Prompting your CRM in plain English

Natural-language querying was a demo feature for about six years. You would type a question, the system would produce a chart that answered a different question, and everybody would agree it was promising.

It works now, mostly, and the reason is unglamorous: the models got good enough to translate a vague human question into a precise database query and then admit when they could not. That second part is what changed. A system that says "I do not have a field for that" is usable. A system that guesses is worse than a spreadsheet.

Here are the five I actually ask, every morning, in the order I ask them.

The five

"What changed on my accounts since Friday?" This is the one that replaced twenty minutes of scrolling. Not new leads, changes: a stage moved, a reply landed, a signal fired on a company in my pipeline, somebody I called last month posted a new role. The value is that it is a diff rather than a list, and a diff is short enough to read before coffee.

"Which deals have not moved in fourteen days?" Staleness is the cheapest thing to detect and the most consistently ignored, because a stalled deal looks identical to a healthy one in every pipeline view ever designed. Both are just a card sitting in a column. Asking for the ones that have stopped moving surfaces the four that need a decision today, and the decision is usually to kill one.

"Who did I promise something to and not deliver?" This is the highest-value question of the five and the most uncomfortable. Follow-through rate is most of what trust is in a commercial relationship, and the failures are almost never deliberate. They are things said on a Thursday call that got buried by Friday. A system that can read call notes and surface unmet commitments is doing something no human process reliably does.

"Which of my open deals is single-threaded?" One relationship is a coin flip against any organizational disturbance. Asking this weekly turns multithreading from a thing you know you should do into a specific list of four accounts where you should do it this week.

"What did I lose last month, and why?" Monthly rather than daily, and the point is the why. Not the count, which is on a dashboard, but the reasons in aggregate, which are usually in prose notes that nobody reads because reading two hundred notes is a Saturday.

Why these five and not a dashboard

Every one of these questions is technically answerable by an existing report. Most CRMs ship something adjacent to each. Almost nobody uses them, and the reason is worth stating because it explains why the natural-language version is different rather than merely nicer.

A dashboard requires you to know in advance which question you will want, and then requires somebody to build the view. That cost is paid once, which sounds fine, and the result is a fixed set of questions that made sense to somebody in a planning meeting eighteen months ago. Anything outside that set costs a request to operations and a two-week wait, which means it never gets asked.

Asking in plain English removes the setup cost entirely, and the effect is not that you get the same answers faster. It is that you ask different questions, including bad ones, which is how you find the good ones. I did not know "who did I promise something to" was my most valuable question until I could ask it idly.

The honest limits

It cannot answer what is not recorded. This sounds obvious and is the single largest disappointment people hit. If your reps log outcomes and not reasons, no amount of querying produces insight about why deals die, because the information was never captured. The quality ceiling is set by your data hygiene, and a natural-language layer over bad records is a very articulate way of being wrong.

It is bad at questions with ambiguous definitions. "Which deals are at risk" requires somebody to define risk. The system will happily pick a definition and not tell you which one, which is how you end up confidently acting on an arbitrary threshold. Ask precise questions and you get precise answers; ask vague ones and you get a plausible sentence.

And it should not be trusted with anything it can act on. Reading is safe. A system that answers "which deals are dead" and then marks them dead is one bad inference away from deleting your quarter. Query freely, and keep the write path behind a human.

One habit worth forming early: when an answer surprises you, ask the follow-up rather than accepting it. The systems are good at translating a question and mediocre at telling you when your question was slightly wrong, so a number that looks off is often a definition mismatch rather than a finding. Asking "how did you count that" is a legitimate question now, and the answer usually explains the surprise in one sentence.

Ask the diff. Ask what stalled. Ask what you promised. The rest is a scrolling exercise.

— DealArena Team