Skip to content
AI explainer series

The software you’d actually want
has never been for sale.

Anyone who runs a team has a list of what they’d change about their software. No vendor has ever built it, because it’s only true of you. Here’s why that piece was never for sale, why building it stopped being a big-company move, and where AI fits once it’s yours.

How the work runs today
Customer systemHolds the customer and the price listNorthstar Group Ltd.$86,000
  1. 1Gather the detailsAn email thread
  2. 2Price itA spreadsheet tab
  3. 3Approve itIn somebody’s head
  4. 4Send itAn email thread again
  5. 5Chase itA reminder somebody sets

Nobody sells this piece

AccountingHolds the invoice and what was paidInvoice 4470Paid
Two bought systems, one holding Northstar at $86,000 and one holding invoice 4470, paid. Between them, a gap in the exact shape they leave, and five steps floating in it with nothing underneath.
Two grey puzzle pieces that do not meet. One is the customer system, holding the customer and the price list, and a record reading Northstar Group Ltd. at eighty-six thousand. The other is accounting, holding the invoice and what was paid, and a record reading invoice 4470, paid. Where their edges fail to join, a gap is outlined in a dashed line, in the exact shape they leave. Five step cards float inside it with nothing underneath: gather the details, an email thread; price it, a spreadsheet tab; approve it, in somebody’s head; send it, an email thread again; chase it, a reminder somebody sets. A label under them reads nobody sells this piece.
Why it happens

Software you buy has to suit everyone who buys it.
What only you do gets left out.

A product is built for every company that might buy it, so what goes in is what all of them share, and what only you do gets left out. The settings menu is how far it bends, and there’s no switch on it for who signs a quote over the limit on a Friday afternoon.

One product, sold to everyone who buys it
What every customer shares
  • Customers, contacts, invoices
  • The screens they drew
  • Twelve reports
What you can change
  • Rename a field
  • Hide a tab
  • Set who sees what
The real menu is longer. It still ends at the wall.
What only you do
  • Gather the details
  • Price it
  • Approve it
  • Send it
  • Chase it

No setting in there reaches any of these

Inside the box, what every customer gets: customers, contacts, invoices, twelve reports. On the wall, three switches, all already on. Outside, the five steps only you run.
One product drawn as a box. Inside it, behind a wall, what every customer gets: customers, contacts and invoices, the screens they drew, twelve reports. On the wall, the settings menu, and it is three switches long: rename a field, hide a tab, set who sees what, all three already on, with a line under them saying the real menu is longer and still ends at the wall. Outside the box, the same five step cards from the first picture, drawn with dashed edges, under a line reading no setting in there reaches any of these.

Two companies can run the same three products and work nothing alike, and the difference is all outside the box.
Ask the vendor for it. Here’s what happens.

The roadmap

Every business has a list of things it asked for.
Somebody else sets the date.

Ask a vendor for the part that’s yours and it joins a roadmap every other customer is writing to, ranked by how many of them asked. At a company we worked with, the top of that list was approving a batch of quotes in one go, first asked for in 2023. Still a maybe.

  1. Approve a batch of quotes in one go20233 years
  2. One view of a customer across the three systems20242 years
  3. Send the invoice from the signed quote20251 year
Three requests, first made in 2023, 2024 and 2025, each bar as long as the wait behind it, and the date column empty on every row.
A list of three requests made to vendors, each with the year it was first asked and how long the vendor has had it. Approve a batch of quotes in one go, first asked in 2023, three years. One view of a customer across the three systems, first asked in 2024, two years. Send the invoice from the signed quote, first asked in 2025, one year. Each request carries a bar that grows with the wait, and a third column, the date the vendor gave, holds an empty dashed slot on every row.

Three years is what a request costs when somebody else ranks the list.
So why did nobody just build it?

Why nobody built it

Building your own was priced in 2006.
Nobody has priced it since.

