AI Strategy Is a Capital Allocation Problem

Brad Wolfe leads with a statement on his LinkedIn instead of a CV: AI strategy is a capital allocation problem. That sentence is doing a lot of work. It tells you who he thinks should own the AI agenda inside a private equity-backed business — the CFO — and why everyone else who tries to own it is, in his view, setting the firm up to fail.

Brad has been a CFO for 30 years. His firm, Wolfpacks Consulting, has spent the last year embedded in roughly 75 portfolio companies on AI and AI-related projects. For 17 of them, his team runs the entire back office — finance, legal, admin, IT — and treats it as a live sandbox for figuring out what AI can do across the lead-to-cash motion. We sat down with him to ask the question every fund and management team is asking right now: how do we actually do this without blowing ourselves up?

The conversation is worth watching in full. The takeaways below are the ones that have been repeating in our heads since.

"If your CFO isn't on the front end, you can go to bed and wake up out of business"

The case for putting the CFO at the center of AI is not philosophical. It is a risk argument.

AI compresses three things that used to live in different timeframes — your forecast, your operations, and your business model — into one real-time loop. That is a feature when it works. It is catastrophic when it doesn't. A revenue recognition rule applied incorrectly by a human costs you a quarter of finance time. The same rule applied incorrectly by an AI agent across every customer in the system costs you the company.

AI will send out a million wrong bills. The force multiplier is so much greater that it takes what would be minor explosions and turns them into nuclear explosions. That's also why you should treat it with great respect — because it will kill you if you don't know what you're doing.

— Brad Wolfe

The CFO is the only role in the company that owns the controls, the corporate governance, and the audit trail across every workflow AI touches. Hand the agenda to the CIO or a head of innovation and you get something that is technically competent and operationally invisible — until it isn't.

Three ways to deploy AI. Two of them are a disaster.

Brad sorts AI deployment patterns into three buckets. Two of them, in his words, you should "stop immediately."

Individual. Hand every employee a co-pilot and let them experiment. The problem is that prompts are directions to the system — they are code, not creativity — and an ungoverned set of prompts is an ungoverned set of automations sitting on top of your processes. It does not drive efficiency. It introduces ambient risk.

Departmental. Each function — finance, sales, marketing, HR — builds its own AI stack. This is the most dangerous pattern, because AI's value is in compressing models together. Seven departments each building their own version multiplies the integration debt you'll eventually have to rip out. You inherit seven incompatible stacks instead of one.

Transformational. A single, company-wide playbook that defines how AI gets deployed, where the data is clean enough to feed it, and what controls govern its outputs. Departments still do their work — but they do it as part of an architecture that's been designed to connect.

The bridge analogy Brad uses in the conversation is the one to remember. I'm building half my bridge from the east, you're building half from the west — when we meet up, they connect. If you don't have a shared blueprint, the two halves don't meet.

The 80/40 number — and what it really means

Across the 17 portfolio companies running AI through Wolfpacks' shared services model, AI has reduced back-office expense by roughly 80% over the last eight months. But 40% of that expense came back. It didn't disappear — it moved.

AI works really well when you have clean data that's organized in a certain way and it feeds through your machine. So you've got to make sure you're getting clean data. You also have to make sure you have corporate governance, controls, the setup in place.

— Brad Wolfe

The savings come from automating reconciliation, billing, reporting, document handling, and the mid-cycle finance work that consumed the most analyst time. The added-back cost goes to the front end of the data pipeline — to the people, processes, and tooling that ensure AI is being fed clean inputs and that anomalies get caught before they propagate. Net-net, the cost base is dramatically lower. But the shape of the cost base changes. The boiler room shrinks. The bridge gets bigger.

"We won't lay anyone off because of AI."

The single most counterintuitive piece of advice in the conversation is also the one most likely to be ignored.

Brad's funds have committed publicly to their portfolio company employees: AI will not cause layoffs. Attrition will be backfilled by AI before it's backfilled by hiring. Roles will move — from the back office to the front, from execution to oversight — but no one loses their job because a model can do part of what they used to do.

The reason is not sentimental. It is operational.

We need our employees to be excited about AI. AI is the future, whether you like it or not. We want them to learn, even if their job's going to change. If they're afraid of it, they sabotage it.

