What to Do With Ten Years of Data Nobody Uses

An established business holds an enormous amount of history. Every job, every quote, every invoice, every client conversation, going back as far as the systems do.

Almost none of it has ever been looked at as a whole. It exists as individual records that were useful once, retrieved occasionally, and otherwise stored.

Why it has never been examined

Not lack of interest. Three practical obstacles.

It is inconsistent. Ten years of records reflect ten years of changing processes, categories, naming and people. A job in 2016 was recorded differently from one in 2024, which makes comparison difficult before it makes it interesting.

Much of it is unstructured. The valuable detail sits in free text notes, email threads and documents rather than in fields. Analysis tools cannot read those.

Nobody has a spare month. The exercise has never had an owner or a deadline, and it competes with work that does.

The questions it can actually answer

Worth being specific, because the promise of historical data is usually oversold.

• Which clients were genuinely profitable, once hours and rework are counted rather than revenue alone.

• How long work actually takes, by type, against what was quoted. This is the single most useful output for most businesses.

• What predicts a job going wrong. Certain clients, certain job types, certain times of year, certain combinations.

• Which enquiries became clients, and what they had in common. This is your ideal customer profile, derived rather than assumed.

• What you have priced inconsistently. Almost every business finds similar work quoted at meaningfully different rates depending on who quoted it.

What it cannot tell you

It describes a business that no longer exists in exactly that form. Your costs, your market and your capability have changed.

So it is evidence about patterns rather than about current numbers. How long a job type takes relative to others is durable. What it cost in 2017 is not.

Treat it as a source of questions to verify rather than answers to act on directly, and it becomes considerably more useful.

Where AI makes it tractable

The obstacles above are precisely the ones that changed.

Inconsistency across years is a reading problem rather than a data problem. AI can recognise that a job recorded one way in 2016 and another way in 2024 is the same category, which no rule based tool could do without somebody writing every mapping by hand.

The unstructured material is where the value concentrates. The note explaining why a job overran, the email where scope was expanded, the reason a client left. Turning that into something comparable was previously a manual reading exercise nobody would fund.

And it is now a job of days rather than months, which is the difference between a project that gets approved and one that stays on a list.

Start with one question

The failure mode is treating this as an exploration. Somebody is asked to see what the data shows, produces an interesting document, and nothing changes.

Start with a decision you are currently making on instinct. Whether a service line is worth keeping. Which client type to pursue. Whether your quoting is systematically wrong on a job type.

Answer that one question, act on it, then ask another. Businesses that do this get value from three questions. Businesses that commission an analysis get a document.

Clean going forward, not backward

One temptation to resist: fixing the historical records so they are consistent.

That is an enormous job with almost no return, because the past is not going to be added to. What matters is that records from today onward are consistent enough to be compared next year.

Extract what you need from the history, then spend the effort on capture going forward. In two years that becomes a clean dataset without anybody having reworked anything.

Check what you are allowed to keep

Before mining a decade of records, one obligation is worth checking.

Under privacy rules in Quebec and elsewhere, personal information should be kept only as long as the purpose it was collected for requires. A ten year archive of client records probably contains material you no longer have a basis to hold.

That is not an argument against the exercise. It is an argument for doing both at once: extract the patterns you need, then apply a retention rule to what remains. Businesses tend to discover their retention exposure precisely when they start looking at old data, and it is better found deliberately.

Watch for the survivor problem

One analytical trap worth naming, because it produces confident wrong answers.

Your history contains the clients who stayed and the jobs that completed. It contains far less about the enquiries that never converted, the quotes nobody accepted and the clients who left quietly.

Drawing conclusions about what makes a good client from the ones who remained tells you about the survivors rather than about the population. Where possible, look at what you lost alongside what you kept, and if that data does not exist, that is itself worth fixing going forward.

The output should be a decision, not a report

Judge the exercise by whether something changed.

A price adjusted. A service line dropped. A client type deprioritised. A quoting assumption corrected. If a month of analysis produced insight and no decision, the analysis was not the constraint and repeating it will not help.

That test also keeps the scope honest, because it forces the question of what you would do differently before the work starts rather than after.

Do it before you sell, not during

One situation makes this urgent rather than useful.

Anybody buying a business, lending to it or investing in it will examine the operating record, and they will ask questions the accounts cannot answer. Client concentration over time, retention by cohort, margin by service line, how reliably the business forecasts.

Assembling that under a deadline, from inconsistent records, while running the business, is where transactions slow down and value gets discounted. Doing it two years early costs the same work and produces a considerably better position.

AI Optimize reads the history you already hold, including the notes and threads no tool could parse, and answers one question at a time. That work sits under Reporting & Data.