A European composite insurance group received a priced plan for finding out where analytics could pay inside its own business. Three phases, three months. Costed at the rates on the schedule of charges it came to roughly 36 person-weeks of senior consulting time and EUR 252,000 in professional fees, before travel. What all of that effort would leave behind was a short list of structured fields: sixteen value drivers, eight candidate cases, five maturity dimensions and one ranked roadmap per country business. Thirty-five fields. The reading cost more than the answer, which was the normal condition of enterprise analysis at the time, and is the part that has changed most in the years since.
The challenge
The group was standing up a group-level analytics centre of expertise, and the centre had no consolidated view of what its own country businesses actually did. Those businesses ran their own markets under their own leadership, life and non-life alike. The group could publish a target, and the proposal was explicit that analytics value had to be traced to the group's own published financial targets rather than to generic efficiency claims. Which claims operation had a fraud problem, which had a cycle-time problem, which had already tried something: establishing that, country by country, was the job the plan was written to do rather than a finding it started from.
So the plan bought sight. The schedule of charges priced it phase by phase. Preparation: EUR 67,500, about 2.1 full-time equivalents over five weeks, run from the head office. Exploration: EUR 24,500 for each country visit, about 3.2 full-time equivalents for one week per country, with six countries in the delivery plan. Prioritization: EUR 37,500, about 2.1 full-time equivalents over three weeks, quoted on the understanding that four to six country businesses would be in scope. Materials, travel and subsistence were extra and charged as incurred.
Multiply the indicative team by the duration and read the result as a stack: 10.5 person-weeks of preparation, 19.2 of exploration, 6.3 of prioritization, which rounds to 11, 19 and 6. The exploration phase is both the widest line on the invoice and the bulkiest block of time, because it is the part where humans physically travelled to where the knowledge was and asked for it out loud.
The approach
Four deliverables were defined, one per phase except the country visits, which carried two.
Preparation would produce a value map for the group. The proposed method was two workshops to tailor a generic insurance value map to the products and segments the insurer actually sold, then five to ten validation interviews with a friendly country business spanning product development, marketing and sales, underwriting, service, policy management and claims processing, plus desk research on what the newer entrants in the market were doing.
The map itself is the artefact worth reading closely, because it is where claims enters. It splits the insurance value chain into four blocks, one of which is claims processing, and joins that chain to a value tree of sixteen drivers: six on revenue, four on cost, six on risk. Claim cost is one of the four cost drivers. Direction of effect is written on each driver, including the ambiguous ones, so pricing and personnel carry a plus and a minus rather than an assumed upside. The map is labelled indicative, and it excludes asset management. It is a proposed structure, not a measured finding, and the deck says so.
Eight candidate cases were plotted across that chain, and two of them land on the claims block. One would predict claims behaviour to inform pricing, underwriting, product design and the handling approach, on the argument that behaviour and attitude patterns correlate with both fraud and propensity to renew, so detecting one improves the other and the whole process gets faster. The other is a value chain optimiser, and the worked example the deck attaches to it is claims management: mine the usage, click and call data of a service operation end to end, find where work bounces between departments, and take steps out of the process. The remaining six sit elsewhere on the chain, in social trend spotting, churn and influence scoring, risk modelling for high-value commercial assets, connected home, next best action and price sensitivity.
Exploration would run three-day visits to each country business. A short leadership session, three to four hours, to surface local candidate cases in the local market context. Alongside it a maturity scan, sent out a week ahead of the visit, scoring five dimensions: data management, people, technology, organisation and users. Prioritization would aggregate the country outputs into one ranked log with values and efforts harmonised across countries, and a roadmap, submitted as an investment plan to the board.
Every one of those steps is a reading job. Read the market. Read what the local leadership already knows. Read the systems estate. Read six sets of answers and reconcile the units they came in. The proposal priced reading by the person-week because at the time there was no other way to price it.
The outcome
This was a competitive proposal, so the honest outcome is a design and a price. No delivered benefit at this insurer is claimed here, because the source contains none. What the document does contain is a fully costed picture of what enterprise intake used to require, and a claims agenda written under an assumption that has since expired.
The assumption is in the wording of the claims case. Predict claims behaviour, detect fraud, speed up the process. All of it starts from claim records: fields already keyed, amounts already entered, a case already open in a system. The intake problem inside claims was invisible because it was somebody's job description. A first notification of loss arrives as a phone call, a form, photographs of a damaged vehicle or a flooded floor, a repair estimate on a garage letterhead, a police report, a medical letter, sometimes handwriting. A person read all of that and typed the fields that the analytics in this plan would later model.
Vision models read those at arrival. Damage photographs classify and localise, an estimate parses into line items, a form scans into fields, and the mismatch between the vehicle on the photograph and the vehicle on the policy surfaces at intake rather than three weeks into the file. The fields exist before anyone types them, which means the fraud case and the cycle-time case stop being downstream models fed by a data project and become properties of the front door.
The same shift applies to the engagement itself. Preparation was desk research and framework tailoring, and that is a loop: propose a draft map, test it against the insurer's own product and claims taxonomy, read what fails, revise, stop when a written check passes. Exploration was six repetitions of one procedure with a report at the end, which is what an agent run against six estates does without the flights. Prioritization was aggregation and unit harmonisation, which is code with a review step, not three weeks of senior time. What stays human is the leadership session, because its output was agreement rather than analysis.
Wire those together and they stop being three separate improvements. An agent reads the claims estate and proposes candidate cases against the value map. A second agent prices each one against the group's published financial targets. A third scores the maturity dimensions from what the systems actually do rather than from what a survey respondent believes. The roadmap regenerates when the underlying evidence moves, instead of ageing from the day the board approved it. That is an interconnected pipeline with agents carrying real work in it, each scoped to a task, each logged, each stopping at the gate where a person decides. Autonomy is graded along the same line the proposal already drew: reading and drafting run unattended, and the investment decision does not.
The remaining cost is not analysis, it is engineering. A loop that plans, runs against the real claims and policy systems, and tests its own output against a written acceptance check. A harness that holds the fixtures, the evaluations, the permission boundary and the audit trail so the loop can run unattended overnight. Build it that way and the reading work this proposal priced at 36 person-weeks becomes machine time plus the workshop the leadership were always going to want; we have not run that against this insurer's estate, and the claim here is about where the effort sits, not a measured speed-up. The one thing that does not compress is the decision at the end, and that was never the expensive part.