I'm curious how other business owners approach this decision. Off-the-shelf software is usually the obvious starting point, but sometimes a company ends up changing its procedures to fit the software rather than the other way around. After adding workarounds, spreadsheets, manual data entry, and integrations, the original solution may not be as simple or inexpensive as it first appeared. At what point do you think it becomes worth considering a custom solution? Is it mainly the cost, the time saved, or the need to support a workflow that existing products cannot handle well?
In the era of AI, it's a no-brainer. With effective prompting, you can build custom, lightweight software that does exactly what you need. Existing solutions are bloated because they try to satisfy everyone. If you only use 10% of their capabilities, why load a thousand files to use a hundred? Go with a custom AI solution.
I’m not sure I would call it a no-brainer yet. AI can certainly help build parts of a custom application, but building something that actually matches a company’s workflow can be a different matter. For example, take a payroll/timekeeping application that imports badge records, compares them with employee timesheets, applies company-specific overtime and approval rules, handles exceptions, and produces a report in the format the company needs. I’d be genuinely interested to hear how you would approach something like that with AI. What would you expect AI to handle automatically, and where would you still need a developer to design and maintain the system?
You're making a very fair point: AI isn't a magical, "prompt it and forget it" replacement for software development, but it is underrated in its ability to write, integrate and execute complex business logic. It will do 90% of what you need; you just fill in the remaining 10%. What AI will do: 1. Generate the complete data pipeline. 2. Build the rule engine. 3. Draft tests and documentation. Where you come into the picture: 1. Oversight. 2. Fill in the specific details. 3. Human verification (where needed). P.S. I am not trying to convince you that AI will do 100% of the work, it still needs you. But if you have software development skills, you can make magic happen. Off the shelf, bloated enterprise solutions are costly and can become a nightmare to maintain down the road.
That's exactly the distinction I had in mind. I agree AI can handle a lot of the implementation work, especially when the requirements are clearly defined. The difficult part, in my opinion, is defining all the exceptions and business rules correctly. For example, two companies might use the same badge system but have completely different overtime rules, approval processes, and ways of handling missing or conflicting records. A payroll validation application that compares badge records with employee timesheets and applies company-specific rules is a good example of this type of challenge. The code itself may not be the biggest issue; making sure the system reflects how the business actually operates can be.
My formal career was with a company that took existing software that we knew inside out and making it fit our client's requirements. The per-seat fees made my eyes water - I doubt any of us earned the equivalent of 2 seats! My freelance career ended up with a single non-profit client and their CRM/CMS. It was only created because we asked a company to build it, they failed and I stepped up from adviser to doing it all. Over a 15 year period we frequently went "back to market" to get a commercial product to migrate to. Nothing existed. 5 years on, and they're still looking. Part of me regrets not building my own and selling it as a service; there is a market. I don't know how the other organisations are muddling along. I'm currently working on an in-house system and regretting not using something like CakePHP for it. There's nothing off the shelf that would do what we want but starting from nothing was a bit silly.
The real answer right now is: it depends. And that is a big change from just 5 years ago, when the clear answer was: buy it, don't build it. AI has changed that answer, but it has not yet completely become build it, don't buy it. It will likely change to that answer in the months and years ahead. I started my software sales career back in the days of mainframes and minicomputers in which most software was custom-written and the processing took place in a machine far from the user (anyone besides me remember time-sharing services offered by companies like CDC?) Then came workstations and eventually PC's where the processing took place at the user's desktop. Now, with the Cloud, we have come full circle with the only remaining piece being the switch back to primarily custom software. Of course, that, too, will change as the hottest discussion topics in hardware engineering circles these days is around physical AI, in which the model runs locally and not in the Cloud.
It comes down to whether the software supports your core competitive advantage. If a process is just standard admin (payroll, basic CRM), adapt to off-the-shelf. But if a workaround slows down the actual service/product you sell, build custom. The moment spreadsheets start replacing your main database, you've crossed the threshold.
That's a good way to look at it. Once spreadsheets start replacing the main database, the business is probably spending too much time working around the software. A real example of this can be seen in timekeeping, where companies may need to compare badge records with timesheets and apply their own overtime and approval rules. Standard software often handles the basic process, but the exceptions are where things become more complicated.
With AI now, I'd say very quick. Like a couple months. If I started now I would custom build everything. Because I start in 2001, I'm stuck in old software that keeps raising the prices and adding more up-sells than making it run better.
It purely depends on your needs. If you think the existing software doesn't serve your purpose then build a custom one. However given that a start-up has frugal resources it make more sense to limit your focus on the problem statement you are trying to address then trying to work on everything.
A good middle ground is to start with existing software and customize only where the workflow really needs it. Once the workarounds start taking more time than the software saves, a custom solution can be worth considering.
You'd still have to know coding, because if AI messes up somewhere you should be the one to rectify it.