Why AI Should Be Implemented End to End, Not One System at a Time

Adding AI to one system at a time makes individual steps faster, but the real value only appears when it is embedded across every system a process runs through, from the first customer request to the final invoice.

Picture a customer order as it moves through a company: it starts in the CRM, becomes a sales order in the ERP, moves on to the warehouse system and then to transport, and finally ends up as an invoice in finance. Now picture a company that adds AI to each of those systems separately, so that the CRM scores leads, the ERP forecasts demand and finance matches payments automatically, with every project reporting a success. The order still waits between every system, people still rekey data from one screen into the next, and the customer still receives the same delivery date as before.

One system at a time optimises the parts

Each AI project is scoped to the system it lives in, which means it only ever sees its own step in the process. It cannot see that the warehouse is waiting for a credit check, or that the invoice is wrong because the order changed two systems earlier, because the delays and errors that matter most sit in the handoffs between systems and no single system owns them. Optimising every part on its own does not optimise the whole, and in practice it often just moves the queue to the next system along.

The value is in the whole process

In research published in September 2026, McKinsey surveyed more than 700 executives and placed their organisations in three horizons: in the first, employees get access to general purpose AI tools; in the second, AI automates and improves existing cross functional workflows at scale; and in the third, the company redesigns roles, workflows and the operating model around AI. Only 13% of leaders in the first horizon report meaningful enterprise value from AI, rising to 24% in the second and 48% in the third. The Hackett Group reports the same pattern in its 2026 benchmarks, which show that organisations can achieve a 75% performance advantage over their industry peer groups through process led AI transformation.

Why piecemeal projects stall

The MIT NANDA initiative found in 2025 that about 95% of generative AI pilots stall and deliver little or no measurable impact on the P&L, and its authors did not blame the models. They blamed flawed integration, because tools that are not connected to the surrounding workflow cannot learn from it or adapt to it. A pilot inside one system can prove that the AI works, but it cannot prove that the business works better, because the business runs across all of its systems at once.

What end to end looks like

Embedding AI end to end means following one process through every system it touches and connecting them, so that the systems share the same data and a customer or an order means the same thing in the CRM, the ERP and finance. A coordinating layer passes work from one system to the next without anyone copying it across, and one person owns the outcome of the whole process rather than one system inside it. Governance has to run across the whole chain as well, because the EU AI Act requires high risk systems to record their events automatically and to be designed for effective human oversight, which is only possible when you can trace a decision from the first system to the last.

Where to start

Do not start with a system; start with a process that carries a P&L line, such as order to cash or procure to pay, and map every system it runs through and every handoff in between. Then embed AI across all of those systems together, and measure the result on the process as a whole rather than on any single step. It is a bigger first project than adding AI to one system, but it is also the only kind that changes the numbers.

All articles on this site are written by me. I use AI to assist with final formatting and editing before publication.