From Data to Decisions: Getting Straight Answers Out of Your Own Numbers

Most businesses do not have a data problem. They have a definitions problem wearing a data costume.

The symptom is familiar. Someone asks a straightforward question. How many jobs did we close last month, what did a client cost us to win, is the new channel working. Three people produce three different answers. Each one is defensible. Each one came out of a different system with a different idea of what counts. So the meeting becomes an argument about the numbers rather than a decision made with them, and eventually people stop asking and go on instinct.

This is a solvable problem, and AI solves most of it. The order is what matters. Settle what the words mean first, because a system pointed at three definitions that disagree will produce the disagreement faster. Settle them, and AI then does the part that was never realistic by hand.

Why the dashboard did not fix it

Most companies have already tried. Somebody built a dashboard, it looked impressive for about six weeks, and then people quietly went back to their own spreadsheets.

The reason is almost never the tool. It is that the dashboard was built from the data that happened to be available rather than from the decisions anyone actually makes. It showed twenty things, four of which mattered, and none of which were defined tightly enough for two departments to agree on them.

The second reason is staleness. A dashboard that requires somebody to refresh it by hand is out of date within a fortnight, and a number that might be old is a number nobody will act on.

Definitions come first

This is the unglamorous part, and skipping it is why the expensive version fails.

Before any of it is automated, a small number of terms have to mean one thing across the business. What is a lead: anyone who filled in a form, or someone who met your criteria? When is a deal closed: when they verbally agree, when they sign, or when they pay? Does revenue count on the date invoiced or the date received? Is a repeat client a new sale?

None of these have universally correct answers. They only need consistent ones. An hour spent writing them down and getting agreement removes the majority of the disagreements that make reporting useless, and it costs nothing.

It also has to be written somewhere people can find. A definition living in one person's head is the same problem in a different form.

Start from the decision, not the data

The right way to build reporting is backwards. Not what can we measure, but what do we decide, how often, and what would we need to know to decide it well.

For most owner-run businesses the list is short. Where should next month's marketing budget go. Do we need to hire. Which clients are quietly unprofitable. Is the pipeline enough to cover the quarter. Which service line is actually carrying the business.

Five questions. Each one needs perhaps three numbers behind it. That is your reporting requirement, and it is far smaller than the dashboard you were about to commission. Everything else is interesting rather than useful, and interesting is what makes people stop opening the page.

What AI actually contributes

Once definitions are settled and the questions are known, there are three places it does real work.

Pulling data out of systems that were not designed to share it. A great deal of business information sits in formats that resist reporting: PDF invoices, email threads, free-text notes in a CRM, supplier statements. Reading unstructured material and turning it into structured fields is genuinely difficult to do any other way, and it is often the step that was blocking everything downstream.

Explaining the change rather than displaying it. A chart tells you cost per client rose eighteen percent. It does not tell you that it rose because one channel scaled while converting worse, and that the other three were flat. That second sentence is the useful one, and producing it consistently is work a system can do every week where a person does it when they have time.

Watching, so nobody has to. Most reporting is reviewed on a schedule, which means a problem beginning on the second of the month is found on the thirtieth. A system checking continuously and saying something when a number moves outside its normal range converts reporting from a monthly ritual into something closer to an alarm.

The weekly summary

The single highest-value output is not a dashboard at all. It is a short written summary that arrives before the week starts.

A dashboard requires someone to remember to open it, and busy people do not. A summary in an inbox on Monday morning, carrying five numbers, what changed, what caused it, and what needs attention, gets read, because reading it is easier than not reading it.

It also changes the meeting. When everyone has seen the same five numbers before they sit down, the hour goes on what to do rather than on establishing what happened.

Attribution, honestly

Connecting revenue back to what produced it is the thing most businesses want and the thing most likely to be oversold to them.

Some of it is straightforward. If enquiries carry their source and that source survives into the CRM and onto the closed deal, you can answer which channel produced the revenue. That alone is more than most companies have, and it is mostly a plumbing exercise.

What is not straightforward is credit across several touches. A client who saw an ad in March, read something in May and was referred in July is not cleanly attributable to any one of them, and any model claiming otherwise is making a choice you should know about. Use attribution to decide where to put more money, not to prove which channel deserves it. The first is a decision. The second is an argument.

What to watch for

Research published by MIT's NANDA initiative in 2025 found that the large majority of enterprise AI projects produced no measurable financial return, and a recurring reason was that nobody had established a baseline, so there was nothing to measure against afterwards. Reporting projects are especially prone to this, because the thing being improved is measurement itself.

Two other failure modes are worth naming. Reporting on everything, which guarantees nobody reads it. And trusting a number nobody can trace. If a figure on a dashboard cannot be followed back to the record it came from, the first time it looks wrong the whole system loses credibility, permanently.

A reasonable first version

Agree the definitions. Pick the five decisions you actually make. Connect the systems that hold those numbers. Put them in one place that updates itself. Send a short summary before the week starts. Make every figure traceable to its source.

That is a modest project and it is more than most companies your size have. The elaborate version can wait until the simple one is being used.

Where the numbers actually live

Before anything can be connected it helps to know what you are connecting. For most companies of this size the map is shorter than expected and always includes at least one surprise.

Sales and pipeline sit in the CRM, assuming people fill it in. Money sits in the accounting system and rarely agrees with the CRM, because one records invoices and the other records optimism. Marketing spend sits in the ad platforms, each with its own idea of what a conversion is. Delivery sits in a project tool, or in a spreadsheet, or in somebody's calendar. And a surprising amount sits in email, which is where the exceptions get agreed and never recorded anywhere else.

Doing that inventory is worth an afternoon on its own. It usually reveals that one number everybody quotes has no system behind it at all.

Who owns it afterwards

Reporting decays faster than almost anything else, because the business changes underneath it. A new service line appears and does not fit the categories. Somebody renames a pipeline stage. A definition drifts.

Someone has to own the definitions. Not the dashboard, the definitions. Their job is to notice when the business has changed shape and update what the numbers mean before the reporting quietly becomes wrong. In a small company this is fifteen minutes a month and it is the difference between a system people trust in year two and one they abandoned in month seven.

AI Optimize starts from the decisions you make each week and works back to the numbers that answer them, connected, live, traceable and summarised before Monday. That work sits under Reporting & Data.

Sources

• Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, The GenAI Divide: State of AI in Business 2025, MIT Project NANDA, July 2025. Based on a review of over 300 publicly disclosed AI initiatives, 52 structured interviews and 153 survey responses from senior leaders, conducted between January and June 2025.