The other end of the financial year list
In April we wrote about an old system at the start of a new financial year — what to fix before you add more. Half a year later the honest question is not “what else could we innovate”. It is which of those decisions quietly doubled. Three of them compound whether you notice or not: who holds the source, whether you have one system of record, and whether the first slice had a number. Everything else is commentary.
This closes that H1 arc. It does not reopen a manifesto about transforming an industry. If you want a pep talk, the site already has one and we are not going to rewrite it. If you want to know why September feels heavier than April, read what to fix when the year is new and the software is old, then this. The cost bands for the next slice still sit in how much custom software costs. We will not print a rupee total here that will be wrong for your connector.
Six months is long enough for a spreadsheet to become the real system, for a vendor to become the only person who can deploy, and for a pilot to become furniture. It is also long enough to have been right. If you hold the source, one record of an order, and a measured slice, you are in a different year than the firm that bought a platform login in May and is still reconciling in WhatsApp.
Source you do not hold
In April “we should get the source” sounds like paperwork. In September it is a negotiation, because the system has grown and the only build machine is a laptop you do not own. Compounding is the extra features that landed in a repo you cannot clone. Each of those features is a reason the vendor will quote a migration as a new project — which it is, if you start from a zip.
Hold means: organisation account, history, the ability to build what is in production, and a second person who has done it. A copy from June that nobody compiled is not hold. An escrow you have never tested is not hold. An employee who will only push from home is not hold. The April version of this sentence was cheaper. Pay it in September anyway. It will not be cheaper in March.
Hosting and registrar compound the same way. A personal card that still pays the VPS is a resignation away from a dark site. If you moved hosts in H1 and the DNS still sits on a free mail, you have a September job that looks like an April job, only now with festival traffic in the log.
If a partner still holds production-only knowledge — the real connection string, the certificate trick, the Tally flag — you have been renting architecture. Write it down or buy the week that writes it down. IT systems consulting is sometimes exactly that week: not a new system, a picture of the one you already depend on.
One system of record, or a reconciliation career
Two systems of record for the same fact will not stay “temporary”. They grow a third — usually a sheet — because neither system is trusted for the argument. After six months the sheet has columns the software does not, and the software has statuses the sheet ignores. Staff become the integration. That is the compounding: not technical debt as a slogan, but hours that now exist only because you refused to name the record in April.
Pick the fact. An order number is invented in one place. Stock is decremented in one place. A GST invoice is posted in one place. Everything else reads. If Tally is the invoice, the website is not allowed to invent a second series. If the custom app is the job card, WhatsApp is not allowed to be the real job card. Saying that in April is a meeting. Saying it in September is a cleanup.
Hybrid is allowed. Counter sales in a product, the bay in a custom system, accounts in Tally. Hybrid is not the same as three writers for the same qty. The April post asked you to notice the sheets. The September test is whether a new sheet appeared beside the system you shipped. If it did, the slice was wrong or the trust was. Do not add a fourth writer to paper over that. Measure, then cut.
Permissions compound here too. A shared admin that was “fine for go-live” is, by September, a former staff login and a branch that can see another branch’s rates. The boring security list from last month is the operational version. The compounding version is: every month you postpone named accounts, the revoke job gets larger. Do it while the user list still fits on a page.
The first slice you measured — or did not
A slice without a number becomes a platform by accident. In April you could say “warehouse will record dispatches here instead of WhatsApp”. In September you either have a percentage or you have a feeling. Feelings fund the next module. Percentages fund a fix or a stop. That is the whole of compounding on scope.
If you never wrote the sentence, write it now for the thing that is live. Then look at two weeks of real use. Staff complete that job in the system most of the time, or they do not. “They said it looks nice” is not a pass. If they do not, the expensive lesson is available: wrong job, or not a software problem. A nine-month build that learns the same thing in Q4 is just the April mistake with interest.
If you did measure and the number moved, protect that slice. Do not bury it under reports a director asked for once. Compounding works in your favour when the trusted path stays narrow and used. It works against you when every success is an excuse to add an unmeasured second path.
AI features follow the same rule. A pilot you kept without change-rate and hours is furniture that now has a bill. A pilot you killed is a lesson you can reuse. See the August post on keep-or-kill if that button is still sitting on a screen with no owner. Do not open a second model on the same unmeasured job.
People leave. The decisions stay.
H1 is long enough for a developer to resign, a vendor to rotate the bench, and a finance lead to hand over. What compounds is not the resignation. It is the architecture that was never written down, the backup that was never restored, and the exception rules that lived in one head. September is when you find out whether April’s “we will document later” was a plan or a wish.
If someone who mattered left, run the handover test even if you are not angry. Clone, build, deploy to a non-prod slot, restore a backup, change a rate, undo a dispatch. Where that fails is the project. It is cheaper than a surprise in January when the same person is not taking calls.
If nobody left, pretend they will. The sheet from the boring security list, the architecture page, the named exception owner — those are September maintenance, not a new programme. Firms that only write these down after a resignation always write them in a hurry, and the hurry misses the Tally flag.
Cost is now a trail
April’s quote was a shape: discovery, a slice, assumptions. September’s number is a trail: what you added, what you deferred, what you paid in staff time to keep a sheet alive. If the trail and the quote cannot be put in the same table, you do not have a vendor problem only. You have a scope that was never a slice.
You do not need a rupee figure in this article to see the shape. A low first number with empty cells for migration and handover is now a change-request pile. A measured slice with a change log is a pile you can defend. The cost guide is how to read the next estimate. This post is why the next estimate should start from what you actually run, not from the April slide.
Year-three cost was on the mid-year review for a reason. Licences, a connector that needs a person, a host that grew because nobody archived. Six months is early for year three and late enough to see the line. Cut one unused login. Archive one environment. Name the connector owner. Those are compounding in reverse.
What to reopen this month
| April question | September test | If you fail |
|---|---|---|
| Who holds source and host? | A second person builds and deploys what is live | Buy the week that makes that true |
| What is the system of record? | One writer per fact; no new sheet beside the app | Stop the second writer before the next module |
| What did the first slice prove? | A number you wrote in advance still exists | Do not start slice two |
| Who owns exceptions? | A named person, still in the building | Name one, or switch the feature off |
Four rows. You can run them in a morning. Consulting is useful when two people in the room disagree on what the record is. Building is useful when you already agree and the software does not match. Do not buy building to settle the disagreement. That is how H2 becomes another H1.
Close the half-year without a slogan
If the four rows pass, you are allowed a second slice. Keep it as narrow as the first. Write the assumption. Measure. Hold the source. That is the whole method, and it is enough for a year.
If they fail, September is a repair month, not a launch month. Repair is unfashionable while competitors post about platforms. Repair is also why you still have a Tuesday process. The firms that skip it will spend next April writing the same list with worse constraints — a vendor who knows they are the only build, a sheet that has become policy, a pilot nobody will kill.
We will keep saying this in plain language because the other version of this ending is a paragraph about innovation that does not name a record. You already know whether the dispatch is in the system or in a group chat. Name it. Fix the one that compounded. Then stop. The year still has a half to run, and the useful work is the next measured slice, not a new category of software you do not own.
If you want a partner for the picture, start with systems consulting. If you want a partner for the .NET slice after the picture agrees, that is a build with a definition of done. Either way, bring the September table, not a blank “we need to digitalise”. The table is the brief. The brief is how you avoid paying twice for the same six months.