Why does AI do badly at exact facts specifically?

Language models generate plausible text, and a plausible-sounding fact is not the same as a verified one. Without grounding a claim in the actual record, a model can produce a date, a name or a number that reads correctly but was not checked against anything real.

Is scoring with reasons more trustworthy than a bare score?

Yes, in practice. A bare numeric score gives no way to sanity-check it; a score with the specific signals behind it, this contact opened three emails and visited pricing twice, lets a person quickly judge whether the reasoning holds up before acting on it.

What is the actual failure mode when AI output is not reviewed?

Not that the output is obviously wrong, since obviously wrong output gets caught. The failure is confidently wrong output that reads just as polished as correct output, which is why the 'needs review' features are exactly the ones producing text someone acts on directly.

What does AI actually do well in a CRM?

Summaries, classification, drafting, scoring with reasons, and extraction: anything that turns messy text into structure or structure into a first draft. It does badly at exact facts, at deciding to act, and at anything nobody reviews.

app.salescrew.io/inbox

The unified reply inbox with classified threads

The short answer

AI in a CRM does well at tasks that turn unstructured text into structure, or structure into a first draft: summarizing a deal's history, classifying an inbound reply, drafting a follow-up, scoring a lead with the reasons behind the score, and extracting fields from a messy email.
It does badly at producing exact facts without grounding, since a language model generates plausible text and a plausible date or number is not the same as a verified one pulled directly from the record.
It does badly at deciding to act on its own, since that requires judgement about consequences the model has no stake in and often no visibility into beyond the immediate text.
The real failure mode is not obviously bad output, which gets caught quickly. It is confident, polished, wrong output that nobody reviewed, because it reads exactly like the confident, polished, correct output next to it.

The pattern behind what works and what does not

The tasks AI handles well share a structure: there is a clear input, the input is usually text that already exists somewhere (an email thread, a set of notes, a record's timeline), and the output is either a smaller, structured version of that input or a first pass at new text a person will read before it matters. Summarizing shrinks a long history into a shorter one. Classification maps a message to a known category. Extraction pulls specific fields out of a paragraph. Drafting produces a starting point, not a finished, unreviewed artifact.

The tasks it handles badly share a different structure: either the output needs to be exactly correct with no room for a plausible-sounding approximation, like a date or a dollar figure, or the task itself is a decision with consequences, sending a message, changing a deal's stage, that the system has no way to weigh the way a person accountable for the outcome would.

Task by quality today, and whether it needs review

Task	Quality today	Needs review?
Summarizing a deal's history	Generally strong when grounded in the actual record	Spot-check, not every instance
Classifying an inbound reply	Generally strong for common categories	Spot-check; escalate low-confidence cases
Drafting a follow-up message	Strong as a first draft, weaker as a final send	Always, before it sends externally
Scoring a lead with reasons	Reasonably strong when the reasons are shown, not only a number	Spot-check the reasoning, not only the score
Extracting fields from a messy email	Strong for well-defined fields (name, date, amount mentioned)	Spot-check on ambiguous or low-confidence extractions
Deciding whether to act (send, change a stage)	Weak as an unattended decision	Always
Producing an exact fact without grounding	Unreliable; risk of a plausible but invented detail	Always verify against the source record

Making the good tasks actually good in practice

The tasks that work well in principle still need to be grounded properly to work well in practice. A summary built from the actual timeline, with citations back to specific lines, is trustworthy in a way a summary built from a vague recollection of "what this deal is generally about" is not. The difference is not the task, it is whether the model was given the real record to work from or asked to reconstruct it from memory.

Disclosure: SalesCrew is our product, and its AI features are built around this pattern: record summaries and lead scores cite the specific evidence behind them, classification runs rules first with a model fallback rather than a model-only guess, and drafts always land in a review queue before an external send. The tasks AI does well are the ones the product leans on; the ones it does badly at stay behind a person.

The failure mode is confident output nobody reads

A wrong summary or a wrong score is not dangerous because it is obviously wrong. It is dangerous because it looks exactly like every correct one next to it, and the habit of skimming AI output instead of reading it is what lets the wrong ones through.

Questions

Why does AI do badly at exact facts specifically?: Language models generate plausible text, and a plausible-sounding fact is not the same as a verified one. Without grounding a claim in the actual record, a model can produce a date, a name or a number that reads correctly but was not checked against anything real.
Is scoring with reasons more trustworthy than a bare score?: Yes, in practice. A bare numeric score gives no way to sanity-check it; a score with the specific signals behind it, this contact opened three emails and visited pricing twice, lets a person quickly judge whether the reasoning holds up before acting on it.
What is the actual failure mode when AI output is not reviewed?: Not that the output is obviously wrong, since obviously wrong output gets caught. The failure is confidently wrong output that reads just as polished as correct output, which is why the 'needs review' features are exactly the ones producing text someone acts on directly.