The 60/40 split

Most AI use cases in a business fall into one of two buckets. The first bucket — about 60% — is standard stuff. Email classification. Basic document OCR. Simple chatbots that answer FAQs from a knowledge base. Off-the-shelf tools handle these well enough. You sign up, configure some settings, maybe do a bit of prompt tuning, and you're running.

The second bucket is where things get interesting.

That remaining 40% involves processes that are specific to how your company operates. Your claims adjudication rules. Your underwriting exceptions. The way your logistics team handles partial shipments when a carrier misses a pickup window and two customers need priority rerouting on the same lane. No SaaS product was built for that. No SaaS product will ever be built for that, because the market for "your exact operational workflow" is a market of one.

The mistake companies make is trying to force that 40% into the 60% bucket. They buy a general-purpose tool and then spend 6 months trying to configure it into something it wasn't designed to be.

Where generic tools break

There are four situations where off-the-shelf AI consistently fails:

  • Your process has exceptions that matter. Generic tools are trained on common patterns. When 18% of your cases follow non-standard paths — and those cases represent 35% of your revenue — an 80% accuracy rate on the standard path isn't good enough.
  • You need to connect 4+ internal systems. Most SaaS AI tools offer integrations with popular platforms. But when the workflow requires pulling data from your policy admin system, cross-referencing it with your claims database, checking against a compliance rules engine, and writing back to your ERP — you're looking at custom integration work regardless. The "off-the-shelf" part becomes a thin wrapper around a custom build.
  • Accuracy below 95% creates real costs. If a chatbot gives a wrong answer, someone corrects it and moves on. If your document processing system misreads a coverage limit on a commercial insurance policy, that error can cost six figures when a claim hits.
  • Compliance requires audit trails. Regulated industries need to show exactly how a decision was made, what data was used, and when. Black-box SaaS tools rarely provide this level of traceability.

A real example

A P&C insurer we worked with tried using a general-purpose document AI platform for claims intake. The platform was well-regarded — good reviews, solid customer list, reasonable pricing at $4,200/month.

On standard personal auto claims, it performed at 82% accuracy. Acceptable for a first pass, with adjusters reviewing the output.

Then they tried it on commercial lines. Commercial property claims. Contractors' liability. Inland marine. These forms are longer, more variable, and use terminology that differs between carriers. The same platform hit 41% accuracy on commercial lines documents. Forty-one percent. That's worse than a coin flip on some field types.

The problem: commercial lines represented 70% of their premium volume. They'd evaluated the tool on the easy cases and assumed it would scale.

They ended up building a custom document processing system trained on their specific form types, with extraction rules mapped to their underwriting guidelines. Total build cost was around $175K. It runs at 94% accuracy on commercial lines and 97% on personal lines. The general-purpose tool couldn't touch those numbers because it was never trained on their data.

The cost comparison people get wrong

Here's the math that trips up most decision-makers. They look at the sticker price: SaaS tool at $50K/year vs. custom build at $200K. The SaaS tool looks like the obvious choice.

But they're not accounting for the real total cost:

  • Integration costs. Connecting the SaaS tool to your internal systems typically runs $30-80K in consulting and development time, even with "pre-built connectors."
  • Workaround labor. The 18% of cases the tool can't handle still need people. If you have 10 FTEs doing this work today, the SaaS tool might reduce that to 6. A custom system might reduce it to 2.
  • Opportunity cost of accuracy gaps. Every error the system makes costs something — rework time, customer complaints, compliance risk. At scale, the difference between 82% and 95% accuracy is hundreds of hours per year in corrections.
  • Vendor dependency. When the SaaS vendor changes their pricing, deprecates a feature, or gets acquired, you're exposed. With a custom build, you own the system.

When you add all of this up, the $50K SaaS option often costs $120-180K per year in total. The $200K custom build, with $30K/year in maintenance, starts looking different.

When to buy off-the-shelf

Custom isn't always the answer. Buy when:

  • The process is standard across your industry — email triage, basic scheduling, FAQ chatbots
  • Accuracy above 80% is good enough for your use case
  • You don't need deep integration with proprietary internal systems

These situations are real, and plenty of companies waste money building custom solutions for problems that a $200/month SaaS tool solves perfectly well.

When to build custom

Build when the process is your competitive differentiator. When you need 95%+ accuracy because errors have direct financial consequences. When the workflow touches 4+ internal systems that need to stay in sync. When you're in a regulated industry and need full auditability.

The pattern we see repeatedly: companies start with the generic tool, hit its limits within 6 months, and then build custom anyway — having spent both the SaaS fees and the time. If your use case clearly falls in the "build" category, starting there saves you a year.

If you're evaluating whether a custom AI solution makes sense for your operations, we can help you run the numbers. See our solutions or get in touch to talk through your specific use case.