Some questions about your business can be answered as far back as 2023. Others can only be answered starting six months from today, and which kind you have was decided years ago by a field somebody did or did not fill in. How much did that customer actually bill us this year? Which of my people keeps ending up on the jobs that come back? Did anyone confirm this work with the client, anywhere? Owners ask questions like these out loud all the time, and what separates those questions is not difficulty. It is where the answer is already written down. A question whose answer sits in QuickBooks or HubSpot carries one price. The same question, if its only trace is a Slack thread from March, carries another. And if it only ever existed as an hour of talking that nobody recorded, there is no price, because there is nothing to read.
That price is knowable before you spend anything. Any question you can think to ask lands in one of three rows, and the row tells you what the answer costs and which way its clock runs. Some rows answer backward, into years you have already lived. Others only answer forward, from whatever day you start them. Most of what gets sold as an AI decision is really a sorting job you can do yourself, on a napkin, in about ten minutes.
Which Questions AI Can Answer About Your Business: The Three Rows
Where does the cost of an answer come from? From where the record sits today, not from the model doing the reading. Three rows cover it.
| What you ask | Where the answer lives | What it costs |
|---|---|---|
| How much did we invoice this customer this year, and what did it cost us to serve them | One system you already run | A connection. Works on records going back years, from the first day |
| Which of my people handled the jobs that came back twice | The job was recorded, the person on it was not | One field, filled in every time. Counts forward from the day you add it |
| Did anyone confirm this work with the client, anywhere | Split across a plan, a chat thread, and a phone call | A build, and sometimes a decision about who is allowed to read what |
The rows are not a maturity ladder and they are not about your industry. A twelve-person shop can sit in row one on money questions and row three on scheduling questions on the same afternoon. Which row a given question falls into came out of how the work happened to get recorded, what somebody made a point of typing in and what stayed in conversation. Sorting your questions is the part nobody does, and it is the part that turns a vague ambition into a dated plan.
Row One: The Answer Already Sits in a System You Run
What does it cost to answer a question whose record already exists in a system? A connection, and nothing else. This is the row owners consistently underestimate, because the payoff is retroactive: on the day the connection is made, the answers reach back three years, not just forward from today.
The money and office side of your stack is where most businesses are already sitting without knowing it. QuickBooks, Xero, Stripe and Square all ship connectors built by the vendor itself rather than by hobbyists. HubSpot's went generally available in April 2026. Pipedrive opened its own to every plan, including the cheap ones, on June 30, 2026. An owner running a normal small-business stack in 2026 can often ask a question about last quarter's margins and get an answer the same week, with no new software bought and no process changed.
Two things spoil the row, and both are worth checking before you count on it. The tools closest to the actual work are the least connected: field service software still leans on community-built connectors, and access is often gated by plan tier rather than by technology. One popular field platform exposes its API only on the top plan. One popular sales platform gives its entry tier no programmatic access at all, which no third party can work around. The current per-tool state is worth reading directly in the tool-by-tool list of what connects today before anyone promises you a timeline.
Row one is not theoretical. One plumbing company records every finished job in the system its crew already uses, and a system of mine has been reading those job records month after month.
Row Two: The Record Exists, but One Piece Was Never Written
What happens when the job was recorded but the exact thing you want to ask about was not? The cost is one field and the habit of filling it in, which is cheap. The catch is the timing, and it is the costliest misunderstanding in the whole subject.
The usual shape: closed jobs are all in the system, but the person who did the work was never attached to the record, or the type of work was never categorized. Everything else about the job is there. Add the missing field today and you can ask the question six months from now about those six months. You cannot ask it about last year, ever. No model reconstructs a fact that nobody wrote down, and if a product answers that question anyway, it is producing a plausible number rather than a true one.
That asymmetry is why row two deserves a decision this week rather than next quarter. A row-two question is a clock that only starts when someone starts it. The cheapest version is a single mandatory box on the screen your crew uses to close a job, chosen so it cannot be skipped and does not need explaining. A record kept that way compounds: each month adds to what the business can ask about itself, which is the argument in why the data you own is the moat a generic model cannot copy.
