← The journalPLATFORM / AI

Asking your business questions in plain language

What it takes for a plain-language answer from your ERP to be trustworthy, where it goes wrong, and how Sundial is approaching it.

Road and trees with directional camera-motion blur.PLATFORM / SUNDIAL
Photography: Ben Wicks / Unsplash ↗ · Sundial motion treatment
THE PRACTICAL TAKEAWAY

A plain-language answer is only useful if it shows how it was calculated, respects permissions and links to the records. Treat any answer that cannot do this as a guess.

Why ask instead of building a report

Most business questions are one-offs. "How much did we buy from this supplier last quarter?" "Which items have not moved in six months?" "Who are our ten biggest customers in the west region this year?" Each is easy to answer if you know which report to open and which filters to set. Many people do not, so they ask someone in finance or operations, who stops what they are doing to run it.

Plain-language questions aim to remove that step. You type the question; the system works out which records to look at, applies the filters, and returns the answer. When it works, it saves time for both the person asking and the person who used to answer.

What makes an answer trustworthy

A language model can produce a fluent answer to almost any question. Fluency is not the same as correctness. For an answer about your business to be trusted, it needs a few properties:

  • It shows its working. Which records were used, which date range, which filters, and which definition of each term.
  • The numbers come from the ledger. The model decides what to query; the ERP does the arithmetic.
  • It links to the records. You can open the invoices or stock entries behind the figure and check.
  • It respects permissions. Someone who cannot see payroll cannot get payroll figures by asking cleverly.
  • It says when it does not know. If the question is ambiguous or the data is missing, it asks or says so rather than guessing.

Where it goes wrong

The most common failure is ambiguity. "Sales last year" could mean the calendar year or a fiscal year that ends in June or September. It could mean invoiced or ordered, gross or net of returns, with or without tax. A careless system picks one meaning silently. A careful one states its assumption or asks.

The second failure is missing context. If a credit note was entered as a journal entry instead of a credit note, a question about returns will miss it. The answer is only as good as the consistency of the data underneath.

The third is overreach: asking "why" questions. "Why did margins fall in October?" invites a plausible story. The system can show which items, customers or costs changed; the explanation still needs a person who knows what happened.

Questions that work well, and ones that do not

Questions that work well are specific and factual: totals, lists, comparisons between periods, rankings, and lookups of a particular document or balance. Questions that work badly are open-ended, predictive or about intent: "Should we drop this product?" or "Will this customer pay?" Those deserve analysis by people, using the factual answers as input.

It also helps to be precise in the question itself. "Net sales by customer for January to June this year, excluding sales tax" will get a better answer than "How are sales going?" The more of the definition you supply, the less the system has to assume. Over time, a team that asks precise questions also tends to agree on shared definitions, which improves its reports as well.

A good habit is to ask the factual question first, check the answer against the linked records, and then do the judgment yourself.

What Sundial is building

Asking the business a question in plain language is part of what Sundial is designed for. It is in active development, not a finished feature, and we are building it around the principles above: answers grounded in the ERP's own records and calculations, with the filters and definitions shown, links back to the documents, and the same permissions the person asking already has.

Because Sundial runs on one system of record, there is a single place to ask. Sales, stock, purchases, production and accounts sit on the same ledgers, so a question that crosses them does not depend on stitching together exports from separate tools.

We will say clearly what it handles well and where it falls short as it reaches customers. In the meantime, the groundwork that makes it useful is the same groundwork that makes reports reliable: clean masters, consistent documents and agreed definitions.