In 2006 a business of a few hundred people had two choices: five developers for eighteen months, or a subscription. Only one was affordable, so it bought the product built for everybody and kept a spreadsheet for the rest. Then AI started writing much of the code, and two developers now do in four months what five did in eighteen.

To build the same thing
When it was decided2006
Five developers
Eighteen months

Cheaper parts every year. Then AI started writing the code.

The same thing today2026
Two developers
Four months
Five developers and eighteen months in 2006. Two and four in 2026, drawn in the same places, so what is no longer needed is left as an empty outline.
The same job, counted twice. The first count is 2006, when the decision was taken: five developers, drawn as five figures, and eighteen months, drawn as eighteen blocks in a row. The second count is 2026: two developers and four months, drawn in the same eighteen and five places, so the three figures and the fourteen months that are no longer needed are left as empty outlines. Between the two counts runs a rail of twenty-one marks, one for each year, and not one of them carries a count. Buying is not in this drawing, because only the cost of building is being counted here.

Ninety months of work is now eight. About $100,000 once, plus hosting, what the AI uses and somebody to keep it current, on top of the $60,000 a year in subscriptions, because those systems are worth keeping.
AI turned up inside the products too, and not where anyone asked for it.

Then AI arrives

The AI in the corner of a product
can only see that product.

Most products now come with a chat box in the corner, and each one reads only the records its product holds. The question that runs your Monday crosses three of them: what’s waiting for approval, what does it bill this month, who’s delivering it. Each box answers the third it can see.

What is waiting for approval, what does it bill this month, and who is delivering it?

  1. Customer systemNorthstar Group Ltd.Quote, $86,000
    What it can answer
    • What is waiting for approval

    It holds the quotes. It’s never seen an invoice.

  2. AccountingNorthstar Group Ltd.No invoice yet
    What it can answer
    • What it bills this month

    It holds the invoices. A quote nobody signed never reached it.

  3. Project toolNorthstar Group Ltd.No job typed in
    What it can answer
    • Who is delivering it

    It holds the people, and only once somebody typed the job in.

One question in three clauses, put to three products. Each one ticks a single clause and cannot answer the other two.
One question, put to three grey product boxes: what is waiting for approval, what does it bill this month, and who is delivering it. Each box holds one record for the same customer, Northstar Group Ltd.: a quote at eighty-six thousand in the customer system, no invoice yet in accounting, and no job typed in to the project tool. Each box has a small chat bubble with a sparkle in its corner. Under each box the same three clauses stand in the same order, and each box marks the one it can answer. The customer system ticks what is waiting for approval and strikes out the other two, because it holds the quotes and has never seen an invoice. Accounting ticks what it bills this month, because a quote nobody signed never reached it. The project tool ticks who is delivering it, and only once somebody typed the job in.

Three corners, three answers, and the question you asked sits in none of them.
The AI is fine. So who decides where it stands?

Where AI plugs in

AI is only as good as where it’s plugged in.
In a bought product, somebody else chose the spot.

Three decisions make an AI useful or useless, and none is technical: which step it sits on, which facts it reads, and who checks the result. In a bought product all three were made for every customer at once. In software you own they’re yours, and most systems already have a door built for other software to read through.

One step, wired the way the job needs it
  1. Price listCustomer system
  2. Last three quotesCustomer system
  3. Margin floorAccounting
2Price itdrafts
  1. $86,000drafted
  2. Over the limita person signs instead
One step. Three facts it reads, each named with the system it comes from, one draft it writes, and one rule for when a person signs instead.
One step from the quoting work, price it, drawn large with the AI chip on it and the word drafts. Three facts feed into the step along dashed lines, each named with the system it is read from: the price list and the last three quotes from the customer system, and the margin floor from accounting. Out of the step, along a solid line with an arrowhead, comes one draft, a quote for eighty-six thousand. A second line leads to a person with a rule: over the limit, a person signs instead. Once seen, a bead travels each fact into the step in turn, the AI chip breathes once, the draft appears, and then the route to the person.

