AI Services HoldCo Returns: An Investment Memo

Buy a service company at a service multiple, add artificial intelligence (AI), and hold or sell a business with software economics. That is the pitch. The return depends on who captures the gain after acquisition price, engineers, debt, integration, and time.

An AI services holding company (HoldCo) acquires labor-heavy service businesses and rebuilds their workflows with software and AI instead of reselling them intact. This memo tests AI services HoldCo returns through seven drivers and the public evidence from Sequence Holdings, Long Lake, and Enam (Sequence, Long Lake, Enam).

It is a framework, not an investment recommendation or a forecast of private-company performance. The conclusion is direct: the operating model is credible, but software-like portfolio returns are not yet proven in public.

Key takeaways

• AI services HoldCos buy labor-heavy service companies at service multiples and aim to earn software economics by rebuilding workflows, but no public portfolio proves that outcome yet.
• Long Lake agreed to acquire American Express Global Business Travel (Amex GBT) for about $6.3 billion, a 60.2% premium to the previous close, before any transformation gain exists (Amex GBT).
• Operating alpha in an AI roll-up comes from growth, throughput, quality, and reuse, while cheap entry, debt, and multiple expansion are financial mechanics that belong in a separate line.
• Long Lake's Nexus platform is the category's clearest named shared infrastructure, with chief executive Alex Taubman claiming roughly 80% reuse across verticals and no published cost data (Travel Weekly).
• Underwrite an AI services HoldCo on a flat exit multiple, delayed benefits, and funded human review; if the deal still works, AI is upside rather than a requirement.

The base case should work before the terminal multiple expands. A buyer pays cash or issues debt today, absorbs integration and engineering costs, and waits for operating changes to reach customers. If benefits arrive late or service quality falls, financing can consume the upside. The right underwriting model separates ordinary acquisition returns from AI-driven growth and margin, then tests whether the latter survives central costs and continuing human review.

The return equation: What has to be true

Software-like returns means revenue and profit that grow faster than the labor and capital needed to deliver them, not simply a higher multiple of earnings before interest, taxes, depreciation, and amortization (EBITDA) at exit. The ownership structure behind that ambition is the AI services HoldCo model and the three companies testing it.

The return equation is:

Entry value + organic growth + labor productivity + shared-platform reuse - central engineering - integration - financing cost = equity value before exit multiple or permanent cash yield.

The components fall into two groups.

Operating alpha comes from growth, throughput, quality, automation, and reuse. Buying cheaply, using debt, and selling at a higher multiple are financial mechanics. Both can create value. Counting them together as AI performance makes the thesis impossible to test.

The size of the prize is what makes the category interesting. An analyst walkthrough on The Long Circuit sketches the arithmetic: where 60 to 70 dollars of every 100 in service revenue goes to frontline labor, pushing that share down to 30 to 40 dollars turns a roughly 10% operating margin into 30% to 40%. The same explainer puts the capital behind the bet at a reported $1.5 billion General Catalyst fund for AI-enabled service businesses and more than $1 billion at Thrive Holdings (video explainer). That is an illustrative model, not a measured portfolio result. The distance between the two is what this memo tests.

No company publishes enough portfolio data for a complete calculation. Sequence explains its permanent-owner design, Long Lake discloses a signed large transaction and its Nexus platform, and Enam describes a central transformation engine (Sequence, Amex GBT, Enam). Each criterion therefore gets an evidence judgment and a list of missing variables.

1. Entry multiple: Did the buyer pay for the AI upside in advance?

The easiest transformation gain to lose is the one paid to the seller at entry.

Start with revenue and EBITDA multiples, then test the quality underneath them. A low multiple can reflect an under-managed asset. It can also reflect a structurally difficult business.

The Amex GBT agreement is the only transparent purchase-price example in this group. Long Lake agreed to pay approximately $6.3 billion in cash, or $9.50 per share. That represented a 60.2% premium to the previous close and 65.1% to the 30-day volume-weighted average price. The deal was expected to close in the second half of 2026 and remained subject to approvals and conditions (Amex GBT). Long Lake did not yet own the company. The announcement does not provide the post-deal EBITDA, equity contribution, debt ratio, or financing-cost inputs needed to calculate an entry multiple and an equity return. The premium does show that the transformation starts with a meaningful hurdle.

Practitioner numbers sit at the other end of the size range. Jeremy Yamaguchi, founder of the pool-services roll-up Cabana, told the Verticals podcast that he buys operating companies at three to four times EBITDA and recycles their cash flow into more acquisitions and product development. He also puts the risks in order: buy at sensible multiples first, integrate second, expand margins and centralize the back office third, grow organically fourth, and only then build proprietary vertical software (video interview). That is one operator describing small private deals rather than a category benchmark, and integration is the step he calls harder than acquisition. The same private-multiple arithmetic sits under the MSP roll-ups Shield and Titan.

