Who Actually Buys the First Analytics Pilot

I keep the proposals I have put my name to. They are more useful now than they were at the time, because a proposal records what somebody believed before any of the work happened, without the tidying a closing report does.

The one in front of me is a proposal for analytics support at a European food and consumer goods manufacturer. I co-wrote it and I signed it. Judged as method it still reads well: pick one to three areas, deliver a first proof case with a concrete result in a short window, and only then work out what platform and what skills to build. I would argue the same sequence today.

Read instead as a sales document, which is what it was, it says something the method never mentions. A sales document is written to persuade one person, and once you know who that person is, you can see them in the shape of every page.

One addressee, and what that role owned

The cover page names a single recipient. Not a steering group, but one role, which held IT services for both marketing and sales across the group.

That is where the first analytics money came from. Not a data office, which at that point in most manufacturers did not exist as a line on the chart, and not the supply chain director who would end up living with several of the answers. It came out of a shared IT services line, and much of what follows was decided by that.

The opening letter names four places where analytics would pay: returns on marketing, predictability in the supply chain, the quality of what suppliers deliver in, and the services provided back out to those suppliers. Read that against the addressee's remit and the arithmetic is one in four. The systems behind marketing sit under that role. Supply chain predictability does not, and neither supplier-facing item does.

Then come the proof points, prior work with other clients and stated as such: budget allocation across online marketing channels that improved conversion ratios by 20 percent, and a demand and supply programme at a manufacturing client that produced a 30 percent lead-time reduction, a 25 percent inventory improvement and $150M of cash freed up. None of those is a result at this manufacturer, and the letter says so.

But notice where they point. One at marketing spend, which the buyer owns. Three at a supply chain the buyer does not own. A letter that arms its reader with somebody else's numbers is a letter that expects its reader to go and repeat them in a room it will never see.

Fourteen candidates, three bets, one remit

The opportunity map is the most quotable page. Fourteen named candidates in four groups: four under marketing, three under sales, four under operations, three under a heading of Other. Every one is named as a product rather than a project, which matters, because a business owner can ask for a product by name.

Seven of the fourteen fall inside what the addressee's role covered; seven fall outside, in operations, finance and asset work.

Three are taken forward into worked detail, two labelled sales cases and the third a marketing case. All three land on the inside of that seven-seven split.

Nobody in the room decided that cynically. Look at how the three pages are written and the reason surfaces on its own. Two of them set out a method pointed at the buyer's own systems. One reconstructs the effort and efficiency of channels and of the sales force by analysing and visualising system log files from every channel. The other joins online advertising channels, banner and social and search, to the back-end transaction systems, so that a per-product funnel and a fact-based cost per sale fall out.

Both of those run on data the buyer's own function administered. No negotiation with a plant, no request to a finance controller, no conversation with whoever holds the supplier records. The buyer could hand over the inputs on their own authority in week one.

The third page is the exception that makes the point. It is the demand case, filed under sales but answerable to the supply chain, and it describes what was built somewhere else, an information flow across a global supply chain with a data lake and a prediction layer over it, rather than which of this buyer's systems it would draw on. The one candidate whose inputs the sponsor could not release on their own signature is also the one the document never gets specific about.

That is the mechanism. A pilot gets scoped against the data its sponsor can reach without asking permission, and that reach is set by the org chart, not by the margin.

What a pilot bounded this way can conclude

A case built on advertising channels joined to transaction records can conclude a great deal about how spend should be reallocated. It cannot say whether the product should have been made in that quantity, in that plant, that week. That question needs data held by people who were not in the room when the work was bought.

So the finding would come back true, useful and small, landing inside the sponsor's own function and confirming what it already suspected. Nothing structural would move. That is the ordinary end of a first-wave analytics programme, and weak modelling has nothing to do with it.

The document knows it. Its method section lists what breaks this kind of work before it starts: lack of senior sponsorship, no candidate clearing a bar of five times return on the effort spent, data not available across silos or actively protected by the stakeholders holding it, a missing combination of business people and data scientists, no discipline to keep reporting the benefit after the closing meeting. Not one item is technical. Two are the same problem from opposite ends: the sponsor cannot reach the data, and the people who can are not the sponsor.

14

Named candidates across four functions, seven inside the buyer's remit

3

Carried forward as worked cases, all inside it

80%

Client share of facilitating the stakeholder dialogue on the business case

6-8 weeks

Quoted duration, against 4-6 for an established data estate

The price is set for one signature

The commercial page makes the sponsorship shape explicit, because prices always do.

The entry option is a single proof case, quoted on 1.3 full-time equivalents at 7,300 euros a week across roughly six to eight weeks. Multiply it out and the engagement lands between 43,800 and 58,400 euros, which is arithmetic on the stated weekly rate rather than a total the document prints. The second option, two cases plus a supplier-provided analytics environment and its licensing, is quoted on 2.1 full-time equivalents at 15,000 euros a week.

The first number is priced to be signable by exactly the role it is addressed to, without a board paper and without the governance conversation a larger figure would trigger. The scoping session at the front is free provided the buyer proceeds with one of the paid options, which turns a workshop into a commitment.

A footnote carries the most honest line in the document: a case of this kind typically runs four to six weeks where the data environment is established, and six to eight was quoted here. The extra time is the price of an estate that is not ready, and it sits in small type rather than in the summary, which is where I would put it now.

The eighty-twenty that gives it away

Then there is the effort table, which assigns every delivery objective a numeric split between supplier and client before anything is signed.

Two rows are correct and unremarkable. Choosing which analytics platform to use is 90 percent client. Providing the security and technical safeguards before data is gathered is 90 percent client. The party that lives with the platform and carries the exposure should choose it and secure it.

The row I would now read out loud in the first meeting is the next one. Facilitating the dialogue with stakeholders about the business case and the benefits is 20 percent supplier, 80 percent client. Monitoring the benefit afterwards is 40 percent supplier, 60 percent client.

We were telling the buyer, in a table, that selling the result to the rest of their own organisation was mostly their job. That is honest, because external credibility does not transfer across a function boundary. It also means the pilot's fate rested on one person carrying a finding into rooms where they had no authority, and nothing else in the document treats that as the hardest task in the engagement.

A proposal is answerable to whoever can sign it. The shortlist follows the signature, the reachable data follows the shortlist, and the conclusion follows the data. By the time anybody argues about the model, the question has already been narrowed.

What I would change

Not the method. Small first, platform later, one falsifiable question rather than a data lake with a hopeful roadmap attached. That has aged fine, and the current run of language model pilots crossing my desk would be better for it.

The addressing is what I would change. Write down two people, not one: the person who can sign, and the person whose margin the answer is supposed to move. Where they are the same, the pilot has a chance of changing something. Where they are not, say so on the first page and design the handover, because the handover is the risk and it currently sits in a table as an 80 percent line item with no owner and no method.

Then check the shortlist against the second person. If every surviving candidate is one whose data the sponsor happens to control, the org chart made the selection. Pick at least one case that needs somebody else to release something, and treat getting that release as part of the deliverable rather than a precondition of it. That is how RealAI's Platform work sequences a first engagement now: the release negotiation is scheduled work with a named owner and a date, not an assumption sitting in the first paragraph.

Drawn from a proposal for analytics support at a European food and consumer goods manufacturer, which I co-wrote and signed: its cover page, opportunity map, worked cases, effort splits and priced options. The percentages in it were achieved with other, unnamed clients, and the durations and prices are an offer rather than a record of delivery. Reading its addressing as the thing that bounded the work is mine.