Move the plug, and the same AI stops answering a third of the question and starts drafting the quote.
That’s the piece nobody sold you. Here’s what it looks like built.

The fix, part one

Keep the systems.
Build the thin part on top.

Nothing moves. What gets built is a small piece of software of your own: a site your team signs into, a screen for each step in between, AI on the steps that repeat. It keeps no records of its own and asks the systems below for every fact it shows, so there’s no migration.

Your own software, on top of the systems you keep
The steps in between
  1. 1Gather the detailsreads
  2. 2Price itdrafts
  3. 3Approve itA person
  4. 4Send itA person
  5. 5Chase itreminds
  1. Customer systemCustomers, pricesNorthstar Group Ltd.$86,000
  2. AccountingInvoices, paymentsInvoice 4470Paid
  3. Project toolJobs, people, datesNorthstar installTwo people
AI on three steps, a person on two, and every fact asked of the system that holds it. One quote for $86,000 walks all five.
Three grey system boxes, each holding one record: the customer system with customers and prices and the Northstar quote at eighty-six thousand, accounting with invoices and payments and invoice 4470 marked paid, the project tool with jobs, people and dates and the Northstar install staffed by two people. Above them, a thin band labelled the steps in between. The five steps sit on the band in order. AI takes the three that repeat: it reads the details in, drafts the price off the price list, and sends the reminders. A person approves and a person sends. Dashed connectors run between the systems and the band, which is how the band asks for every fact it shows. One quote for eighty-six thousand travels the five steps in turn. Gathering the details asks the project tool, pricing asks the customer system, and chasing asks accounting; each time, the record in that system lights up and a mark travels the connector.

The five steps that lived in an email thread, a spreadsheet tab and somebody’s head now have a home, and it’s a thin one. What changes is the screen each person opens in the morning.
And nobody needs the same one.

The fix, part two

One piece of software.
The screen each job always wanted.

Everyone’s working on the same quote. Operations wants a queue. Finance wants an invoice run. The CEO wants one line. None of it was ever coming out of a settings menu, because each screen is one job’s view of steps the product never had, and all three read the same record from the systems below.

  1. Operations
    Waiting for approval
    • Northstar Group Ltd.$86,000
    • Two more quotesin the queue
    Ops approves
  2. Finance
    This month’s invoice run
    • Northstar Group Ltd.$86,000
    • Ready to sendonce approved
    Finance sends
  3. The CEO
    This week, in one line
    • Northstar Group Ltd.$86,000
    • Three quotes waiting$214,000 in all
    One number, three screens
The same piece of software, under all three
One quote at $86,000, read three ways: an approval waiting, an invoice ready to send, and one of three lines adding up to $214,000.
Three screens reading from the same thin band. The operations screen shows a queue of quotes waiting for approval, with Northstar Group Ltd. at eighty-six thousand at the top. The finance screen shows this month’s invoice run, with the same quote at the same figure, ready to send once approved. The CEO’s screen shows the week in one line each, where that quote is one of three waiting, totalling two hundred and fourteen thousand. The figure lands in all three at once.

Same facts on all three screens, and not one of them was built for somebody else’s job.
Then the work changes, because it always does.

Over time

The same list, on your side of the table.
Week twelve, not year three.

The list of things you asked for doesn’t go away. It changes hands. The people who do the work rank it, the developers take the top of it into the next release, and it ships on a date with a note and one training. The 2023 request goes out in week twelve.

  1. Approve a batch of quotes in one go2023Week 12Release 1.1
  2. One view of a customer across the three systems2024Week 6Release 1.0
  3. Send the invoice from the signed quote2025Week 17Release 2.0
The same three requests, on a list the company owns: one view of a customer in week six, batch approvals in week twelve, the invoice in week seventeen, each trained once.
The same list of three requests as before, now with the week each one shipped and the release that carried it. Approve a batch of quotes in one go, asked in 2023, shipped in week twelve as release 1.1. One view of a customer across the three systems, asked in 2024, shipped in week six as release 1.0. Send the invoice from the signed quote, asked in 2025, shipped in week seventeen as release 2.0. Each bar is as long as the weeks it took, and where the empty date slot was there is now a release.