Row Three: The Answer Is Scattered Across Tools and Conversations
The question where every piece exists somewhere but no single place holds the whole thing costs a build, because the pieces have to be pulled into one place a question can read, and occasionally it costs a decision before it costs any work at all.
Scattered comes in three shapes, and they carry different prices. The plan lives in one tool and the confirmation lives in a message thread: joinable, ordinary work, priced by how many systems have to agree on what a job is. The record exists in two systems that disagree about it: also work, plus a ruling from you on which one wins. And the briefing that only ever existed as an hour of talking: that is row two wearing a costume, and no build reaches it. Sorting which shape you have takes one conversation and kills the wrong project before it starts.
Row three is also where the choice of what you plug in starts to matter, because a build that assembles your scattered records into a vendor's system solves the question and creates a new dependency in the same move. That tradeoff has its own procedure in how to evaluate AI tools before you move your records in. Sort the row first. The evaluation only earns its keep once you know a build is genuinely what you need.
The Permission Wall: Answerable Questions That Stay Blocked
A question can be technically answerable today and still not get answered. Answering it means reading what people on your team write to each other, and no vendor can make that decision for you.
Take the third question from the table. For any system to answer "did anyone confirm this work with the client, anywhere", it has to read the messages where somebody would have said so. The chat platforms have landed on opposite sides of that line. One hands out access a channel at a time; another covers an entire organization with a single administrative approval. Which model each platform uses is set out in the platform-by-platform breakdown of who can read what. Either way the decision lands on your desk, and on some platforms it lands as a switch with only one setting.
That decision is not a technical step and should not be treated as one. Whoever signs it is choosing how much of the company's internal conversation becomes readable, and that choice outlives the project it was made for. The people whose messages are in question have a stake in it, and in some places more than a stake: a works council or a data protection officer may have to sign before you do. The honest version comes with terms attached. Mine are stated the same way on every project where the records are sensitive.
Names and identifiers are replaced before a record ever reaches the model, so what gets read is a work history rather than a person's history. The data layer sits apart from the answering layer, and a question pulls only the slice it needs instead of opening the whole store. Nothing is absorbed into the model itself, so there is nothing searchable sitting inside it. Access runs through authenticated systems: granted for a stated purpose and logged. It can be withdrawn.
Terms like those are what make the conversation possible at all. An owner who opens with "the AI is going to read our chats" gets a no from the room, and deserves one. Nobody in that room is objecting to the capability. They are objecting to an open-ended grant that names no purpose and comes with no way back, which is exactly what that sentence sounds like. The same capability, with the terms named first, becomes a normal decision that a normal team can say yes to, and one you can put in writing before anyone opens a console.
Trace One Question to Its Row
The move is small and you can make it today. Pick the one question you would most like a straight answer to, and go find where that answer is sitting right now. If it lives in a system you run, you are looking at a connection and answers that reach backward through your history. If the record is there but one piece is missing, you are looking at one field and a date about six months out. Scattered across tools and conversations means a build, and possibly a conversation about permission before anyone writes code. That earlier question, whether the record exists at all, is the test that runs before any tool evaluation. The row tells you what an answer costs.
What surprises most owners is how much of what they assumed needed a new system turns out to be row one. The records are already in the business. Settle the terms and make the connection, and the answers are already there, built out of what you have been writing down for years. The rest is a habit you start on a Monday, and a clock that only counts from the day you start it.
Nobody sold you those rows. They came out of how your own people recorded work, back when none of this was on the table. That also means you are the one who can move them.
Provenance: the connectivity facts and dates were verified against primary vendor sources on 2026-08-27, and the platform-by-platform message-access rules are taken from those vendors' own published permission models. The governance terms in the permission section are the author's standing practice on projects handling sensitive records, described from his own architecture. No client, vertical or account identifiers are disclosed, and no figures appear that are not first-party or sourced.
If you're trying to sort which questions about your business are actually answerable today, a 30-minute call can sort your specific list into the three rows.