A proposal written a decade ago for a European food and consumer goods manufacturer carried one page that reads differently from everything around it. Not the price page and not the benefits page. It was the example effort plan: fourteen delivery activities, each with a percentage split between the delivery team and the client's own people. Formulate the hypotheses: 50/50. Crunch the data so it can be analysed: 90 percent us, 10 percent them. Agree which analytics platform to use: 10 percent us, 90 percent them. The rest of the document sold a possibility. That page divided the labour.
The challenge
The company was standing up a BI and data analytics initiative, and the sponsor for it sat in IT services covering marketing and sales. Building the capability was the point, which meant the pilot could not assume an environment that was already running. The pricing footnote said as much in one line: a pilot of this kind runs 4 to 6 weeks in an established data analytics environment, with the range dependent on the data available and the support the client itself puts behind it. This one was quoted at 6 to 8. Two extra weeks, written into the price, as the cost of starting cold.
The more interesting honesty sat on the method page. Alongside a five-stage delivery model, the deck listed the conditions that break each stage, written flatly as things the company lacks. At the hypothesis stage: lack of senior sponsorship, and no genuine case identified, where "genuine" was defined as returning five times the effort put in. 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 the first answer arrived.
Most of those six failure modes are organisational rather than technical. Sponsorship, staffing, ownership of data, reporting discipline. None of them is solved by a model. A pilot scoped only as a technical exercise would have run straight into all of them, and the argument afterwards would have been about who was supposed to have done what.
The approach
The pilot was designed as five stages: hypotheses and business case, data preparation, analysis, validation, then benefits realisation. Stages one through four were grouped as value generation; the last stood on its own. Time was allocated one week for hypotheses, four to six weeks for the analytical middle, one week for benefits. The pilot itself sat inside a larger method of six named components across three phases, with the deliberate position that exploration comes first.
Then the effort table did the real work. Read as a whole, the fourteen rows encode a philosophy about where a partner belongs.
The heavy technical rows went to the delivery team. Understanding the data model and assessing source quality: 80/20. Crunching data into analysable shape: 90/10. Integrating and transforming it to test the hypotheses and refine them on the results: 90/10. Defining the approach across stakeholder engagement, data sourcing and enabling technology: 70/30.
The thinking rows were split down the middle. Formulating the business question and its hypotheses: 50/50. Working with technical engineers to source the data: 50/50. Analysing and validating: 50/50.
The rows the delivery team refused to own went the other way. Choosing the analytics platform: 90 percent client. Providing security and technical safeguards before any data was gathered: 90 percent client. Running the internal conversation with stakeholders about the business case and the benefits: 80 percent client. Monitoring benefits realisation afterwards and steering on it: 60 percent client.
Provide security and technical safeguards before gathering the data. 10 percent us, 90 percent you.
The commercial page followed the same logic. A single pilot at 1.3 FTE and 7,300 euros per week. Two pilots plus a sandbox platform with licensing at 2.1 FTE and 15,000 euros per week. An inspiration session and maturity scan offered separately, the session at no cost if either delivery option was taken. Per-week pricing against a stated FTE count, not a fixed fee, which is only defensible when the activity split is written down.
One row in the source table lists 20 percent and 0 percent and does not sum to 100. It stayed that way in the deck.
- 14
- Activities with a proposed split
- 90/10
- Data preparation, consultant-led
- 6-8
- Weeks proposed
The outcome
This was a proposal, so the honest outcome is a design rather than a result. The example output page sketched return on marketing investment rising through 10, 24 and 40 percent, and labelled itself an example. No measured benefit at this company belongs in this write-up, and none is claimed here. What the document put on the table was a definition of who does what, written down before any money moved.
That is still the right instinct, and a decade on, the table needs rewriting rather than retiring.
Start with the rows that agentic systems absorb. Data crunching at 90/10, integration and transformation at 90/10, data model comprehension at 80/20: these are the three heaviest partner rows in the original table, and every one of them was priced as hands. They are now a loop and a harness. The loop is an agent that reads the source schema, writes the transformation, runs it, checks the result against a test that states what a correct number looks like, and revises itself when the test fails. The harness is everything that decides what that agent may touch: which systems, on whose identity, with which fields masked, and what is written to a log a stakeholder can read afterwards. Build the harness once across the manufacturer's sales and marketing sources and the analytical middle stops being bounded by how many analyst-weeks you bought and starts being bounded by how fast the loop closes. The one-week hypothesis phase and the one-week benefits phase barely move, because neither was ever labour-bound.
The 50/50 rows survive as the scarce work, and their share of the engagement goes up. Deciding which question is worth five times the effort, and deciding when an answer is true enough to act on, are judgements that need someone who owns the margin. When generating a plausible analysis becomes cheap, validating one becomes the expensive step, not the formality.
There is also a row the original table could not have written. All fourteen activities assume the fact base is already tabular: sources located, integrated, transformed, analysed. The deck's own data-stage failure mode says the sources sit across silos, or behind the stakeholders who own them. In a marketing and sales estate, a good part of what sits behind those stakeholders was never in a table to begin with. Trade terms signed and filed as scans. Promotion claim forms. Pack artwork through its revisions, and photographs of what the shelf looked like the week a promotion ran. Computer vision turns that material into fields an agent can query, which changes what the data preparation row is for. The original row could only reach what someone had already put in a table, which is why the 90/10 split on it was a statement about how many hands were free. Its replacement today reaches the evidence that never got there.
The rows the delivery team pushed to the client get heavier, not lighter. Platform selection then meant picking a data stack. Now it means model choice, agent identity, tool permissions, retention, and the question of where data is allowed to leave. Security safeguards before data is gathered becomes security safeguards before an autonomous process gathers on a schedule nobody watches. Agentic and autonomous are two different purchases, and the difference belongs on the table. An agent working inside a loop with a person reading its output is a throughput decision. A process that sources, transforms and reports on its own schedule is an operating decision about how the company runs. Both stay 90 percent client-side, and both now carry consequences a warehouse decision never did.
One row does not move at all. Running the internal conversation about the business case, 80 percent client, is still 80 percent client. No agent persuades the owner of a category to change how that category is funded.
The row that improves most is the last one. Continuous benefit reporting was named in the proposal as a discipline the company lacked, and a missing discipline is the usual reason pilots stop being measured a quarter after they end. Nobody has to locate that discipline now. Sourcing, transformation, the reruns of the analysis and the benefit report itself can sit in one connected pipeline where agents do the steps between them, with exceptions routed to a person instead of the entire report. The 40/60 split stops describing an intention and starts describing a job that either ran last night or did not.
For a manufacturer commissioning agentic work today, the practical move is the same as it was then, with one change. Write the activity table before the statement of work, and give it three columns rather than two: the client, the partner, and the system. Every row that moves into the third column still needs a named human owner for the day it is wrong. A pilot that cannot say who owns each activity has not been scoped, whatever the technology in it.
Redraw the table that way and the price page reads differently too. Two of the six to eight weeks in this proposal were bought to cover starting cold, with no analytics environment already running. That premium was real when the proposal was written and it is avoidable now, because the environment a pilot needs is mostly harness, and a harness is built once and reused. The part of this plan that ran longest is the part that compresses hardest.