Same three requests, same words. The only column that changed is who set the date.
Here’s what’s different on a Monday.

What changes

Same systems, same records, same people.
Only the software in front of them changed.

What moved is who decides what the software does next, and where the AI is allowed to stand.

The AI reads all three systems.

On the step you chose, with the facts it needs and a person on the check. One answer to the Monday question, instead of a third of one from each corner.

The list is yours to rank.

A request goes to the people who do the work, and what ships next is decided by them, not by how many other customers asked for the same thing.

The data and the instructions stay yours.

Records stay in the systems that hold them. The code and the written instructions stay with you, whichever AI you’re using next year.

Common myths

Four myths about building your own.

You might think

“Waiting for our vendor is the safe option.”

Actually

Waiting is a decision too, and it’s the one you control least. Somebody else decides what goes on that list, somebody else decides the order, and the date, when it comes, is theirs to set.

You might think

“The right tool is out there. We just haven’t found it yet.”

Actually

Shortlists get long when nobody has written the job down. Write one sentence of it and most of the list answers itself. The same goes for the automation tool you may already pay for: it copies a fact from one system to another well, and it has nowhere to put who approves, what happens when a number looks wrong, or what the person doing the work sees.

You might think

“Custom means rebuilding the system that runs our finances.”

Actually

Accounting carries on being accounting, and the customer system carries on being the customer system. The piece on top holds the steps they never covered and asks them for every fact it shows. It doesn’t store one record of its own, so switch it off tomorrow and everything is exactly where it was.

You might think

“We could build this ourselves over a weekend with an AI tool.”

Actually

A weekend gets you something one person can use alone, which is genuinely useful. The moment four people are on the same deal you need control over who can see and change what, one place for the facts to live, and something that outlives whoever built it. That’s a different job.

Keeping it safe

The part you build stays thin.
All of it stays yours.

The fear about building is rarely the build. It’s the year after: a thing only the people who made it understand, nobody inside the company deciding what it does next, and nobody sure who keeps it running. That’s settled by four decisions, all of them taken before anybody writes anything.

Never rebuild what already works.

The systems that hold your records keep holding them. The piece on top holds only the part they never covered, and asks them for every fact rather than keeping a copy. If it can be bought, don’t build it.

Ship the changes in numbered releases.

A memo is how a process changes. A release is how software does: group the changes, send one short note of what changed and why, and train people once on that note. Not a memo on Tuesday and a correction to it on Thursday.

Own the code and the instructions.

Both, in writing, whoever builds it. Keeping it running is an occasional job, not a department: when a system underneath changes, whoever picks it up next can read the instructions and fix it. The records never left your systems.

One person inside the company owns the list.

Not the agency, not the vendor. Somebody who does the work decides what goes in the next release, and nobody has to be talked into it.

How to start

Start with the request
that’s been waiting longest.

Not because it’s owed to anybody. The request that has waited longest is the one people have already built a workaround for by hand, so they can tell you exactly what the software has to do before anybody designs it.

  1. Day one

    Write down everything you’ve asked a vendor for, with the date.

    The date is the column that does the work, not the feature. If no such list exists, an hour with your department heads usually builds the whole thing.

  2. Week one

    Draw the whole thing on one page.

    Anything that holds records stays where it is. Anything that is a step between them goes on top, with where AI sits and who approves. Four boxes and some arrows, not a specification.

  3. Weeks two and three

    Prototype the first release on real records.

    One job, your own data, and the people who do that job watching it run. That’s the shape of an AI Jumpstart.

Next step

Stop guessing where AI belongs.
We’ll show you.

AI works better when you choose the step, the facts and the check. An AI Jumpstart is two to four weeks, fixed scope, and starts there: what your systems hold, which steps live between them, and what the first release would be, proven with a working prototype on your real records.