch it making it up, and from then on every response is suspect.
An AI that says "I don't have ticket data for that show" is more useful than one that invents a number. The gap is actionable. The fiction is misleading.
The gap-vs-guess problem
An AI trained on general knowledge will produce plausible-sounding answers about music industry metrics. A monthly listener count. An estimate of tour revenue. None of it real, all of it confident.
The hard part is not training the AI to be careful. It is making it impossible for the AI to know numbers except from actual data. If the only numbers in its context come from real sources, it cannot invent. The refusal is not a choice the AI makes, it is structural.
Why answers diverge
The most damaging failure is not hallucination. It is disagreement between two surfaces that are supposed to be showing the same data.
A manager sees 214,000 streams on the dashboard. They ask the assistant the same question and get 219,000. Both numbers came from real queries. One excludes a data source the other included. One corrects for reporting lag, the other doesn't. Now the manager trusts neither number.
This happens when two parts of a system build answers independently. They use slightly different logic and nobody notices until a customer sees the contradiction.
What honest answers look like
Admitting gaps. "Two of the six shows have no ticket data uploaded, so I can only tell you about four." Naming the gap is more useful than quietly working around it.
Refusing attribution. No streaming service tells you which campaign drove a play. An assistant that claims a campaign generated 50,000 streams is inferring, and it should say so. Or refuse to claim it at all.
Labeling uncertainty. An estimate is not a fact. If a figure is calculated from partial data or a shortened window, that matters.
Refusing ambiguity. When a question could mean multiple things - which ad account? which time period? - the honest answer is to ask. Not to guess and hope.
What it is useful for
Triage and discovery, not deep analysis. Which of my eight projects moved unexpectedly. Which shows are behind pace. What is new since last month.
The heavy analysis is still yours. The assistant's job is to save you opening eight dashboards to find the two that need attention.
A tool refusing rather than guessing
IllustrativeHow much did we spend on ads for Vela Court last month?
Data readget_ads_performance
This project has three ad accounts connected, so I need to know which you mean before giving a spend figure.
Meta, account ending 4471, labelled Main
Meta, account ending 9032, labelled Legacy 2024
Google Ads, account ending 118, labelled Label partner
Tell me which accounts to include, or say all of them, and I will pull the numbers. Each account reports in its own currency, so a combined figure will need a conversion rate that I will state.
Illustrative example using a fictional project. The assistant is not being cautious by choice: with more than one account and no selection, the tool returns the account list and no numbers, so a combined total is not available to guess with.
Common questions
- What is the most dangerous thing an AI assistant can do?
- Invent a number that sounds plausible. A manager will test it, catch it making something up, and never trust the system again. An AI that says 'I don't have that data' is worth more than one that guesses confidently.
- Why is disagreement between the dashboard and the AI worse than missing data?
- Because missing data is actionable and disagreement is corrosive. A manager learns to work around missing data. But if the dashboard says 214k and the AI says 219k, the manager stops trusting both.
- Can an AI tell you what caused a stream increase?
- Not reliably. No streaming service reports why a play happened. An honest AI describes what signals moved together and in what sequence. An AI that claims a campaign caused X streams is guessing, and it should refuse to do so.
Sources
- 1Anthropic, Reviewed August 2026. Tool use with Claude
- 2Model Context Protocol, Reviewed August 2026. Introduction to MCP
Danny Starr
Co-founder, Backline
Danny Starr is a co-founder of Backline and builds the platform. He writes about the data engineering behind music analytics: ingestion, identity, honesty in charts, and the AI layer on top of it.
Backline does this for the projects you run
Streaming, audience, social, advertising, website, search, ticketing and press data in one dashboard per project, with an AI assistant that answers questions about your own connected data. Invite-only.
Keep reading
AI and automation
Tool design for an analytics agent
The quality of an AI assistant over your data is decided by the tools, not the model. Nine rules learned from building twelve of them over music data sources.
Danny Starr · 3 min read
AI and automation
Why analytics data should be reachable by any AI client
A platform that only works with its own AI assistant is limiting. Why exposing analytics to external AI clients matters, and what the risk model is.
Danny Starr · 2 min read
AI and automation
Scheduled reports without a data team
A recurring report written from a project's own connected data, sent to whoever needs it. What makes one worth reading, and the scheduling details that decide whether it arrives on time.
Danny Starr · 3 min read
