Document Intelligence and Spreadsheet Intelligence: The Business AI That Pays Off First

Every established business, whatever it sells, is quietly running on two things. There is a pile of documents: the invoices and contracts, the intake forms, the PDFs a supplier emails over, the scans in a shared drive nobody has opened in a year. And there is a stack of spreadsheets: the one that tracks the thing the real software cannot, the pricing sheet three people are afraid to touch, the master list that somehow became the source of truth for the whole department. This is not a failure of organisation. It is what running a business looks like. The documents are where information arrives, and the spreadsheets are where it gets held together.

That is also, as it happens, exactly where the most useful artificial intelligence lives. Not in a chat window, and not in a grand plan to reinvent the company, but in reading the paperwork you already receive and making the sheets you already trust actually trustworthy. There are names for these two disciplines, document intelligence and spreadsheet intelligence, and they describe the business AI we focus on, because they share the two properties that separate AI that pays for itself from AI that impresses in a demo and dies in six weeks: the raw material already exists, and every answer can be checked against a source.

This piece is about what those two things actually are, in concrete terms, why they are the place to start rather than a place to end, and where they go wrong so you can tell the honest version from the sales pitch.

What document intelligence actually is

Start with the pile of paper, because it is bigger than you think and less used than you would ever guess.

By the most widely cited estimates, 80% to 90% of the data in a typical organisation is unstructured: documents, emails, scans, notes, transcripts, the text and images that do not sit in neat rows and columns. For most of computing history that data was, in practice, unreadable by software. A person had to open each document, read it, understand it, and type the important parts into a system that could actually do something with them. The document was where information came in, and a human was the only bridge from the document to the database.

Document intelligence is what removes that bridge. In plain terms, it is applied AI that turns an unstructured document into structured, verified data and answers. It does a few distinct jobs, and it is worth separating them because people tend to collapse them into one vague idea of “AI reading stuff.”

The first job is extraction. Take the invoice, the contract, the intake form, and pull out the specific fields your business runs on: the amount, the dates, the party, the account, the line items, the clause that matters. Not a summary of the document, the actual values, written into the system where they belong, with anything the document left ambiguous flagged for a person rather than quietly guessed. Crucially, a well-built extraction step cites where each value came from, so a person can confirm the number against the source in one click instead of taking it on faith. This is the plainest version, and it is the one that ends the daily indignity of retyping the same information from a PDF into three different screens. It is the front end of the admin pipeline that quietly runs every business.

The second job is classification and routing. Given a stream of documents that arrive in no particular order, work out what each one is, an invoice versus a purchase order versus a complaint versus a renewal, and send it where it needs to go, with the urgent ones surfaced and the routine ones filed. The value here is not the sorting itself. It is that nothing sits unseen in a shared inbox for a week because the one person who triages it was on holiday.

The third job is the one businesses almost never get, because it was never possible before: answering questions across everything. Not “read me this contract” but “which of our supplier agreements auto-renew in the next ninety days,” “which invoices this quarter are missing a purchase order,” “what did we actually commit to across all the statements of work we signed last year.” These are questions whose answers were always technically present in the documents and practically unreachable, because getting them meant a person reading hundreds of files. When the documents become searchable by meaning rather than by keyword, the answer becomes a report instead of a week of someone’s life. That searchable-by-meaning part rests on a specific technique worth understanding on its own, which is what embeddings do: they let a system find the contract that means what you asked, even when it uses none of your words.

Notice what all three have in common. The document was always the raw material, never the result. It is the same law that governs what a call transcript is worth: the capture is cheap and getting cheaper, and the value is entirely in what the next step makes of it. A summary in your inbox is a paragraph. A structured record that flows into your systems, a routed document that reaches the right desk, an answer pulled from ten thousand pages in seconds, those are products. The difference between the two is design, not luck.

What spreadsheet intelligence actually is

Now the other stack, which is where it gets interesting, because the spreadsheet is not just a place information sits. It is a place decisions get made, and it is far more dangerous than anyone treats it.

Here is the uncomfortable research. When auditors actually sit down and check operational spreadsheets, the better-designed studies keep finding errors in the overwhelming majority of them. Raymond Panko, who has spent a career on this, summarises the field audits bluntly: the more rigorous ones find errors in the large majority of the spreadsheets they examine, upward of nine in ten once someone looks carefully. Not toy spreadsheets. The real ones, the ones companies run on.

And these are not harmless. In 2020, Public Health England lost track of 15,841 positive COVID-19 cases over about a week, because the results were being collated in an old .xls file that silently stops accepting rows near sixty-five thousand. The cases were not deleted. They just fell off the bottom of a spreadsheet that had quietly run out of room, and nobody was told. In 2012, JPMorgan disclosed a $6.2 billion trading loss in which, according to the bank’s own internal report, a risk model built in a spreadsheet divided by a sum instead of an average and understated the risk by roughly half, for years, before anyone caught it. And one of the most cited economics papers of its decade, the one used to justify austerity across governments, turned out to rest partly on a spreadsheet that accidentally left several countries out of an average. A graduate student found it by asking to see the file.

The point is not that spreadsheets are bad. They are one of the most useful tools ever built, and businesses reach for them precisely because they are flexible and immediate in a way that real software often is not. The point is that a spreadsheet gives you no way to know whether it is right, and the more important it becomes, the more people touch it, and the more places a single wrong cell can hide while quietly steering a decision.

Spreadsheet intelligence is the response, and like document intelligence it has two halves that are worth keeping separate.

