Allocating capital across your portfolio, with AI as your analyst
Every business-unit head argues for more, the confident case wins, and last year's budget quietly decides this year's. Here is how to make AI cross-examine each business case before it ranks anything, so the money follows the economics rather than the loudest voice in the room.
Good AI for capital allocation does not rank your portfolio for you. It cross-examines every business case before any comparison: where the optimistic assumptions sit, what would have to be true for an NPV to hold, which number is unsupported. The Havruta Methodology (formerly the Think Partner Methodology) installs that discipline. You instruct the machine to interrogate each submission rather than tidy it into a list, and you anchor it in one verified set of actuals, capacity and assumptions instead of each unit's own template. That is the Flip working over Ground Truth. The result is a defensible comparison, where the money follows the economics rather than the most confident voice. This is not a scoring tool. It is how the leader reasons with AI on the allocation they have to defend.
The situation
The funding round comes round again. Each business-unit head arrives with a case for more, and the numbers on your desk lean one way: a peer-reviewed survey of executives found that more than half say cash flow and NPV forecasts are biased upward, and only about a third consider the forecasts reliable (Hoang, Gatzer and Ruckes, 2024). So the case that wins is often the one argued most confidently, not the one with the best economics. In the same study, soft factors quietly outrank the maths: the strategic story headquarters likes, and your read of the manager, rank above IRR and NPV in the actual decision. Underneath it all, inertia does the deciding. The split drifts incrementally year on year, and almost nobody spends even a day asking whether last year's allocation still fits the plan. You have the bandwidth of one person and a portfolio that needs an analyst.
What commodity AI does with it
Hand an assistant the business cases and ask it to rank them, and it returns a clean league table in seconds: scores, a weighted matrix it invented, a confident order. It reads like analysis. It is the vending machine: a tidy answer to the wrong request. The risk is sharper here than with most tasks, because people tend to follow an algorithm's recommendation more closely than a human's and scrutinise it less, and a controlled study found that putting a human back in the loop raised uptake while lowering average decision accuracy (Sele and Chugunova, 2024). A black-box ranking does not strip the politics out of the room. It rubber-stamps the same politics with a veneer of rigour, and you will trust it more for looking certain. AI is now a top strategic priority for finance leaders, yet many say they are unsure they can govern it (Coupa via CFO.com, 2026), and this is exactly why: the easy use of it makes the decision worse.
The Flip: cross-examine each case, do not rank it
The move that changes the task is to stop asking the machine to order the portfolio and start making it interrogate each case before it answers anything. Instead of "rank these", the instruction becomes "question every one of these: where is the optimistic assumption, what would have to be true for this NPV to hold, which number is unsupported by the data we have". An AI made to think that way does not hand you a league table. It behaves like a sceptical analyst who cross-examines every submission, which is precisely what strips confidence-as-currency out of the room. The loudest case no longer wins on volume. It has to survive the questions.
What the machine must ask before it ranks
Cross-examination is only as good as the facts behind it. Each unit brings its own template and its own optimistic assumptions, and data quality is the most-cited barrier to getting value from AI in investment work. Feed the model that mess and it launders inconsistent inputs into one confident answer. So before it compares anything, it has to request what no submission alone can give it:
-
One verified set of actuals: the same numbers, the same definitions, across every business unit, not each unit's own template.
-
For each case, the assumptions the NPV depends on: which are evidenced, which are hopeful, and what would have to be true for them to hold?
-
The real capacity behind each bet: can the unit actually deliver it, and what else is already drawing on the same people and cash?
-
The net economics, not the gross: the return from funding each case set against what it costs, and against the next-best use of the same capital.
-
The do-nothing baseline: what last year's split has actually returned, so inertia is a choice you make, not a default you inherit.
That request is Ground Truth, the methodology's discipline of anchoring AI in your verified numbers rather than its generic knowledge or each unit's self-interest. It is also where the Mirror Principle earns its keep: if the comparison comes back bland, or simply re-ranks by who argued hardest, that is not a sign the units are equal. It is a sign the input was generic, the prompt lacked Ground Truth or a real Flip. The mirror tells you to sharpen the question, not to sign the answer.
Rank these six business cases for the funding round and tell me where the money should go.
Before I rank anything: are these built on one consistent set of numbers, or did each unit use its own template and its own assumptions? If it is the latter, any order I give you is just their optimism, sorted.
Each unit built its own. The flagship case has the highest NPV by a distance.
