If AI keeps getting your business wrong, the data is usually why: every fact lives in several places, and the AI picks one and sounds sure. Here is how to think about your data, and the two fixes that make AI both right and safe.
How most businesses run today
Customer system$86,000Typed by sales, Tuesday
Project tool$86,000Typed by ops, last month
Someone’s laptopNorthstar_final_v3.xlsxOut of reach
What’s Northstar’s contract value?
?
It picks one
$68,000 a year.
Read from the Data Lake, tab 7. Four numbers to choose from, and no way to tell.
Accounting$68,000Retyped by finance
Data Lake, tab 7$68,000Pasted three weeks ago
An inbox“Northstar numbers” attachmentOut of reach
Ops asks what Northstar’s contract value is. The AI chip in the middle reaches out along four dashed wires to four copies of the number: the customer system says eighty-six thousand, typed by sales on Tuesday; accounting says sixty-eight thousand, retyped by finance; the project tool says eighty-six thousand, typed by ops last month; the Data Lake’s tab seven says sixty-eight thousand, pasted three weeks ago. Two more copies sit out of reach with no wire to them: a spreadsheet on someone’s laptop and an email attachment. A question mark appears on the chip, and it answers sixty-eight thousand a year, read from the Data Lake, with four numbers to choose from and no way to tell.
How AI uses your data
AI doesn’t know your business. It looks things up.
An AI knows nothing until it reads something. Every answer is a lookup: find the fact, read it, answer or act on it. So what it can reach decides what it can do, and the state of that fact decides how good the answer is.
Customer system
Contract value$86,000 a year
Changed Tuesday, by sales
Ops asks
What’s Northstar’s contract value?
looks it up
Answers
$86,000 a year.
Read from the customer system, changed Tuesday
Or acts
Drafts the invoice for $86,000 in accounting
One question, one lookup, one answer with its receipt. The same fact can also start work: an invoice drafted from the real number.Ops asks what Northstar’s contract value is. The AI chip in the middle sends a mark up a dashed line into the customer system, where the contract value row reads eighty-six thousand a year, changed Tuesday by sales. The mark comes back down and the AI answers eighty-six thousand a year, read from the customer system, changed Tuesday. A second card shows what else it could do with the same fact: draft the invoice for eighty-six thousand in accounting.
So everything rests on where your facts live and what state they are in. Which raises the question of how they ended up in six places to begin with.
Why it happens
Systems arrive one problem at a time. None of them knows the others exist.
Each system was bought to fix one department’s problem, and each only knows what somebody typed into it. Nobody asked them to talk to each other. So the same customer gets typed in four times, four ways, and a spreadsheet grows in every gap.
Customer systemBought for lost leadsCustomerNorthstarYear 1
Export, paste
AccountingBought for late invoicesCustomerNorthstar Grp LtdYear 3
Export, paste
Project toolBought for missed datesClientNS GroupYear 6
Export, paste
HR systemBought for holidays on paperAccountNorthstar (C-1048)Year 8
One customerFour systemsFour spellingsWires between them 0
Each system arrived to fix one thing, and somebody typed the customer in again. Nothing passes a fact from one to the next except a person with a spreadsheet.A timeline of four systems arriving over eight years. Year one, the customer system, bought for lost leads, with the customer typed in as Northstar. Year three, accounting, bought for late invoices, with the customer typed as Northstar Grp Ltd. Year six, the project tool, bought for missed dates, with the client typed as NS Group. Year eight, the HR system, bought for holidays on paper, with the account typed as Northstar C-1048. In each gap between two systems sits a small spreadsheet labelled export, paste. A tally beneath reads: one customer, four systems, four spellings, wires between them, zero.
Excel is the software that runs between your software. It came out in 1985, the year of Back to the Future, and it still fills every gap, because somebody fills it by hand. Here is what that costs.
What it costs
Nothing connects the systems, so people copy facts by hand.
One number, one week: typed once, exported, pasted into the spreadsheet, emailed, typed again. At a company we worked with, forty-five people spent half their week doing this, because nothing passed a fact from one system to the next.
