№

§ Comparison

Custom AI vs off-the-shelf software.

Cost curve, time to value, accuracy ceiling, data ownership and lock-in — compared without a sales agenda.

↓ scrollitresearches.com

The question is not which approach is better. It is which one is better for a specific decision, at a specific volume, on a specific dataset. Off-the-shelf software buys you the average of everyone else's problem; custom AI buys you the shape of yours. Both are correct answers in different rooms.

The short version: buy when the process is standard, the data is unremarkable and speed dominates. Build when the decision is close to your margin, your data is genuinely proprietary, or a generic model's accuracy ceiling sits below the threshold at which the system starts paying for itself. Most mature estates end up doing both — commodity layers bought, the thin differentiating layer built.

Total cost of ownership over five years
Time to first measurable value
Accuracy ceiling on your own data
Data residency and model ownership
Integration depth with existing systems
Exit cost and vendor lock-in

§ 01

The five-year cost curve is the whole argument

Most build-versus-buy debates compare a purchase price against a project quote, which compares the wrong things. A SaaS licence is an operating cost that grows with headcount or transaction volume; a custom system is a capital-shaped cost followed by a small maintenance line. Plotted over five years, they cross — and where they cross is decided by scale, not by preference.

As a rough guide: below roughly fifty users or a hundred thousand decisions a month, off-the-shelf is normally cheaper for the full period. Above that, the licence line keeps climbing while the custom line flattens, and the crossover usually lands somewhere in year two to year four. Any vendor or consultancy that will not draw this curve with your actual volumes on it is selling, not advising.

§ 02

Accuracy ceilings are set by data, not by vendor

Off-the-shelf models are trained on a distribution that is an average of many customers. If your data resembles that average, the generic model will be excellent and no amount of custom work will beat it by enough to matter. If your data is unusual — a rare defect class, a regional dialect, a product hierarchy nobody else uses, a regulatory field structure specific to your market — the generic model hits a ceiling that no configuration screen can lift.

This is testable before you commit. We routinely run a two-to-four-week evaluation that scores an off-the-shelf option and a quickly trained baseline on the same held-out slice of your real data. The result frequently ends the debate in one direction or the other and costs a fraction of either commitment.

§ 03

The hybrid answer, and when it is a cop-out

The recommendation we make most often is hybrid: buy the commodity — OCR, transcription, general-purpose language models, vector storage, observability — and build only the thin layer that encodes your specific advantage. That layer is usually small, often a few thousand lines and a well-curated dataset, and it is the only part a competitor cannot also purchase.

The hybrid answer becomes a cop-out when it is used to avoid a decision. If the differentiating layer cannot be described in a sentence, there is no differentiating layer, and you should buy. If it can, and you buy anyway, you have outsourced your advantage to a supplier who will sell the same capability to your competitor next quarter.

What you get

Outcomes, not just outputs

/ 01

Cost shape, not cost number

Custom is front-loaded and cheap to run; SaaS is cheap to start and scales with your success. The crossover typically falls between year two and year four.

/ 02

Accuracy ceiling

A generic model is trained on someone else's distribution. On unusual data the ceiling can sit 10–30 points below what a model trained on your own history reaches.

/ 03

Speed to value

Off-the-shelf can be live in days. Custom takes weeks to months — but the value keeps rising afterwards rather than plateauing at the vendor's roadmap.

/ 04

Ownership

Custom means you hold the weights, the pipelines and the documentation. SaaS means your leverage at renewal equals your migration cost.

/ 05

Operational burden

A bought system is operated by the vendor; a built system needs an owner. If nobody on your side can own it, buy it — honestly.

/ 06

Regulatory posture

Where model cards, lineage and rollback authority must be demonstrable, custom systems are usually easier to evidence than a vendor black box.

FAQ

Common questions, candid answers.

Is custom AI better than off-the-shelf software?+

Neither is universally better. Off-the-shelf wins when your process is genuinely standard, the vendor's training data resembles yours, and speed matters more than the last ten points of accuracy. Custom AI wins when the decision is core to your margin, your data is unusual, or the accuracy ceiling of a generic model is below the threshold at which the system becomes useful.

What does custom AI cost compared with a SaaS licence?+

Custom AI is front-loaded: a data readiness assessment plus a first production deployment typically lands between a mid-five and low-six figure sum, then costs little to run. SaaS is back-loaded: low entry, then a per-seat or per-transaction fee that scales with your success. The crossover usually falls between year two and year four; below roughly 50 users or 100,000 transactions a month, SaaS is normally cheaper.

How long does each option take to deliver value?+

Off-the-shelf software can be live in days to weeks, with value limited to what the vendor already models. Custom AI takes two to four weeks for a data readiness assessment and eight to sixteen weeks to first production deployment, with a materially higher accuracy ceiling afterwards.

Who owns the data and the model?+

With custom AI built by an independent research centre, you own the code, the model weights, the pipelines and the documentation, and the data never leaves your chosen residency. With most SaaS, the vendor owns the model, often trains on aggregate customer data, and your leverage at renewal is limited by migration cost.

Can you combine both approaches?+

Usually yes, and often you should. Buy the commodity layers — OCR, transcription, general-purpose language models, storage — and build custom only the thin layer that encodes your specific advantage. That hybrid is the most common recommendation we make.

When is off-the-shelf clearly the right answer?+

When the process is regulated into a standard shape, when no proprietary data exists to learn from, when the team has no capacity to operate a model, or when the total decision volume is too low for a bespoke model to pay back. We tell clients this regularly and it costs us projects.

↳ Next step

Build, buy, or both?

Send us the decision and the volumes. We'll draw the five-year curve with your numbers, including the case where the answer is 'buy it'.

Get in touch

AI◆Big Data◆Vision◆Forecasting◆Knowledge◆Research◆Strategy◆Delivery◆