Then that NPV is the first thing to question, not the reason to fund it. What is it resting on, a market-growth rate, an adoption curve, a cost saving? Which of those is evidenced in the actuals, and which is the case author hoping? And what does the flagship cost net, against the next-best use of that capital?
The growth rate is double what that market has ever done. The quieter unit's case is smaller but every number traces to last year's results.
Then on the evidence you have, the quiet case is your fundable one, and the flagship is parked until that growth assumption is defended, not cut. Let me write the comparison with each case's load-bearing assumption and net return attached, so it survives the room.
What you walk away with
Not a league table. A defensible comparison where every case carries the reasoning that placed it: the load-bearing assumption it rests on, whether the data supports it, the net return against the do-nothing baseline. It survives a peer or a board member asking why this and not that, and it tells the unit whose case was parked exactly what it would take to fund it. The bandwidth you never had, doing the first-pass interrogation across every case at once, is what gives you back the time to revisit the split against strategy. That compression is Decision Velocity, worked on your portfolio rather than a template.
- Fund now
The quiet unit's case
Numbers trace to actuals Net return clear Capacity to deliver - Parked, pending defence
The flagship case
High NPV Growth assumption unproven - Held, with the reason
The headquarters favourite
Liked by HQ Unsupported number
The 4-Lines you can run yourself
-
Act as a sceptical capital-allocation analyst who assumes every business case is biased upward until the numbers prove otherwise.
-
Goal: a defensible comparison of these cases, where each is ranked by net economics against the do-nothing baseline, not by how confidently it is argued.
-
Interrogate each case before you rank anything: ask me for one consistent set of actuals, then name the load-bearing assumption in each NPV, which numbers are evidenced, the real capacity to deliver, and the net return against the next-best use of the capital. Do not conclude until you have what you need.
-
Ask one question at a time, step by step.
Frequently asked questions
How do you use AI for capital allocation across a portfolio?
Not by asking it to rank the portfolio. You instruct the machine to interrogate each business case first: where the optimistic assumptions are, what would have to be true for the numbers to hold, which figure is unsupported. Before any comparison, you anchor it in one verified set of actuals, capacity and assumptions rather than each unit's own template. The output is a defensible comparison you can take to the table, not a confident ranking that launders the same politics into a tidy list.
Can AI remove bias from capital budgeting?
It can expose bias, if you use it as an interrogator rather than a ranker. A peer-reviewed survey of executives found more than half say cash flow and NPV forecasts are biased upward (Hoang, Gatzer and Ruckes, 2024). An AI made to cross-examine each submission surfaces where a case is propped up by an optimistic assumption. But an AI you let rank the portfolio unchallenged can swap one bias for another, since people follow algorithmic recommendations more closely and scrutinise them less (Sele and Chugunova, 2024).
Should I just let AI rank my investment options?
No. That is the failure the evidence warns about. In a controlled study, participants followed an algorithm's recommendation more closely than a human's and adjusted it least where the error was largest; adding a human monitor raised uptake but lowered average accuracy (Sele and Chugunova, 2024). A black-box ranking can rubber-stamp the same politics with a veneer of rigour. Use AI to question each case and stress-test the comparison, then make the call yourself.
What stops AI from laundering my business units' optimistic numbers?
Ground Truth. Before any comparison, you anchor the model in one consistent set of verified actuals, capacity and assumptions rather than each unit's self-serving template. Data quality is the most-cited barrier to getting value from AI in investment work, and for good reason: feed it inconsistent inputs and it returns a confident-sounding answer built on them. Anchored in a single verified picture, the same case can land in a different place because it is now measured against reality.
How does this speed up the allocation cycle?
By doing the first-pass interrogation across every case in parallel. With the Flip and Ground Truth cross-examining each submission at once, the comparison compresses from a months-long incremental ritual to a fast, evidence-led review. That is Decision Velocity: you reclaim the analyst bandwidth you never had, so you can actually revisit the split against strategy rather than defaulting to last year's budget.
Where do we start?
With the Eye-Opener Workshop, a half-day where your leadership team sees the shift on a live allocation question of your own. A Strategic Briefing maps the right entry point.
References
- Hoang, D., Gatzer, S., Ruckes, M. "The Economics of Capital Allocation in Firms: Evidence from Internal Capital Markets." Management Science, November 2024.
- Sele, D., Chugunova, M. "Putting a human in the loop: Increasing uptake, but decreasing accuracy of automated decision-making." PLoS One, February 2024.
- Coupa, Strategic CFO Report, reported by CFO.com. "AI investment among top strategic priorities for CFOs: survey." CFO.com, April 2026.