What you should take away
- Small headcount means labour arbitrage is a weak argument; capacity and coverage are the real ones.
- The binding constraints are data licensing, information-barrier design and verification time.
- Cost per decision-grade output is the only unit that connects AI spend to investment process.
- Operational workloads, reconciliation, allocations, LP reporting, diligence extraction, often pay faster than idea generation.
A bank automating client service is buying cost-to-serve reduction across millions of interactions. A hedge fund has no such volume. Applying the same framing produces a rounding error, which is why AI cases at funds often collapse into an unfalsifiable claim about edge.
The better framing is capacity. If a research process can cover 40 names properly and AI-assisted workflows take that to 65 at constant headcount, the question becomes whether 25 additional names of real coverage justify the spend. That is a question an investment committee can actually answer.
The constraints that bite first
- Data licensing: market data, broker research and expert-network content frequently prohibit ingestion, retention or derivative use. This is a contractual gate, not a negotiation with your engineers.
- Information barriers and MNPI: a retrieval index that spans compliance-walled content is a control failure in waiting. Entitlement must be enforced at retrieval.
- Verification: an unverifiable model output cannot enter a position decision, so review time is part of the unit cost, permanently.
- Concentration: if several funds run the same models on the same public data, the resulting analysis is unlikely to be the source of differentiated edge.
Interactive model
Cost per decision-grade output at fund scale
Illustrative fund: 1,200 research and operational requests a month, currently 90 minutes of professional time each at £120/hour fully loaded.
Total cost per successful outcome
£183.54
Current annual human cost
£2.59m
AI-attempted workflow cost
£1.02m
Modelled annual operating difference
£797k
Where the cost sits
- Human escalationCalculated£318k31%Calculated
- Implementation, annualisedCalculated£250k25%Calculated
- Model inferenceCalculated£140k14%Calculated
- Retrieval & dataCalculated£120k12%Calculated
- Platform, evaluation & monitoringCalculated£90k9%Calculated
- Tools & APIsCalculated£60k6%Calculated
- Failures & retriesCalculated£40k4%Calculated
Largest modelled component: Human escalation (£318k). Outcome measured per decision-grade output.
Illustrative assumptions for a fictional fund. At low volumes, fixed lines, platform and amortised implementation, dominate unit cost. That is the central fact of AI economics at fund scale.
Why fixed cost is the fund-specific problem
Spread £500,000 of build and platform cost across 5 million resolutions and it disappears. Spread it across 14,000 outputs and it is £36 each before a single token is consumed. The implication is straightforward: at fund scale, buy before you build, keep the fixed base small, and prove capacity value on a narrow workload before committing to a platform.
And state the benefit in the fund's own terms, coverage breadth, faster diligence, fewer operational breaks, rather than in FTE reductions you have no intention of making.
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.