What Strategy Firms Get Wrong About AI Deployment for Portfolio Companies

In the last six months, every major strategy firm has announced an AI deployment practice for private equity. The pitch sounds familiar. Big name, global footprint, relationships with the sponsor, a packaged offering for portfolio companies. Operating partners are taking the meetings.

We think the math breaks down at the portfolio level. Here is why.

The strategy-firm pattern

Strategy firms run AI deployment the same way they run everything else. A senior partner sells the engagement to the GP. A team lands at the portco for ten to twelve weeks. They produce a roadmap, identify use cases, write a maturity assessment, and present at the next board meeting. The slides are excellent. The recommendations are sound.

Then they leave.

What stays behind is a document. Sometimes a pilot built by a contractor the firm subcontracted to. Sometimes a Snowflake account and a Databricks workspace that nobody on the portco team knows how to operate. Often a partnership announcement with a single foundation model vendor that locks the portco into one stack for the next three years.

The portco CTO now owns a strategy without the team to execute it. The use cases that looked clean on a slide hit the messy reality of a fifteen-year-old ERP, a sales team that lives in Excel, and a data warehouse that nobody loaded last Tuesday because the cron job failed and the engineer who wrote it left in 2023. The strategy firm is already onto the next engagement. The bill is paid. The AI program is not.

What PE portcos actually need

A portfolio company in year two of a hold period has a different problem than a Fortune 100 enterprise. The exit clock is running. The team is lean. The data platform is whatever the founder built in 2019. The board wants AI in the next earnings narrative, not in a 2028 transformation roadmap.

Three things matter at this size:

Production, not pilots. A use case that ships to production in eight weeks and generates measurable revenue or margin impact in the next quarter beats a twelve-month transformation every time. PE boards reward shipped systems, not strategy decks. The metric a sponsor cares about at year-end is whether the EBITDA bridge has a real number on the AI line, sourced from a real system, audited by a real CFO. Pilots do not bridge.

Model optionality. The frontier model market moves faster than any single vendor partnership can keep up with. The right answer for fraud detection at a fintech portco is rarely the same as the right answer for clinical documentation at a healthcare portco. Lock-in to one model family at the portfolio level is a risk that compounds over the hold. A partnership announcement that looks like a procurement win in year one becomes a migration project in year three when the model that wins the workload sits behind a different API.

ISV-native architecture. Most portco software runs on a known stack: Snowflake or Databricks for data, dbt for transformation, Airflow or a managed orchestrator for pipelines, Salesforce or HubSpot for revenue, NetSuite or QuickBooks for finance. AI that lives inside that stack ships faster, costs less to maintain, and survives the next platform decision. AI built as a parallel system on a strategy firm's preferred infrastructure usually does not. When the next platform RFP comes around, the AI built into the warehouse comes along for the ride. The AI bolted on from the outside gets thrown out.

What we do differently

Blue Orange Digital has been building production AI inside PE portfolio companies since before AI deployment was a category. We do not write a roadmap and leave. We build the system, we operate the system, and we hand it off when the portco team is ready to own it.

Our typical engagement looks like this. Two or three senior engineers embed inside the portco for a quarter. They work alongside the existing team. The first deliverable is not a slide deck. It is a working data model in production, instrumented, monitored, and generating output that the business uses to make decisions.

From there, we layer on AI where it earns its keep. A pricing model that updates daily. A demand forecast that drives inventory decisions. A clinical workflow that compresses chart review from forty minutes to four. A claims triage system that routes the right cases to the right adjusters. A property leasing assistant that handles the first three rounds of tenant questions before a human gets involved. Each one ships to production, each one has measurable business impact, each one is owned by the portco team within six months.

We work with whatever model fits the use case. Sometimes that is OpenAI. Sometimes that is Anthropic. Sometimes that is a Databricks foundation model, a Llama variant, or a smaller domain-specific model that the portco can host inside its own VPC for compliance reasons. The choice is driven by the workload, by the data residency rules, and by what is actually cheapest at the volume the portco runs. Not by a partnership announcement.

When we hand off, the portco engineers know how the system was built, how to debug it at 2am, and how to retrain it when the data drifts. The runbook is short and concrete. The model registry is named. The eval set is checked into the repo. The hand-off is not a transition, it is a non-event, because the portco team has been doing the work alongside us the whole time.

Why this matters for the exit

Strategy firms talk about AI transformation. PE talks about exit multiples. The two are connected, but not in the way the strategy decks suggest.

What lifts a multiple at exit is operational evidence. A buyer wants to see margin expansion that survived a leadership transition, revenue per employee trending up, a data platform that scales without a rebuild, and an AI program that compounds rather than depreciates. None of that comes from a roadmap. All of it comes from systems that have been running long enough to prove they work.

The diligence question we hear most from buyers is the one the seller least wants to answer: show me the AI in production and walk me through the team that runs it. A buyer who hears "the engagement just wrapped, here is the deck" reads that as project cost without enterprise value. A buyer who hears "this model has driven thirty-eight percent of pricing decisions for the last fourteen months and the portco team owns the eval pipeline" reads that as a defensible operating capability.

Portcos that hire strategy firms for AI deployment usually have a deck. Portcos that hire builders usually have a track record. At exit, the deck does not bid up the multiple. The track record does.

What to look for

When you are evaluating an AI partner for a portco, the questions that separate builders from advisors are simple:

• Show me a system you built that is still running three years later, and the portco team operates without you.
• What does your team look like on day ninety, six months in, and at handoff?
• How do you decide which model to use for a given workload, and have you ever changed that decision mid-engagement?
• What is the smallest measurable business outcome you have shipped in the first quarter of an engagement?

Strategy firms struggle with these questions. Builders answer them in the first call.

Where to start

If you are an operating partner or a portco CTO evaluating the AI deployment landscape, the cheapest way to get clarity is to run a Blueprint assessment. We spend two weeks inside the portco. We look at the data platform, the use case backlog, the team, and the existing tooling. We tell you what is buildable in the next quarter and what is not. No deck. A working spec, with implementation paths, costs, and the team needed to execute.

Start a Blueprint assessment.