The folder of ideas that were quoted and then dropped usually holds the same things: an internal tool, an integration between two systems that never justified the build, an AI pilot that looked good in a vendor demo until someone priced running it properly. Each failed a cost-benefit test in 2023, and the test was fair. Run it again today and some of them pass, because the cost of building has fallen.
Why the 2023 no made sense
In 2023, building anything with AI meant absorbing several costs at once. Model access was expensive per call, especially for anything beyond simple classification. The tooling to connect a model reliably to your existing systems — your CRM, your inbox, your booking system — barely existed, which meant most of the work was bespoke integration, built by a developer or an agency, on a timeline measured in months. And once it was built, someone had to maintain it: watch the model's outputs, patch the connections when an API changed, own the thing indefinitely.
Weighed against that, most shelved projects were reasonable decisions. The value on offer — faster response times, less manual data entry, a slightly better customer experience — didn't clear the bar set by months of specialist build time and an open-ended maintenance commitment. Saying no wasn't caution for its own sake. It was correct, given what building actually cost.
What's changed since
Three things moved, and they compound.
Model pricing has come down repeatedly and by a wide margin since 2023, as providers have competed on cost per token rather than just capability. Tasks that used to require a large, expensive model — drafting, summarising, classifying, extracting structured data from messy text — now run on smaller, cheaper models that are good enough for the job.
The orchestration layer has matured. Tools like n8n, Zapier, and various agent frameworks now let you wire a model into your existing systems without writing a custom integration from scratch. The plumbing that used to be the majority of the build — connecting the model's output to your CRM, your ticketing system, your accounting software — is increasingly configuration work rather than software engineering.
And the skill required to prototype has dropped. A technically curious generalist can now put together a working version of an idea in days, using off-the-shelf tools, to find out whether it actually works before anyone commits to a proper build. In 2023 that same exploratory step usually required hiring a specialist or briefing an agency, which meant you paid discovery-phase costs just to find out if the idea was worth pursuing.
None of this makes AI free, and none of it makes engineering judgement optional. What it does is lower the cost of finding out — and for a lot of shelved projects, the original no was really a no to the cost of finding out, not a no to the idea itself.
How to re-test a shelved idea
The practical move is to go back through the list and ask a different question than the one you asked originally. Not "would this be useful" — you already answered that, or you wouldn't have considered it in the first place. Ask: does this idea depend on the cost structure that existed when you shelved it?
Some will. A project that needed deep integration with a regulated core system, or that required guaranteed accuracy in a compliance-sensitive workflow, still needs careful engineering and proper oversight — the model getting cheaper doesn't remove that work. If the original no was about risk rather than cost, it may still be a no.
But a lot of shelved ideas were cost calls dressed up as strategic ones. The customer support triage tool that didn't justify a developer's time. The lead-qualification workflow that would have needed a specialist to build and babysit. The internal reporting tool that got quoted at a price nobody wanted to defend to the board. These are exactly the categories where the underlying task — reading unstructured text, extracting the useful bit, routing it somewhere sensible — is now the cheap part of the problem rather than the expensive one.
The way to find out is to prototype before you re-scope. Build a rough version with the tools available now, test it against real inputs from your business, and see whether it actually does what the original pitch claimed. This is a materially different exercise than the one you went through in 2023, because the cost of being wrong is now a few days of someone's time rather than a commissioned build.
Where this still goes wrong
Cheaper prototyping doesn't mean cheaper production. A workflow that works well in a two-day test can still fail once it meets real data volume, edge cases, and the parts of your business that don't follow the happy path. The gap between "this worked in the demo" and "this runs unattended in production without someone checking it" is still real, and it's still where most AI projects actually earn or lose their keep. Cheaper tools have lowered the cost of testing an idea — they haven't lowered the cost of running it responsibly at scale, and treating those as the same thing is how a promising prototype turns into an unmonitored liability six months later.
Not every shelved idea deserves a second look. Some were rejected because the business case was thin, not because the build was expensive. Cheaper AI doesn't fix a weak use case — it just makes it faster to discover that it's still weak.
The practical takeaway
If you have a list of AI or automation ideas that got quoted, considered, and set aside, it's worth pulling that list out again — not because AI has become magic, but because the specific cost structure that killed those ideas has changed in ways that are easy to underestimate from the outside.
The way to test that isn't to commission another proposal. It's to build a rough version, cheaply, and see what it actually does against your real data.
If you want a second set of eyes on that list, get in touch. We're happy to look through it with you and tell you plainly which ones we think are worth a prototype and which ones probably aren't.