— Brad Wolfe

The companies seeing the cleanest AI deployments are the ones where the people inside the company are actively looking for places to deploy it. The companies that have stalled are the ones where employees have correctly read the signal that AI is a threat and have responded the only rational way: by hiding the parts of their job that could be automated.

The CFO as Mr. Spock and Scotty

Brad's analogy for the CFO of the next five years is one of the better ones we've heard. The CFO becomes both Mr. Spock and Scotty: half the time on the bridge thinking strategically about where the company is going, half the time in the engine room making sure the machine actually works.

In practice, three things change:

• The CRM moves into the control structure. A CRM is no longer a sales tool. It is a system of record that feeds the ERP and the data warehouse, and that means it needs the same audit trail, governance, and reconciliation discipline as the general ledger. In Brad's portcos, the CFO owns the CRM. Sales gets the output.
• HR shrinks. Finance grows. Once you're managing agents instead of people, you're not running an HR function — you're running a finance function. Agents have utilization, cost, error rates, and ROI. People have managers.
• Audits become continuous. The 15-day close gives way to a real-time loop. Audits don't go away — they shift from validating outputs after the fact to verifying that the processes producing those outputs are correctly designed and continuously monitored.

Humans in the loop, redesigned

One of the most operationally specific points in the conversation is about review.

The default model — AI produces output, a human reviews it before it goes out — is broken, because humans habituate. After 30 days of correct outputs, no one is reading them anymore. The 1% failure case lands in production unread.

The replacement model Brad describes is structurally different: humans are incentivized to go look for anomalies rather than to validate every output. They run the system through targeted queries designed to catch the kinds of errors AI is most likely to produce. It is a posture borrowed from internal audit, applied continuously.

And not everything goes through the AI pipeline at all. Roughly 10% of deals in Brad's portcos bypass the AI system entirely — because the data is too gray, too contractual, too subject to interpretation. Forcing that 10% through a model that's optimized for clean, structured input is exactly how you get a hallucination that costs $100 million.

AI defensibility — the question Brad asks first

Before any AI deployment work begins, Brad asks one question of every portco he walks into: Are you AI defensible, or can you become AI defensible?

The answer determines whether AI is going to compound your moat or commoditize your business out from under you. The most defensible position is proprietary data — data you've collected over years that no competitor can replicate without going through you. Environmental records, customer behavior history, longitudinal performance datasets in regulated industries. AI doesn't help you build that data. AI is what makes the data more valuable than ever.

If you're not AI defensible and can't become so, the next question is how much runway do you have? One year, three years, five, ten? That tells you how to position the asset — for an exit, for a roll-up, for a managed wind-down. What you cannot do is pretend the question doesn't apply to you.

For funds, this changes due diligence. M&A timelines are compressing — five months of diligence becoming five weeks, in Brad's experience — but integration capacity is the bottleneck. You can do 25 deals a year if your portcos can absorb 25 deals a year. Most can't.

Where to start

For a CFO or fund partner staring at a blank page wondering where do we start, Brad's sequencing is unromantic and effective:

1. 01Settle the AI defensibility question first. Until you know whether AI is your moat or your risk, every deployment decision is uncalibrated.
2. 02Run a high-level lead-to-cash assessment. Where is your data quality strong? Where are your processes brittle? Where does AI have leverage? You don't need a six-month consulting engagement to answer this.
3. 03Fix the foundations before you deploy. The shortcut every consultant pitches — skip the data work and deploy the model — is the shortcut that produces the most expensive mistakes. The companies that have leapfrogged foundational work are the ones writing the largest checks to clean it up later.
4. 04Build for connection, not for fit. Every individual AI project should fit inside a unified plan that knows where the data flows next. If a project doesn't fit the plan, change the plan or change the project.
5. 05Pick partners carefully. "If you choose the wrong partner — like if you choose the wrong spouse — life is difficult." Vendor selection is a multi-year decision in a market moving every six months.

The full conversation goes deeper into each of these — including specifics on shared service architecture, AI-native vs. legacy tooling, and how the C-suite changes when accountability becomes real-time. Worth the 47 minutes.