$86,000
MondayBornSales types it into the customer system$86,000
TuesdayExportedOps downloads the customer as a .csv$86,000
TuesdayPastedInto the Data Lake, tab 7$86,000
WednesdayEmailedThe tab goes to finance as an attachment$86,000
ThursdayRetypedFinance keys it into accounting$68,000
Month end: accounting invoices $68,000, sales expects $86,000, and both are “the number”.Five stops in a row for one fact. Monday, born: sales types eighty-six thousand into the customer system. Tuesday, exported to a CSV. Tuesday, pasted into the Data Lake on tab seven. Wednesday, emailed to finance as an attachment. Thursday, retyped into accounting as sixty-eight thousand, the digits swapped. A marker travels the five stops.
Forty-five people at half a week each is twenty-two full-time jobs, and every one of them is retyping. That is margin. It is also the setup an AI walks into on day one.
Then AI arrives
AI doesn’t clean up messy data. It scales it.
Point an AI at the spreadsheet and it picks one of two numbers. Give it the COO’s login and it answers a new starter with the COO’s access. It never sounds unsure. Same cause both times: the fact had no home, and the rule lived in a person.
With the master key
Ops asks
What’s Northstar’s contract value?
The AI answers
$68,000 a year.
Read from
Data_Lake_v14, tab 7
Last changed
3 weeks ago, by someone
Read as
The COO
Wrong number
A new starter asks
What does the delivery team earn?
The AI answers
Read from
Data_Lake_v14, Payroll tab
Rule checked
None on file
Read as
The COO
Wrong personThe small print names both failures: a three-week-old copy, and the wrong badge.Two answer cards, labelled with the master key. Ops asks Northstar’s contract value and the AI answers sixty-eight thousand a year, read from the Data Lake’s tab seven, last changed three weeks ago, read as the COO. Stamped wrong number. A new starter asks what the delivery team earns and the AI answers, the figures hidden here, read from the Payroll tab, no rule on file, read as the COO. Stamped wrong person.
Nobody did anything wrong, and the AI did exactly what it was set up to do. The fix is not a better AI. It is two changes to the data underneath it, and the first has nothing to do with AI.
The fix, part one
Connect the tools you already have.
Your systems could talk. Nobody wired them. An integration is a wire between two tools: change a fact in one and the other hears about it. Close a deal and the project and the invoice appear on their own, with the real number. No export, no paste, no retype, and no AI needed yet.
Systems connected
Customer systemSalesDeal closedNorthstar · $86,000 · starts 14 Oct
Project toolOpsProject createdNorthstar · starts 14 Oct
Data_Lake_v14_FINAL (use this one).xlsxLast opened: 3 weeks ago
One fact, typed once, arriving everywhere it is needed. Thursday’s retype never happens, and nobody opens the spreadsheet.Three systems in a row, joined by solid wires with arrows: the customer system, the project tool and accounting. A deal closes in the customer system, Northstar at eighty-six thousand starting the fourteenth of October. A marker carrying eighty-six thousand travels the wire to the project tool, where a project is created starting the fourteenth of October, then on to accounting, where the invoice schedule is set at eighty-six thousand. Beneath them the Data Lake spreadsheet fades, last opened three weeks ago.
Wired systems don’t need a spreadsheet between them. The fact arrives on its own, so an AI finds one number instead of two. Now it needs to know which system to look in, and who may see what it finds.
The fix, part two
Give every fact a home and a rule.
One record says which system holds each fact, so there is one number and the AI knows where to look. Each fact carries a line saying who may read it, and the AI reads as whoever is asking. Right and safe, from the same fix.
One customer recordNorthstar Group Ltd.
Also known as NorthstarNS Group / C-1048
FactValueLives inWho may read it
Contract value$86,000 a year••••Customer systemOps · finance · leadership
Delivery statusPhase 2, on track••••Project toolEveryone
Delivery team payOn file, 4 people••••HR systemLeadership · people
Reading as
Ops
The COO
Finance
New starter
Five facts, five homes, one line each on who may read it. Nothing moved out of the systems it lives in.The Northstar Group Ltd. record with five facts. Each shows its value, the system it lives in, and who may read it: contract value from the customer system for ops, finance and leadership; delivery status from the project tool for everyone; billing contact, missing with sales as owner, for everyone; margin from accounting for finance and leadership; delivery team pay from the HR system for leadership and people. Beside it, the AI chip wears a badge that changes between ops, the COO, finance and the new starter, and the rows lock and open to match.
