What you should take away
- Research workloads are long-context and document-heavy: consumption per output is orders of magnitude above service chat.
- Licensed data terms, not compute, frequently decide feasibility.
- The binding cost is analyst verification minutes per published output.
- The right denominator is a decision-grade output, not a generated draft.
Investment research is where AI economics inverts. In client service, volume is enormous and each interaction is small. In research, volume is modest and each output is large: a filing pack, five years of transcripts, a sell-side model, a data vendor extract. A single pre-earnings note can consume more context than several hundred service conversations.
And yet inference is still not the dominant line. The dominant line is the analyst who must be able to defend every number in the note.
The four real drivers
- Context volume per output: filings, transcripts, prior notes and data extracts assembled for each question, across multiple reasoning passes.
- Licensed content rights: whether your market data, broker research and news licences permit ingestion into a model workflow at all, a legal constraint that precedes any cost model.
- Analyst verification time: the minutes spent checking citations, figures and the chain of reasoning before anything reaches a portfolio manager or a client.
- Wrong-answer cost: a plausible but false figure in an investment note is not a service inconvenience; it is a decision input and, if published, a conduct matter.
Modelling a research workload
The structure is the same as a service workload, with different magnitudes: fewer 'interactions', far more context per interaction, and escalation replaced by mandatory human review. Below, escalation time stands for analyst verification minutes per output, and volume is research requests rather than client conversations.
Interactive model
Cost per decision-grade research output
Illustrative desk: 8,000 research requests a month, 12 minutes of analyst time per request today at £95/hour fully loaded, agent drafting with mandatory verification.
Total cost per successful outcome
£50.69
Model inference per successful outcome
£5.64
Human escalation cost
£1.22m
Modelled annual operating difference
£3.14m
Where the cost sits
- Human escalationCalculated£1.22m52%Calculated
- Implementation, annualisedCalculated£350k15%Calculated
- Model inferenceCalculated£260k11%Calculated
- Retrieval & dataCalculated£220k9%Calculated
- Platform, evaluation & monitoringCalculated£140k6%Calculated
- Tools & APIsCalculated£90k4%Calculated
- Failures & retriesCalculated£60k3%Calculated
Largest modelled component: Human escalation (£1.22m). Outcome measured per research output produced without deep rework.
Illustrative assumptions for a fictional desk, not a benchmark. Note how sensitive the result is to verification minutes compared with the model bill.
Where research agents pay reliably
- Extraction and normalisation from filings and transcripts, where the output is checkable against a source line.
- Coverage breadth: monitoring names or instruments no analyst currently has time for.
- First-draft structure for recurring formats, where the analyst edits rather than composes.
- Search across an internal note archive, which is retrieval rather than generation and carries the lowest risk profile.
Where they pay least reliably is in conclusions. The judgement is the product, and the verification cost of a machine-made judgement usually exceeds the cost of a human making it.
Apply this to your own workload
The figures above are illustrative assumptions. A BillingEngine assessment replaces them with your company's numbers and shows which assumptions decide the answer.