Why better decisions start with better representations

A tradie with a healthy business applies for finance and gets declined, because a model read three big invoice payments a month as dangerous volatility. A cafe with spotless arrears is eight weeks from collapse, paying its lender first while its buffers quietly drain, and nothing on any dashboard blinks. A credit team buys an expensive score, then discovers every competitor bought the same number, so it bought no edge at all. An executive asks why churn is rising and gets the answer six weeks later, in a deck, after the customers are already gone.

Different rooms, same failure: in every case the decision was made on what the data looked like, not on what the behaviour meant.

Or take two small businesses walking into the same lender. Same industry code, same requested amount, and on the bank statement, almost the same average balance. One of them will quietly fail within a year. The other is about to have its best year yet. Nothing on the surface tells you which is which. Would your model know? Would your score even ask?

This is not a data problem. Both businesses generated thousands of transactions, rich with behaviour: how they pay suppliers, how they recover from a bad week, how their revenue breathes with the season, which obligations they protect when money gets tight. The problem is that almost none of that behaviour ever reaches the model. It arrives as a column of amounts and dates, gets averaged into a few summary statistics, and the story dies in aggregation.

Here is the uncomfortable truth underneath most analytics and most AI: models do not see the world. They see representations of it. And every decision a model makes, every score, every forecast, every approval, inherits the quality of the representation underneath it. Feed a brilliant model a poor representation and you get a confident answer to the wrong question. The industry has spent a decade upgrading architectures. The representations mostly stayed where they were.

The best machine learning teams in the world have already learned this on their own platforms. Ask them where the wins actually come from and they will tell you plainly: from finding signals others missed, not from cleverer architecture. Architecture is now abundant. Meaning is still scarce.

Listen to how the market is asking for it

You do not have to take the argument on faith. Organisations are writing it into their job ads, in their own words, for roles they are struggling to fill. Read the postings from SaaS companies, retailers, adtech platforms and quant funds over the past year and the same questions repeat, almost word for word:

Which of our numbers can we actually trust for company level decisions? Where do our revenue, retention, churn and customer health signals disagree, and why? How do we make trusted data discoverable and usable by our leaders, our teams and our AI tools? How does each team own its own evidence without fragmenting one company truth? How do we reduce the latency from something happening to us knowing what it means? How do we measure the behaviours we care about most, the ones with no column in any database: customer health, loyalty, financial strain, an account quietly disengaging months before churn shows up?

Walk into any lender and the questions wear credit clothing but keep the same shape: which applicants should we acquire, which borrowers are drifting toward trouble while their arrears still look clean, what does this thin file actually tell us. Every one of these is a representation question. The people writing them are asking for a signal layer without yet having a name for it, and they are offering director salaries to anyone who can build one. That is what demand looks like just before a category gets its name.

A feature is implementation. A signal has meaning.

The word I use for a good representation is a signal: a measurable representation of hidden business behaviour. The distinction from a feature matters. A feature is a column, an implementation detail, something a pipeline computes. A signal is a claim about the world: this number measures how quickly a business recovers when a payment bounces, and businesses that recover quickly behave differently from businesses that do not.

One stream of raw data can yield dozens of these claims. Take a single daily cash balance over four months. Its trend tells you one thing. Its volatility tells you another. The persistence of its weekly rhythm, the day the pattern broke, the way expenses have been closing in on revenue: each is a different mathematical lens on the same underlying behaviour, and each carries information the others miss. The surface looks like noise. The representations underneath, read together, tell one coherent story.

The behaviour is in the numbers. It always was. Numbers in, behaviour out: that is what a representation layer does.

And here is where the lenses start climbing toward meaning. Understanding the detailed patterns inside thousands or millions of transactions, across many different business types, requires knowing what each movement of money is actually telling you about behaviour. Because underneath all the mathematics, credit has only ever asked three questions. Can this business pay: does the money coming in actually cover what must go out, with buffer to survive a bad month? Will this business hold: is the machine itself stable, does its rhythm persist, does it recover when something breaks? And the oldest question of all, the character question: will they choose to pay, even when money is tight, even sometimes when it is not?

None of those three is a column in any database. Ability, stability and willingness are hidden traits. But every one of them leaves fingerprints all over transaction data, and each can be modelled from it. Willingness shows itself in the payment hierarchy: watch which obligations a director protects when cash gets scarce, who gets paid first, who gets stretched, and how quickly a bounced payment is made good. Ability shows in the arithmetic of inflows against obligations and the depth of the buffers between them. Stability shows in whether the rhythm of the business persists, how much it shakes, and how it behaves after a shock. These are not scores anyone asserts. They are behaviours, measured from evidence, each built from many signals telling one coherent story about a trait you can never observe directly.

The same signal does not mean the same thing

There is a second layer to this, and it is the one generic scores get wrong. Meaning is conditional on who you are. Revenue arriving in lumps is a warning sign for a cafe whose money should flow daily through a terminal. It is completely normal for an ecommerce brand paid on a gateway settlement cycle. And for an invoice driven tradie, three large payments a month is the business working exactly as designed. A zero revenue week is a Tuesday, not a crisis.

A representation is not finished when it is computed. It is finished when it is interpreted against businesses like the one in front of you. That is why a single score, sold to everyone, reading every business through the same lens, keeps punishing borrowers for the shape of their own industry. Better representations are not just richer mathematics. They are mathematics that knows what kind of machine it is looking at.

