“AI cannot reason about data it cannot see.”
The Integration Tax
Most enterprise AI fails not because the model is weak, but because the data is trapped behind a dozen portals, file shares and manual CSV exports. This is the pattern for connecting any of it to AI.
By Marc Heath · Usermode
The tax is paid in three currencies
The Integration Tax is the cost of leaving your data trapped. It is rarely a single invoice — it is paid slowly, in three currencies, until the project quietly stalls.
Time
Point integrations are slow. Six to nine months of bespoke wiring per system, and the project loses momentum long before the data is connected.
Quality
An agent reasoning over partial data gives confident partial answers. It doesn't know what it can't see — and neither does the person reading its reply.
Trust
When the early demos underwhelm, the sponsor stops answering emails about phase two. The hardest cost to recover is organisational belief.
How you connect any data to AI
Three ideas do the work. Each is a small, durable contract — together they let you wrap any system, contain any failure, and migrate at your own pace.
The vertical slice
Every integration owns its full stack — your system, an API adapter, and the MCP tools on top — as one self-contained package. A failure in one slice can’t cascade into the others. One slice per system, no shared blast radius.
The MCP boundary
Agents never see the messy vendor API. They see clean, named tools on the other side of a wall. That boundary is what lets us wrap any SaaS, file share, database or internal app — and because we build to the open MCP standard, the same tools work in clients like Claude Desktop.
The Strangler Fig
Migrate from SaaS dependency toward full data ownership at your own pace, in four phases. The MCP contract stays stable the whole way, so the agents never notice the system underneath them changing. No 18-month big-bang rewrite.
“The model is rented; it will be obsolete within the year. The moat is everything you build around it that the model alone could never know — your data, your gates, your memory, your tests.”
The whitepaper, in six parts
A practitioner's account of why integration is the real constraint on enterprise AI — and the pattern we use to remove it.
Part 1
The one sentence everything is downstream of
Why data — not the model — is the binding constraint on enterprise AI, and why renting a smarter model never fixes it.
Part 2
The three currencies of the Integration Tax
Time, quality and trust — how each is paid, and why point integrations keep charging the bill.
Part 3
The vertical slice
Every integration owns its full stack — data to API to MCP — as one self-contained package, so a failure in one can't cascade into the rest.
Part 4
The MCP boundary
Agents see tools, not vendor APIs. The contract that lets you wrap any SaaS, file share, database or internal app — and why the same tools work in clients like Claude Desktop.
Part 5
The Strangler Fig
Migrate from SaaS dependency to full data ownership at your own pace, in four phases, with the MCP contract held stable so the agents never notice.
Part 6
The moat
The model is rented and will be obsolete within the year. What you actually own — your data, your gates, your memory, your tests — is the part that compounds.
Stop paying the Integration Tax
The pattern in the whitepaper is what we install. Start with a £2,500 Audit, connect your first system, and see an agent reasoning over data it can finally see.