Most companies already buy all three. They just buy them from three different places.

A consultancy writes the strategy. A systems integrator builds something. A tool, somewhere further down the chain, produces the output. Three engagements, three contracts, three sets of assumptions — and between each pair, a handoff where the person who understood the problem passes a document to someone who did not.

The deck says "automate supplier onboarding." Nine months later, something exists that technically does that, and nobody can point to a number that moved.

This is not an anecdote. It is the shape of the market. IDC's Q1 2026 survey of more than 620 IT leaders found that 81% of organisations have a detailed AI strategy, while 12 to 16% have reached meaningful execution. Read that twice if you are currently commissioning a strategy. Strategy is not the scarce good. It is the most abundant deliverable available.

What is scarce is one pair of hands from the first question to the file that lands on a customer's desk. Kept together, the three words change meaning. Strategy stops being a vision problem and becomes a ranking problem: which decision is made most often, on the worst inputs, by the most expensive people. Implementation stops being a platform and becomes plumbing: turning the files you already receive into something a machine can query. Generation is the only part the business ever sees — the quote, the document, the price file that nobody retyped.

Which means the honest test of the first two is never a milestone review. It is whether a specific person stops doing a specific task.

We put the whole model into sixty seconds.

▶ Watch the film — AiGAIN: Strategy · Implementation · Generation (1 min)
https://youtu.be/oF_zmEh_0PQ

It is the first question in almost every initial meeting, and it is always asked in good faith. The assumption behind it is reasonable: the data lives in the ERP, so connecting an AI to the ERP should produce the result.

It rarely does, and the reason is worth understanding before you budget anything.

The ERP holds what has already been decided. The purchase order exists because someone read a quote, checked a certificate, compared two lead times and made a call. The work that consumes the team happens before the ERP — in the inbox, in the shared drive, in the attachment nobody has opened yet. That is where the cost sits, and none of it is a database problem.

Look at what actually arrives in a buying office in a week: supplier price updates as Excel files with merged cells and three tabs; inspection reports as scanned PDFs; certificates attached to an email thread; a quotation revision buried in a reply chain; an invoice with a currency that changed since the last order. All of it is structured — a human can read it in ten seconds — but none of it is structured as fields.

So the honest answer to the question is: the ERP connection is the easy part, and it comes last. The work is turning what you already receive into something a machine can query. Almost none of it is model work. It is unglamorous, it is fast when scoped narrowly, and it is the step most vendors skip because it does not demo well.

The practical version: before asking which AI to connect to your ERP, count what lands in your team's inbox in a single week that someone has to read and retype. That list is the project. The ERP is where the result gets written, not where it gets produced.

The situation. A buying office receives updated price files from its suppliers every season. They arrive in whatever format each supplier happens to use: an Excel sheet with merged cells and three tabs, a PDF export, an email body saying "+4% on the 200 series, other references unchanged." Someone consolidates all of it into the master file by hand. Every quotation issued between the supplier's update and the end of that consolidation carries last season's prices.

Strategy — the ranking. The question is not where AI could create value; every function answers yes. It is which decision is made most often, on the worst inputs, by the most expensive people. In a buying office, that is pricing. It repeats per supplier, per reference, per season, and the people doing the retyping are the ones who should be negotiating.

Implementation — the plumbing. Almost none of this is model work. It is turning the files that already arrive into fields: reference, currency, price break, MOQ, incoterm, validity date. The merged cells, the PDF, the sentence in the email body. Deliberately unglamorous, and the reason the rest works.

Generation — the only part the business sees. An updated master price file, every changed line flagged against the previous version, every value traceable to the document it came from. Produced in a click instead of a week. Nobody retyped it.

The second turn. This is where the economics change. Once a price history exists as data rather than as a folder of attachments, a question becomes askable that was unaskable a month earlier — not forbidden, simply unanswerable. Which suppliers have raised prices twice this year while lead times got worse? Which references have drifted furthest from the quoted cost base? Those questions produce supplier decisions that no workshop would have surfaced, because the data to justify them did not exist in usable form.

What it does not do. It does not negotiate, it does not set sell prices, and it does not write to ERP master data without human validation. Ask any AI supplier what their system will not do. A specific answer means they have built it before.

Scope. Five days to establish whether it can be built on real supplier files and to put a number on it. $1,500. Deployed on the client's own accounts — the supplier data never leaves their infrastructure.

See the three sourcing automations → https://aigain.ai/sourcing-automation/

If you use AI in a working session — a pricing review, a product decision, a supplier strategy — the order matters more than the prompt.

Have every person in the room write their own ideas down before a single model output is shown. Ten silent minutes, no discussion.

The reason is measurable. Doshi and Hauser, in Science Advances, found that access to AI-generated ideas made written work more creative and more enjoyable — especially for the less creative participants — while making the AI-assisted outputs markedly more similar to one another than the human-only ones. Individually better, collectively narrower.

In a meeting, that convergence is invisible and permanent. Once the model has framed the problem, nobody un-frames it. The three options on screen become the option set, and the fourth one — the one that only the person who has handled that supplier for six years would have raised — never gets said.

Ten minutes of silence is the cheapest safeguard in the entire protocol.

The full five-step version is here: The Fifth Participant → https://aigain.ai/the-fifth-participant/

Eighty percent of people say AI made them more productive. Thirty-seven percent of their companies can find it in the EBIT. The gap is not a technology problem — it is that AI was sent to the back office and never invited into the room where decisions get made… →

The film names the model in sixty seconds. The article says where it breaks: why each stage is individually defensible and the chain is not, what the three words have to mean when nobody hands them off, and why we pointed the model at sourcing rather than at everything... →

Nine in ten procurement leaders say they are ready for AI. One in ten is getting measurable results. The gap is not the models. It is what sits underneath them… →

For decades, supply chain management ran on a logic of compromise... →