The Business Case
A Fund’s Case for AI Is Four Arguments,
and Efficiency Is the Weakest of Them.
Most AI business cases for an investment firm stop at cost saving. Cost saving is real and it is the smallest of the four things recovered hours are worth. What follows is the arithmetic, the sources behind it, and the cells where it is weak.
The Same Recovered Hours Read Four Different Ways.
A fund that puts AI into its document work gets hours back. What those hours are worth depends entirely on what the fund does with them, and the four answers differ by an order of magnitude.
One | Efficiency
The Cost Line Falls.
An equity team recovers roughly 335 document hours a year and a debt team roughly 550, on the modeling set out below. This is the easiest argument to prove and the smallest in value, and it is the one most exposed to the question of what the tools themselves cost.
Two | Growth
The Same Hours Read as Capacity.
A fund does not release an analyst when hours come back. It writes more business. A marginal equity deal costs about 318 document hours once the processes that enter diligence and die are counted, so 335 recovered hours is roughly one more closed deal a year, moving a fund from 3.5 closes to 4.6. A marginal debt position costs about 60 hours, so 552 recovered hours carries about nine more positions, a book of 73 rather than 64.
Growth also means the next fund. Every manager is raising one, and the same capacity goes into the market research behind the thesis, the co-investor mapping and the LP targeting.
Three | New Instruments
Smaller Tickets Become Viable.
When diligence and monitoring cost less per transaction, the minimum viable ticket falls. That opens blended structures with more tranches, outcome-linked and results-based instruments, and smaller checks into smaller companies.
We do not put a figure on this column. Fixed transaction cost is the reason ticket sizes cannot come down, and ticket size decides which enterprises are reachable at all, but nothing published supports a number and we will not invent one.
Four | Evidence
Impact Data Becomes Answerable.
Impact data arrives late, in different shapes, and is assembled by hand. In a July 2026 baseline survey we ran, eight participants across five impact funds rated their own use of AI in impact data collection and reporting at 2.88 out of 5.
A limited partner asking how impact is measured is asking a question about the next fund. Evidence answers it and narrative does not, which is why this argument and the growth argument belong together.
The first argument covers the cost of the tools. The last two decide whether there is a next fund.
Document-Heavy Work Is the Only Work an AI Workflow Touches.
The model below covers a fund of USD 200 million. Field visits and travel are excluded throughout, because they are not work a workflow reaches. Equity value sits in diligence, which is 69 percent of the document load. Debt value sits in monitoring a book, which is 63 percent of it. The two businesses are different enough that blending them into one figure hides the finding.
Equity
1,458 document hours a year
| Diligence on deals that close | 490 hrs |
| Diligence on deals that never close | 518 hrs |
| Portfolio monitoring | 270 hrs |
| Investor and LP reporting | 180 hrs |
Built on 3.5 closes a year at 140 desk hours, and a four to one funnel. Diligence on deals that die is the larger of the two lines, and a fund that does not count them understates its own load.
Debt
2,402 document hours a year
| Underwriting, new borrowers | 420 hrs |
| Underwriting, renewals | 476 hrs |
| Covenant and portfolio monitoring | 896 hrs |
| Annual credit review | 430 hrs |
| Investor and LP reporting | 180 hrs |
Built on about 64 positions at 0.9 deployment, replaced over roughly three years, so 20 to 30 transactions a year. A debt book holds far more positions than an equity portfolio, and far more repeating document work.
Sources. Diligence hours and funnel ratios from Gompers, Gornall, Kaplan and Strebulaev, NBER working paper 22587, a survey of 885 institutional venture capitalists reporting about four full diligence processes per closed deal and 118 hours of diligence per close, uplifted here for emerging market integrity and environmental, social and governance workstreams and reduced for the on-site component. Position count and ticket size from the Symbiotics Microfinance Investment Vehicle Survey, which reports an average of 39 investees at a USD 2.8 million average direct debt position. Underwriting hours from the CGAP tiered due diligence framework, where a Tier II review is 10 to 14 person-days.
A Fund Recovers More Than It Spends on the Tools, in Both Cases.
Hours are valued at a fully loaded blended rate of USD 50, derived from investment analyst total cash across fourteen cities, loaded at 1.5 times to cover payroll, workspace, IT and data subscriptions, over 1,800 productive hours. A fund on New York, London or development finance institution grade scales runs at roughly double. Recovery is modeled at 23 percent, the midpoint of a 10 to 35 percent band we set deliberately below the published range.
Equity
USD 200m fund, globally staffed team at USD 50 an hour
| Line | USD |
|---|---|
| Hours recovered at 23 percent | 335 hrs |
| Gross value of those hours | 16,800 |
| Less the cost of running AI | (11,160) |
| Net gain in the year | 5,600 |
This is the weakest cell in the whole model and it is the honest answer for a small emerging market equity team. At USD 100 an hour the same fund nets about 19,400.
Debt
USD 200m book, globally staffed team at USD 50 an hour
| Line | USD |
|---|---|
| Hours recovered at 23 percent | 552 hrs |
| Gross value of those hours | 27,600 |
| Less the cost of running AI | (11,160) |
| Net gain in the year | 16,500 |
Recurring portfolio work alone recovers more than the entire cost of running AI, before a single new loan is written. At USD 100 an hour the same book nets about 41,100.
What Running AI Costs
Steady state, a team of six, at USD 50 an hour
| Seats, six people | 2,160 |
| Model usage, about 1,500 document runs | 6,000 |
| Maintenance, 60 hours | 3,000 |
| Total | 11,160 |
Seat cost at USD 30 a month per person, in line with published enterprise assistant pricing. Model usage estimated on published token rates. The first year is heavier, around 160 hours of build, and sits outside this figure. Tell that to anyone you are making the case to, because a year-one number is worse and pretending otherwise invites the challenge.
Why the Band Is Conservative
10 to 35 percent, against the published evidence
McKinsey’s 2025 survey of 200 merger and acquisition practitioners found 40 percent of generative AI users reporting deal cycles 30 to 50 percent shorter, and 46 percent reporting faster due diligence. Apex Group’s March 2026 survey of 105 private credit leaders found 30 percent reporting reduced processing time in middle-office operations, which is adoption evidence rather than measured hours.
Both bounds of the band used here sit below those figures. We would rather a fund find the model conservative when it runs its own numbers than find it optimistic.
The Weak Cells Are Named Here Rather Than Found Later.
A business case that only shows its strong cells is a sales document. These are the four places this one can be pushed, and we would rather set them out than have a chief financial officer find them in the second meeting.
- Token cost is the least certain line. Model usage is estimated at 1,500 runs a year. Agentic workflows that loop can multiply that several times over, and doubling the usage line takes the equity case at the lower rate into a small loss. Run your own usage estimate rather than taking ours.
- The hours per process are judgment built on published baselines. They are not measured inside any one fund. The first thing worth doing is counting your own, which takes two weeks of timesheet discipline and settles the argument either way.
- The third argument carries no number. The cost side is modelable and the opportunity side is not. If someone asks for a figure on new instruments, the honest answer is that the first two arguments already recover the cost of the tools without it.
- Development finance co-funding is not in the model. Several institutions cover a share of the cost of external capability building. Where a facility applies, the fund’s own outlay falls and every figure above improves. The rates need verifying per facility, so none is assumed here.
Run the Numbers Against Your Own Fund.
The model is built to be checked. Send us your transaction count, your portfolio size and your view of the blended rate, and we will show you where your fund lands across the four arguments and which of them your team should be making internally.