Written by: Michał Sobieraj, Operations Manager, Digital Colliers

Everyone wants to start with the AI layer. The clients who actually ship one usually spent a year on something that looks nothing like AI. A parts catalogue. A quoting tool. A customer portal that finally replaced the shared inbox. That boring web app is the reason the AI product later has anything to stand on.

Here's the pattern worth tracing, because it's the difference between the operators who end up in production and the ones who end up in the abandonment stats.

The parts platform that wasn't about AI

Imagine a mid-market industrial distributor. Thousands of SKUs, half the pricing logic in a rep's head, quotes handled over email, PDFs flying around. The first project isn't an assistant. It's a plain B2B platform: clean product data, structured customer accounts, quote workflows, an order history that's actually queryable.

Nothing about that phase is glamorous. It's schema design, data migration, permissions, edge cases in tax and freight. But by the end of it, three things exist that didn't before:

A single source of truth for what a part is and what it costs for whom.
An event log of every quote, order, and rejection, with reasons attached.
An access layer that knows which user is allowed to see which price.

Only now does an AI assistant on top of that platform make sense. A rep asks "what did we quote Acme for the 40mm bearing last spring, and what's the equivalent in stock today," and the assistant can actually answer, because the data model was built to answer it.

What happens when you skip step one

Skip the platform and try to bolt an assistant onto a folder of spreadsheets and an ERP export, and you join the graveyard. The numbers on this are grim and consistent. IDC found that 88% of AI proofs of concept never reach widescale deployment. RAND puts overall AI project failure north of 80%, roughly twice the rate of conventional IT work. MIT's 2025 study on enterprise GenAI found 95% of pilots deliver zero measurable P&L impact.

The reasons rarely have anything to do with the model. They're the things the boring web app would have forced you to solve:

No canonical entity for "customer" or "product," so the assistant hallucinates joins.
No permissioned data access, so you can't safely expose it to end users.
No event history, so the assistant has nothing interesting to reason over.
No feedback capture, so you can't tell if it's getting better or worse.

S&P Global's 2025 survey found 42% of companies abandoned most of their AI initiatives that year, up from 17% the year before. The average organisation scrapped 46% of pilots before production. The left-behind risk isn't hypothetical. It's most of the market.

What the boring phase actually buys you

Think of the internal platform as the substrate. A few concrete things it gives the AI layer later:

Grounded retrieval. The assistant queries a real database, not a scraped PDF.
Auditability. Every AI-generated answer traces back to a row and a user who was allowed to see it.
Guardrails you can enforce. Pricing rules, contract terms, and export restrictions live in code, not in a prompt.
A place to measure. You know what a good quote looks like, so you can tell when the assistant helps and when it doesn't.

This matters more as regulation lands. EU AI Act transparency obligations apply from 2 August 2026, and high-risk obligations from 2 December 2027, with fines reaching €15M or 3% of global turnover. If your AI feature sits on a platform with proper access control and logging, compliance is a config exercise. If it sits on a swamp of spreadsheets, it's a rebuild.

How to sequence it if you're starting now

If you're a custom platform operator sizing an AI roadmap for 2026, the order that keeps working looks like this.

Ship the unsexy platform first. Get the entities, the events, and the permissions right.
Instrument everything. Assume you'll want to fine-tune or evaluate later.
Pick one narrow AI use case where the answer lives inside your own data. Quoting assistance, spec matching, ticket triage.
Wrap it in the same auth and audit layer as the rest of the product. No side channels.
Measure against the pre-AI baseline the platform already gave you.

The teams who end up with an AI product in production almost always got there by building something that wasn't AI first. The ones chasing the assistant before the substrate are the ones filling out the failure surveys next year.

The parts platform that wasn't about AI

What happens when you skip step one

What the boring phase actually buys you

How to sequence it if you're starting now

Sources

Key-person risk is an AI use case

The bottleneck is the interface, not the data

The questions your stack cannot answer