Build it, buy it, or have it built

1. Yes, you can just use Claude

We publish every one of our build standards free, with a button that hands the whole page to Claude, ChatGPT or Copilot with the first message already written. That is not a loss leader with a catch in it. Plenty of businesses should take the standard, build the thing, and never speak to us.

So let us be direct about the competition. The most likely alternative to hiring anyone is the owner or the office manager who wires something up themselves in an afternoon. They are not being naive. The models really are good enough, the connectors really do exist, and a capable person with a clear head can get a long way before lunch.

The mistake is to argue that this does not work. It does. The question is a different one, and it only shows up later.

2. The spreadsheet test

You could run an entire business in Excel. People do. Excel is powerful, flexible and cheap, and it will happily hold your payroll, your job schedule and your debtors ledger.

Almost nobody does, past a certain size, and the reason is not that Excel became less powerful. It is that power is not the same as fitness. Excel was not built around what an auditor, an insurer or a partner needs to see before they sign something. There is no record of who changed what, no enforcement of the rule you agreed last year, and no way to prove afterwards which version was true.

A chat window has exactly the same shape. It is enormously capable and it was not built to be the thing your business depends on unattended. It does not know your after-hours rules, it will not refuse to book two jobs into the same slot, it has no opinion about the Do Not Call Register, and it leaves no log an insurer can read.

Which is fine, right up until the job is one where being wrong is expensive. Explaining a variance can be a guess. Balancing a ledger cannot. Neither can sending a van to an address.

3. Build it yourself

This is the right answer more often than a consultant will tell you. Take the route when all four of these hold:

• One system. The workflow reads and writes in a single place. The moment it has to keep two systems agreeing with each other, the difficulty goes up by more than the extra system suggests.
• Cheap exceptions. When it gets one wrong, somebody notices and fixes it in a minute, and nothing left the building in the meantime.
• A named owner. Somebody on staff whose job includes keeping it working. Not “whoever built it”, and not the owner at 10pm.
• No irreversible actions. It drafts, summarises, sorts or flags. It does not send money, commit to a customer, or ring a stranger.

Internal work is the sweet spot: sorting the inbox, drafting the weekly report, flagging the invoices that look wrong for a person to check. Start there. If it works and nothing breaks for a month, that tells you something real about the next one.

4. The failure nobody warns you about

The DIY route fails in a specific and repeatable way, and it is not the one people expect. It is not that the model produces rubbish. It is what happens on the second pass.

You build something. Somebody reviews it and finds a problem. You re-prompt to fix the problem, and what comes back is not the same thing with one correction. It is a new thing, structured slightly differently, which has to be reviewed from scratch. Do that three times and the reviewer has done four full reviews to land one change.

Accountants have been reporting exactly this loop with clients who build their own models: the client re-prompts, a different model comes back, and the review hours land on the professional who signs. The root cause is not model quality. It is regeneration without a fixed structure. No stable shape means no diff, and no diff means oversight can never be by exception. Somebody has to read all of it, every time, forever.

This is the practical argument for a written standard rather than a prompt. A standard fixes the shape: these checks, in this order, with these exceptions and this pass test. Then a change is a change rather than a fresh start, and a reviewer can look at what moved.

5. Buy it

Some steps are not worth hand-rolling at any level of skill, and our standards name them explicitly rather than leaving you to find out. The tell is always the same: the step involves the outside world, where the rules are somebody else’s and the mistakes are public.

Live phone calls to customers are the clearest case. A call has to disclose that it is an AI, respect calling hours, check the Do Not Call Register where it applies, capture consent to record, hand over cleanly when the conversation leaves the script, and keep a recording and transcript afterwards. None of that is hard to understand and all of it is tedious to get right, and getting it wrong is a regulatory problem rather than an inconvenience.

The tell is mechanical, so you can apply it yourself: if your agent can fake the step convincingly with a stub and still pass your own test, build it. If passing honestly would require you to hold a phone number, be the named party on a Do Not Call check, capture consent, or produce a log two years later, buy it. Today that means the after-hours phone, chasing invoices, and working a list you own.

Note what is not on that list. Your board, your queues, your reporting, your quote templates and your job taxonomy are all things you should build and keep. Anyone selling you those is selling you an afternoon’s work on a subscription.

Buying is also simply cheaper when someone has already solved it, and those three are already built. The after-hours phone is Booked After Hours, $119 a month. Chasing open quotes and overdue invoices is Get Paid, $199 a month, for builders and for accountants’ clients. Working a list you own is workmylist, $200 a month per agent. Flat monthly, no call caps, so you can compare each against the cost of your own time.

6. Have someone finish it

Three conditions, any one of which is enough to want help. None of them is a reason to hand the whole thing over.

It spans more than one system. A booking that has to exist in the job system, appear on a calendar, and not contradict the accounting package is three integrations and a reconciliation problem, not one automation. Vendor connectors make the connection cheap: a chat can now create the job. They do not make it hold. The job id has to be the one the quote, the invoice and the chase read, and a second record has to fail.

The exceptions are the job. In plenty of workflows the ordinary case is trivial and the whole difficulty is the minority of cases that are unusual. If listing the exceptions takes longer than describing the happy path, the happy path was never the work.

A pass test will not go green. You built it, it nearly works, and disclosure or write-back or the escalate path will not come good. This is the common one, and it is worth an hour of somebody else’s time rather than a fortnight of yours.

What that should cost you is a fixed price against the specific thing that is broken, quoted off the failing test. What it should not cost you is a monthly fee to have somebody stand next to a build you own. Be wary of anyone, us included, who answers “I am stuck on one step” with a management retainer. The workflow is yours. If it needs maintaining every month, it was built wrong.

What you should expect from anyone who does finish it for you: the workflow written down before it is built, the actions it can take without asking scoped narrowly, look-then-confirm on anything touching money or the schedule, and an audit pack as standard rather than an extra.

7. Six questions that decide it

1. If this gets one wrong at 6pm on a Friday, what has happened by Monday? An apology, or money gone and a customer lost?
2. Does it talk to anyone outside the business? Customers, suppliers, tenants. Outside means rules.
3. How many systems does it have to keep in agreement? One is a project. Three is a different kind of project.
4. Who owns it in six months, by name? If there is no name, the answer is not build.
5. When someone asks why it did that, can you show them? A log, not a recollection.
6. Does a standard already exist for it? If yes, most of the thinking is done and the build is smaller than you think.

Mostly “apology, one system, named owner” means build it. Any “rings customers” means buy that step. Any “three systems” or “no name” means have it built.

8. What each route costs

Building it yourself costs an AI subscription and somebody’s attention. The second one is the real price and it is recurring, not one-off. Budget for the maintenance, not the build.

Buying is a published monthly figure, and in this market the advertised number hides setup fees and per-minute overages. The cost paper shows how to compare quotes properly.

Getting unstuck is a fixed price against the step that is failing, and the $500 audit is credited against any work that follows, so the route that involves us most still begins with the smallest commitment. Our figures are on the pricing page.

9. Where to start, whichever you pick

Write the workflow down before anybody builds anything. Not a prompt: the checks in order, what counts as an exception, who handles each exception, and a test that says whether it worked. That document is the thing that makes the difference between a build you can supervise and one you have to re-read every time.

We publish 24 of those already, free, and they are written to be handed to an AI coding agent as instructions rather than read as argument. If one of them covers your job, you are most of the way there before you have decided anything about who builds it.

And keep a list of what you will not automate. Complex first-time variations, safety assurances, anything implying a licence the AI does not hold, price changes without a person, and legal or tax advice. That list is not timidity. It is the thing that makes the rest of it trustworthy.