Blog · Topic
What AI can and cannot do with sales data, and why the language model should never do the arithmetic: our tools do it, and check it.
An annotated example of a weekly sales digest built from computed tables: the header that states the period and the checks, ten movements each with the measure, the level, the change, the threshold and the citation, the three lists that follow, untouched, slipped and dormant, the line that says what did not move, and the footer with the definition version. What each element is for, and the version of the same digest that nobody reads.
16 Sept 20263 min readWhat is wrong with the red-amber-green account health field most CRMs carry, what a computed health score is made of, five inputs from the ledger and the CRM with stated weights, how the weights are set from the company's own churn history rather than opinion, why the score is shown with its inputs and never alone, the identity that every account has one score from the same rule, and how the score is validated against what happened.
16 Sept 20263 min readHow a team whose data lives in Excel or Google Sheets can ask plain-English questions of the workbook they already keep: reading the structure as it is, mapping columns once, validating the totals, and getting answers with the rows behind them, without pivot tables or a data warehouse.
16 Sept 20263 min readWhy every effect measured from a company's own history, samples and prescriptions, onboarding milestones and renewals, first-touch latency and first-year revenue, delayed installs and non-renewal, is an association and not a cause, the selection that makes it so, the comparison group that removes most of the noise, the sentence that reports it honestly, associated with rather than caused, why it is still enough to direct effort, and the two claims that are never made from it.
16 Sept 20263 min readA number in a report or an assistant's answer is a claim until it carries a citation, and a citation has five parts: the table, the row, the period cut, the definition version, and the upload it was computed from. Why each part is there, what a reader does with it, the difference between a citation and a link, how a citation survives a re-upload, and the three shortcuts that turn a citation into decoration.
16 Sept 20263 min readHow an analytics product uses a language model without letting it touch a number: deterministic code computes every measure from stated definitions, a tool registry names each computation, the model plans which tools to call and reads their outputs, every output carries its citation, and the model's own text is labelled as a reading. The four boundaries that make this hold, what crosses each, and what never does.
16 Sept 20263 min readWhat an answer engine extracts from a page about a commercial measure: the one-paragraph direct answer, the formula in a form it can quote, the worked table, the definitions it can cite, and the questions it can answer verbatim. Why a page written for a person who has to compute the measure is also the page an assistant cites, the structure that serves both, and the four habits that make a page unquotable.
16 Sept 20263 min readWhat separates a question an AI analytics assistant answers correctly from one it answers plausibly: naming the measure, the level, the period and the comparison, asking for the table before the sentence, and checking the citation. Six rewrites of common questions, and the four kinds of question an assistant should refuse.
16 Sept 20263 min readA plain comparison for a commercial leader deciding what to buy: what self-service dashboards do, what an AI layer adds, the parts of the job that neither replaces, the two failure modes specific to AI analytics, generated numbers and unstated definitions, and the test that tells a trustworthy AI analytics tool from a fluent one.
16 Sept 20263 min readHow to decide what counts as a movement worth reporting for each measure in a weekly sales digest: thresholds from the measure's own week-to-week variability rather than a round number, the floor that stops small denominators generating alerts, the cap on items per digest, and the review that tightens or loosens each threshold from how many alerts were acted on.
16 Sept 20262 min readA buyer's checklist for AI sales analytics: where each number comes from, whether the model computes or reads, how definitions are versioned, what the roll-up identity is and whether it is checked, how a wrong answer is caught, what data leaves your building, whether names are needed, what happens when the exports change, what the assistant refuses to do, and what you can take with you. Each with the answer that should worry you.
16 Sept 20263 min readWhy self-service analytics fails without one governed definition of each metric, how to run a versioned registry where net revenue v4 supersedes v3 on a date, how scope is applied in the data layer rather than a prompt, and what an audit trail per answer has to record.
16 Sept 20263 min readCoverage intelligence is the practice of measuring, from a company's own sales data, which customers it reaches, how much of each customer's spend it holds, and where the gaps are, with every figure reconciled to the ledger. This guide defines it, sets out its five measures, coverage, share of wallet, gap at norm, dormancy and concentration, shows how they roll up, and distinguishes it from BI, CRM reporting and sales analytics.
16 Sept 20263 min readAn analytics assistant that reads computed tables and declines arithmetic will refuse some requests, and how it refuses decides whether the reader trusts it. This guide sets out the four requests a well-built assistant declines, add these up, average this, extrapolate that, estimate what is missing, what each refusal says, where it points instead, the tone that keeps the reader, and the refusals that are wrong because the table exists.
16 Sept 20263 min readThe case for taking a scheduled file from the customer's systems rather than connecting to them: what a live connector costs in credentials, per-customer schema and outages, why every system already exports a file, the four ways a file arrives, and when a bespoke adapter is still right.
16 Sept 20263 min readA plain argument for lists over dashboards as the output of sales analytics: what a dashboard asks the reader to do, what a list asks, why the dashboard's forty tiles produce no action and the list's ten rows produce ten, the four properties a list needs to be worked, ranked, valued, reasoned and owned, and how the dashboard survives as the place to check rather than the place to start.
16 Sept 20263 min readThe one rule that makes AI analytics safe to put in front of a sales desk or a bank: the language model chooses the calculation and explains the result, and deterministic code computes every figure. Why the alternative fails, and what the rule changes in practice.
16 Sept 20264 min readA practical answer to whether a general-purpose AI chatbot can analyse a sales export. It can: given a clean file and a precise question, a chatbot with a code tool will compute correct totals, rankings and simple rates. It goes wrong in predictable places: silent assumptions about what columns mean, duplicated customers, joins across files, definitions that change between sessions, arithmetic done in prose rather than in code, and no check against the ledger. This page lists what works, the seven failure points, five checks anyone can run on an answer, and what a purpose-built tool does differently.
17 Sept 20265 min readA checklist for anyone about to put a customer ledger, a CRM export or a contract file into an AI analytics tool. It covers what is actually in a sales export that makes it sensitive, the nine questions to put to any vendor, covering model training, where data is processed, who at the vendor can see it, retention and deletion, tenancy, pseudonymisation, access control, the model provider behind the tool, and what happens at exit, what a good answer sounds like for each, and how to reduce the risk yourself by removing what the analysis does not need.
17 Sept 20265 min readThree software categories with overlapping names. Sales analytics measures what happened in your own sales data. Sales intelligence, as the market uses the term, mostly means third-party data about companies and contacts to prospect into. Revenue intelligence mostly means capturing and analysing calls, emails and CRM activity to inspect deals and forecast. This page sets out what each actually does, the data each runs on, the question each answers, where coverage intelligence sits among them, and how to tell which one a problem needs.
17 Sept 20264 min readA plain explanation of why large language models make arithmetic errors, for people deciding whether to trust AI with business figures. A language model predicts the next piece of text; it does not calculate. It reads numbers as fragments, not quantities, produces answers that look right, and is most likely to be wrong on long numbers, long columns and multi-step calculations, which is exactly what business data involves. This page explains the mechanism without jargon, shows the kinds of error to expect, and sets out the design rule that removes the problem: the model decides what to compute and explains the result, and code does every calculation.
17 Sept 20265 min read