Four people ask the same question. One record answers, and each of them sees exactly their share. Go back to the two questions that went wrong.
Ask again
The AI sees what you’re allowed to see, not what the person who built it could.
The same two questions, looked up through the record.
With one record and rules
Ops asks
What’s Northstar’s contract value?
The AI answers
$86,000 a year.
Read from
Customer system, live
Last changed
Tuesday, by sales
Read as
Ops
One number
A new starter asks
What does the delivery team earn?
The AI answers
Not available to your role People and leadership own that. On Northstar you can see delivery status and the billing contact.
Read from
HR system, rule on the fact
Rule
Leadership and people only
Read as
New starter
Right personThe answers changed because the lookup did: a live system instead of a copy, the asker’s badge instead of the builder’s.The same two answer cards, labelled with one record and rules. Ops gets eighty-six thousand a year, read live from the customer system, last changed Tuesday by sales, read as ops. Stamped one number. The new starter is told the pay is not available to their role, that people and leadership own it, and what they can see instead. Stamped right person.
Nobody had to remember anything. The record knew where to look, and the rule knew who was asking. Once that is true for one record, the week around it starts to change.
What changes
The spreadsheet stops being the connection between your systems.
The same week, with the systems wired and one record underneath them.
Retyping stops.
Sales confirms the billing contact once, in the customer system. The wires carry it to accounting and the project tool. The Tuesday paste and the Thursday retype are gone, and so is the spreadsheet between them.
Work moves on its own.
Delivery marks Phase 2 done. The AI looks up the contract value, the real one, drafts the invoice in accounting and tells finance. Nobody re-keys a number.
Who sees what is written down.
One line per fact, in a place anyone can check. People and the AI read the same line, and nobody has to remember who is allowed what.
Common myths
Four myths about getting data ready.
You might think
“Our data isn’t ready for AI.”
Actually
Ready doesn’t mean clean everywhere. It means one record with a home for each fact. Start with the record everyone retypes and leave the rest where it is.
You might think
“Connecting our systems is a big IT project.”
Actually
Most tools you already pay for have a connector built in. Wiring two of them is a small job. Start with the pair that passes the most facts by hand and leave the rest for later.
You might think
“Connecting AI means everyone sees everything.”
Actually
Only if the rule lives in a person. Put it on the fact and the AI reads as whoever is asking. The master key is what you get when nobody wrote the rule down.
You might think
“Spreadsheets are the problem.”
Actually
The gap between systems is the problem. The spreadsheet is where it shows. Close the gap and Excel goes back to being Excel.
Keeping it safe
Write down who may read what. The AI reads the same list.
The worry is fair. Connect an AI to everything and it feels like everyone can suddenly see everything. The fix is the same good sense you already apply to people, written down once, in a place the AI reads too.
01
The rule lives on the fact, not in the tool.
Whoever sets up the AI, and whichever tool they use, the lock is on the fact. Nothing gets set up around it.
02
Reading isn’t writing.
The AI reads within your role and can draft through the same wires. Changing a fact still goes through the person who owns it.
03
Every answer has a receipt.
Where it was read from, when that changed, and who it was read as. Anyone can check, so nobody has to take the AI’s word for it.
04
The record and its rules are yours.
Change AI provider next year, or add a fifth system. The wires, the record and the read rules stay where they are. Whatever arrives reads the same list.
How to start
Start with the record everyone retypes.
Nobody gets this working by cleaning every system first. They pick the one record that gets typed into the most places and start there.
Day one
Ask every department what one customer is called.
Pick one customer. Ask sales, ops, finance and people what it’s called in their system. Count the names.
Week one
Write down where each fact lives, and who may see it.
For that one record, one line per fact: which system holds it, who owns it, who may read it. A page, not a project.
Weeks two and three
Wire it, then prototype it on real data, with real roles.
Connect the two systems that pass the most facts by hand, put the record and its rules on top, and look at the new starter’s view before anyone else’s. That is the shape of an AI Jumpstart.
Next step
Stop guessing where AI belongs. We’ll show you.
AI works better when your data does. An AI Jumpstart is a short, fixed-scope project that starts there: where your facts live, which systems to wire, who may read what, proven with a working prototype on your real records.