A proposal written for a European food and consumer goods manufacturer contained an effort plan: fourteen delivery activities, each with a percentage split between the outside team and the company's own people. Read across, it is an ordinary staffing table. Read down the client column, it is something else. Average the company's own share over the fourteen rows, each row counting once, and it comes to 44 percent. Before a contract was signed, before a platform was chosen, before anyone had touched a data source, the company was scheduled to carry nearly half of its first analytics project, row for row, with people already on the payroll.
The challenge
The company was standing up a BI and data analytics initiative, sponsored from the IT function covering marketing and sales. The ambition in the covering letter ran the length of the business: returns on marketing spend, predictability through the supply chain, quality in what arrived from suppliers. Capability was the stated object, not a single report.
The deck was unusually direct about what would break the work, and every condition it listed was internal. At the hypothesis stage: no senior sponsorship, and no case identified that would return five times the effort put into it. At the data stage: sources not available across silos, or protected by the stakeholders who owned them, and a company lacking the combination of data-driven business people and data scientists. At the closing stage: no discipline to run continuous benefit reporting, and no habit of improving the fact base once a first answer arrived.
The price page repeated the diagnosis commercially. A pilot of this kind ran 4 to 6 weeks in an established data analytics environment; this one was quoted at 6 to 8, dependent on the data available and the support the company itself put behind it. Two extra weeks, written into the fee, as the cost of starting cold.
That is a capability shortage, and it reads as a general one. The effort plan says something more specific, and more useful.
The approach
The pilot was designed in three blocks: one week to frame the question and the business case, four to six weeks to prepare the data and test what it says, one week to put benefit tracking in place. Fourteen activities were spread across those blocks, each carrying its own split.
Sorted by how much the company kept, the rows fall into a pattern that has nothing to do with difficulty.
It held half or more on seven. It held 90 percent of choosing which analytics platform to use and 90 percent of providing security and technical safeguards before any data was gathered. It split framing the business question and its hypotheses 50/50, interacting with technical engineers to source data 50/50, and analysing and validating the result 50/50. It held 80 percent of running the internal conversation with stakeholders about the business case and the benefits, and 60 percent of monitoring the benefits afterwards and steering on them.
It bought in four. Assessing the data model and the quality of source data: 20 percent in-house. Defining the approach across stakeholder engagement, data sourcing and enabling technology: 30 percent. Crunching data into a shape analytics could use: 10 percent. Integrating and transforming that data to test the hypotheses and refine them on the results: 10 percent.
Two rows sat at 40 percent, both business-case work. One row is printed in the source as 20 and 0 and does not sum to a whole activity. It stayed that way in the deck and it stays that way here.
Averaged by phase, the in-house share runs 52.5 percent in the hypothesis week, 32.9 percent across the analytical middle, and 60 percent in the benefits week. The company owned the start of its own project and it owned the end. Where it was thinnest was the four to six weeks in between, which is also the only part of the schedule with a range on it.
That is the finding the table was carrying without anyone reading it as one. The company had not lost the argument about judgement. It kept the decisions that carry consequence, kept half the hypothesis work, kept half of validation, and refused to let anyone else pick its platform or write its safeguards. What it lacked was throughput: hands to move data into a state where the judgement it already had could be applied to it.
Throughput is the one part of that list where the economics have completely changed.
The outcome
This was a proposal, so the honest outcome is a design and a price rather than a measured result. Three options were tabled: a single pilot at 1.3 FTE and 7,300 euros per week, two pilots plus a supplier-provided analytics environment with licensing at 2.1 FTE and 15,000 euros per week, and an inspiration session with a maturity scan offered separately, the session at no cost if either delivery option was taken. The example output page sketched a return on marketing investment trajectory of 10, 24 and 40 percent and labelled itself an example. No measured benefit at this company is claimed here, because none exists in the source.
What transfers is the shape of the table, and the shape has to be redrawn.
Take the four rows the company bought in. Assessing a data model, judging source quality, crunching data into analysable form, integrating and transforming it to test a hypothesis and refine it on the result. Every one of those is a loop: propose, run, read the error, revise, run again, stop when a written acceptance test passes. An agent with scoped warehouse credentials runs that loop hundreds of times in one analyst's afternoon, and it does not get bored on the fourth schema mismatch. The compression lands on the four-to-six-week middle, precisely the phase whose in-house share was lowest. Loop engineering closes the gap the effort plan had already located.
The harness is the other half, and it is built out of the two rows the company insisted on holding at 90 percent. Choosing the platform back then meant picking a stack. Now it means model choice, agent identity, which tables an autonomous process may read, what it may write, on what schedule, what is retained and where it is allowed to leave. Security safeguards before gathering data become security safeguards before an unattended process gathers on a timer that nobody is watching at 3am. Those rows get heavier, not lighter, and they stay in-house for the same reason they did then: the party living with the consequence should own the boundary.
The covering letter's ambition about the quality of what arrives from suppliers cashes out differently now too. At the time that was a sampling and reporting problem, answered after the fact from whatever was recorded. A vision model on an intake line reads every delivery rather than a sample, produces a labelled record per unit instead of a batch average, and turns a question that used to be answered quarterly into a stream an agent can act on. The judgement about what counts as a reject stays where it was, with the people who own the specification.
The last row moves most. Monitoring the benefits afterwards was split 40/60 and sat next to a written admission that the company lacked the discipline to run continuous benefit reporting. Discipline was the wrong thing to ask for. In an interconnected pipeline where ingestion, feature construction, model runs and the benefit report are scheduled jobs with agents as the workers, continuous reporting is not a habit anybody maintains. It is a scheduled job with an owner and an alert when the number moves.
Three things do not move at all. The five-times-return bar on case selection is still a judgement about where margin sits. Senior sponsorship is still a person, not a system. And running the internal conversation about the business case, 80 percent client in the original plan, is still 80 percent client, because no agent persuades the owner of a category to change how that category is funded.
The instructive part is what the redrawn table does to the in-house column. Loop and harness engineering does not shrink the share a company holds itself. It grows it, because the rows that were bought in for lack of hands come back inside, and the rows that stayed inside get more consequential. A manufacturer that ran this exercise today would find the same thing the original column showed and nobody totalled: the analytics team was already on the payroll. It was short of throughput, not judgement, and throughput is now the cheap half.
The risk the deck named at the data stage was a company short of the combination of data-driven business people and data scientists. That is still the right risk to name, and it is now a training problem rather than a hiring one, because the people who hold the judgement are already there and what they need is the loop-and-harness skill to direct the throughput.