The first half is reading and reasoning across the spreadsheets you already have. Point applied AI at the sheets that run a corner of the business and let it do what no one has time to do: answer questions across them, reconcile the ones that are supposed to agree and do not, and surface the broken references, the stale formulas, the row that got sorted out of alignment, the number that was pasted where a formula used to be. This is the audit nobody ever runs, done continuously instead of never. It turns the spreadsheet from a thing you hope is right into a thing you can actually interrogate.

The second half is the more consequential one: replacing the fragile, load-bearing spreadsheet with a real system that keeps what people liked and removes the risk they never saw. The reason a critical process lives in a spreadsheet is almost always that the spreadsheet was flexible and available when nothing else was. But once that sheet is running a real part of the operation, its flexibility has become a liability: no rules about what can go where, no record of who changed what, no check that the totals mean what they say. The move is to rebuild that specific process as a proper system, one that still feels quick and adaptable to use, but where a mistyped cell can no longer silently become a bad decision, and where the logic is written down once instead of living in the formula bar and one person’s memory. This is often the quiet first step toward a genuine business operating system, built one proven piece at a time rather than in a single risky rebuild.

There is a well-documented reason this matters more than it sounds. When a new system is harder or slower or missing something, people keep the old spreadsheet running beside it, and within a few weeks the spreadsheet is running the department again and the new system is a tab nobody opens. That is the single most common way an AI project fails, and it is why spreadsheet intelligence has to make the new thing genuinely better to use, not merely more correct. Correct and annoying loses to wrong and convenient every time.

Why these two are the business AI worth doing first

Step back and ask why document intelligence and spreadsheet intelligence, specifically, out of everything AI is supposed to be able to do. The answer is the reason they are the business AI we focus on, and it comes down to three properties they share.

The first is that the raw material already exists. You are not collecting anything new, changing how anyone works, or betting on a process that does not yet exist. The documents arrive on their own and the spreadsheets are already open. Starting with what you already have, at a point in the work where a person still confirms the result, pays back far sooner than a grand rebuild in the middle of the business, and it earns the confidence for the next step.

The second is that the work is checkable. This is the property that separates AI you can run a business on from AI you can only demo. A document extraction can cite the line it pulled a number from. A spreadsheet reconciliation can point at the exact two cells that disagree. When the output carries its own evidence, trust stops being a feeling someone has to summon and becomes a step built into the process. That also means it survives staff turnover, which a culture of “remember to double-check the AI” never does.

The third is the property that runs under everything we write, and it is worth stating plainly because it is the thing most people get backwards. The model is a commodity. The frontier system you can use is, give or take, the same one your competitor can use, and it will be roughly the same one everyone has next year. So the model is not where the advantage comes from. The advantage comes from the structure around it: which fields your work actually needs, which spreadsheet errors matter and which are noise, where a person checks the result, what happens when the document is messy or the sheet is malformed. That structure is a design decision made with judgment, up front, and it is the same reason the same AI can cost ten times as much or work ten times better depending entirely on how it was built. The document is the raw material and the spreadsheet is the liability. The judgment applied to them is the product, and it does not come in a box. It is the substance of our work with a business’s own data and the workflow automation built on top of it.

Where it goes wrong, honestly

None of this works if you skip the parts that are unglamorous, so here is the honest version of what makes it fail, because knowing the failure modes is how you tell a real build from a pitch.

It goes wrong when the underlying data is a mess and nobody says so. AI reads your documents and your historical spreadsheets faithfully, including all the errors, gaps, and quiet biases that accumulated in them over the years, and it will reproduce that history with total confidence unless someone designs against it. A system is only ever as honest as the records it learned from, and pretending otherwise is how a confident wrong answer gets treated as gospel.

It goes wrong when the checking step is optional. If verification depends on a busy person remembering to be suspicious, it will not happen, and the one document in fifty that the system misread will flow downstream unexamined. The checking has to be part of the workflow, not a virtue people are asked to have.

It goes wrong when the data is handled carelessly. Documents and spreadsheets carry some of the most sensitive information a business holds, contracts, financials, personal details, and processing them means being deliberate about consent, retention, and where the data physically lives, which in Canada means PIPEDA and often keeping the data in the country. That is a design requirement, not an afterthought, and it is one reason where the processing runs actually matters.

And it goes wrong, most often, when someone buys a tool instead of designing a solution. There are plenty of products that will extract fields from a document or flag anomalies in a sheet in the abstract. Which fields your records actually need, which spreadsheet is load-bearing and must not be disturbed mid-change, which patterns would change a decision and which are merely interesting, those have a right answer for your business and no universal one. The tool is what the judgment uses. The judgment is the work, and it is the difference between something your team trusts and something they quietly route around.

The part worth carrying out the door

The most valuable AI in your business is probably not the ambitious thing you have been picturing. It is sitting in the documents that arrive every day and the spreadsheets you are half-afraid to touch, and it is valuable precisely because those two things are already there and can already be checked.

Document intelligence turns the paperwork from a pile someone has to read into structured data you can trust, with the source attached. Spreadsheet intelligence turns the sheets you run on from things you hope are right into things you can actually interrogate, and, where they matter most, replaces them with systems that keep the flexibility and lose the risk. Both are built around how your business really works, both keep a person on the result, and both pay back fast because you are starting with what you already have.

That is the business AI we focus on. The model is the same one everyone gets. What you do with your own documents and your own spreadsheets is the part that is yours.