A score is one answer. You have many questions.

There is no shortage of providers selling cashflow scores and bureau scores, and they share a quiet limitation: a shipped score answers exactly one question, at one fixed horizon, chosen by the vendor. Usually it is the probability of default over one window. Useful, until you ask anything else.

And you will ask something else, because lending is many decisions wearing one customer. What if you want to understand the characteristics of customers who go thirty days delinquent, versus sixty days, versus those who fail inside twelve months? Those are not the same behaviour at different strengths. Early delinquency is usually a timing and liquidity story, the rhythm wobbling. Twelve month failure is a structural story, the machine itself winding down. One number cannot carry both. What if you want to optimise which customers to acquire in the first place? What if you want to identify early delinquency customers while they can still be helped, and proactively design your collections strategy around how they actually behave, rather than dialling everyone on day thirty? A score bought for underwriting cannot be re pointed at any of this. It was shipped with its question welded on.

Signals work the other way around. A behaviour is measured once, validated separately against each question and each horizon it claims to answer, and then reused everywhere it has earned its place: acquisition, underwriting, pricing, monitoring, collections, renewal. Same vocabulary, many decisions. That is the practical difference between buying an answer and owning a representation: the answer expires the moment your question changes, and your questions never stop changing.

Honest representations know what they did not see

Real world data is full of holes, and the holes have shapes. One account visible out of three. Six months of statements, then darkness, then a renewal decision. A young business with no history anywhere, not because data is missing but because it has not existed yet. Outcomes observed only on the loans you approved, never on the ones you declined.

Most systems interpolate smoothly over holes they cannot see, with total confidence. A trustworthy representation does the opposite. It carries its own coverage: what could have been observed, what actually was, and how much the answer should widen because of the difference. It knows that a quiet account and an invisible account mean different things. It lets a thin file borrow strength from businesses like it, honestly widened, so that nobody is punished simply for being new.

Finding signal in noise is a craft, and a rare one

It is worth being honest about why this layer has been missing. Creating a new behaviour out of raw data is not a technique you look up. It is problem solving, creativity, domain knowledge, statistics, mathematics and machine learning operating together, in one mind, on one question. You need the domain sense to know which behaviour would matter, the creativity to imagine a way it might leave a trace in the data, the statistical judgement to know which mathematics is even legal on that kind of observation, and the modelling craft to know whether the result will survive contact with a real system. Remove any one of those and you get noise wearing the costume of insight.

I have interviewed hundreds of data scientists across my career. Almost none could invent even a few signals that generate real model uplift, or explain why one works. This is not a criticism of them. The default they were trained into is raw data in, model on top, and the craft of representation sits in the gap between disciplines that are usually taught, hired and managed separately. The result is that organisation after organisation sits on valuable patterns that are simply never captured, not because the data was missing, but because nobody was equipped to ask it the right questions. That scarcity is exactly why the capability has to become a platform: the market is already advertising for these people at extraordinary salaries, and mostly failing to find them.

The craft also travels further than people expect, because behaviour has the same mathematical shape in very different worlds. I spent my research years separating signal from noise in chaotic biomedical waveforms, and cashflow turned out to be the same animal wearing a suit. The slow drift that breathing adds to a heart signal is the seasonal cycle that tax dates and payroll timing add to a cash balance, and the same decomposition isolates both. The sharp transient that marks an acute cardiac event is the supply shock or the client default that marks a liquidity crisis, and the same detectors catch both early. The high frequency noise filtered away to reveal a true physiological rhythm is the churn of daily micro transactions hiding a business’s real trajectory. Same shapes, different nouns. Once you have learned to read one chaotic, non stationary system honestly, you can read the next one.

Representations must be earned

Here is where discipline separates a representation layer from a feature factory. It is now easy to generate candidate representations. Modern tooling, agents included, can produce thousands of them before lunch. Generation was never the hard part. The hard part is the gate: which of those candidates actually predicts real outcomes, holds up over time, survives multiple testing honestly, and can explain itself in the language of the people who must act on it.

So every representation worth deploying carries evidence. It is validated against outcomes, versioned, and watched for drift, because meaning that was true last year can quietly stop being true. And it respects one line that too many systems cross: prediction is not intervention. A representation that predicts risk has earned the right to rank and to warn. It has not earned the right to promise that changing the number changes the world. That claim is earned separately, through experiments, or not claimed at all.

This is also why the age of AI agents raises the stakes rather than lowering them. Language models are remarkable reasoners and unreliable accountants. Hand an agent raw transactions and it will hallucinate arithmetic with perfect fluency. Hand it named, validated, explained representations, each carrying its evidence and its coverage, and it can reason the way your best analyst does. Agents are only as trustworthy as the representations they reason over. Documents got their retrieval layer. Behaviour is still waiting for one.

Where decisions actually improve

Better decisions rarely come from a better algorithm applied to the same inputs. They come from the moment a system can finally see the thing that mattered all along: the payment hierarchy under strain, the rhythm dissolving before the arrears appear, the growth that is quietly being funded by stretching suppliers. Once a behaviour is represented, measured, interpreted for the business model in front of you, and backed by evidence, everything downstream improves at once. The models get better. The explanations get honest. The humans and the AI finally speak the same language about the same customer.

That is the whole argument. Decisions inherit representations. Representations inherit discipline. If you want better decisions, do not start with the model. Start with what the model is allowed to see.