Sequence says it takes meaningful ownership stakes but publishes no acquisition prices or portfolio list (Sequence). Enam says it acquires service businesses but does not publish transaction terms or acquired-company identities (Enam).

My base case gives no credit to multiple arbitrage without acquisition cohorts and entry valuations. Underwrite the current cash flow first, then add AI upside only if the price and the operating plan leave room to earn it. I would also price the trusted distribution, not only the dollar of EBITDA. Customer relationships and transaction volume give the central engine somewhere to deploy, but only if contracts survive the ownership change.

Evidence grade: moderate for Long Lake's announced price and premium, low for portfolio-wide entry economics. The announcement supplies consideration, not the capital structure or a return bridge. Sequence and Enam disclose too little to grade acquisition discipline publicly.

2. Organic growth: Does productivity create more demand or just fewer people?

A faster workflow does not tell you who receives the benefit.

Released capacity has four main uses: remove labor, serve more volume, lower prices, or improve service. Only the first appears immediately as cost reduction. The other three can be worth more if they raise customer wins, retention, or willingness to pay, and the choice mostly gets made in the pricing model for AI-delivered services.

Long Lake and Amex GBT describe a forward-looking model in which AI and human agents work together across booking, proactive disruption resolution, and travel administration. The stated goal is faster and less frictional service, not only lower staffing (Amex GBT). Those are intended outcomes after a pending acquisition, not realized results.

Sequence says ownership allows it to redesign businesses with incumbent management, but it publishes no acquisition-cohort growth data (Sequence). Enam says its model makes expertise abundant and delivers outcomes at greater scale, without public portfolio growth figures (Enam).

The strongest thesis is capacity-led growth. If a service business can handle more transactions at stable quality with the same core team, revenue can grow without matching headcount growth. A one-time layoff raises margin faster, but it proves nothing repeatable and it can damage the customer franchise.

Proving demand capture takes a cohort design. Compare customers exposed to the new workflow with similar customers still on the old process, then measure organic revenue growth, win rate, gross and net retention, price, response time, service levels, and satisfaction. Control for market growth and acquisitions, because a portfolio-level growth number can rise on deal volume while the underlying service stagnates. Management should also say where the released capacity went: moving employees from data entry to exception resolution or sales can be rational, but it still needs a path to cash flow.

Evidence grade: low. The public sources establish intended service improvements and ownership theses, not realized organic growth attributable to AI.

3. Labor productivity: How much work disappears and how much human review remains?

Automation percentages are among the least useful numbers in a deal deck. The expensive exceptions determine the economics.

Measure time per transaction, transactions per employee, straight-through processing, exception rate, human review minutes, error, rework, service quality, and total headcount relative to volume. Track them together. Removing 40% of manual steps means little if the remaining work needs senior reviewers or creates customer remediation.

Enam's technical recruiting materials show the intended operating work: engineers integrate legacy data and production systems, automate workflows, increase throughput, and reduce cost inside acquired businesses (Enam careers). That supports a hybrid transformation model, not labor-free software. Long Lake's announced Amex GBT plan similarly combines AI with human agents across core travel workflows (Amex GBT). BankSouth's partnership with Sequence is also framed as human-first, with embedded engineers building tools inside the bank (BankSouth). None of these sources publishes a standardized productivity cohort.

A 20% faster workflow is not a 20% EBITDA gain. Employees may spend the released time on more customers, higher quality, exception review, sales, or new services. Central software costs, model usage, and monitoring absorb part of the saving. The labor-share model quoted earlier assumes frontline cost actually leaves the income statement, and in a hybrid team it usually moves rather than disappears.

The reconciliation should start with one workflow. Count total cases, automated cases, reviewed cases, escalations, errors, and rework, then add every input cost: employee and contractor time, model usage, software, monitoring, and central engineering. Compare the fully loaded cost per correct completion before and after.

Headcount misleads in both directions. A company may retain staff because demand grew, hiding real productivity. It may also cut staff while service deteriorates, creating false efficiency. Revenue per employee has to sit beside customer satisfaction, retention, and error cost.

Evidence grade: low. Recruiting pages and partnership announcements prove that hybrid teams exist. They do not publish the operating measures required to value the gain, and I would not capitalize a productivity claim until those numbers reconcile to the income statement and the customer outcome.

4. Central engineering: Does the platform compound or become a permanent cost center?

The most software-like property in the model is not AI. It is reuse. The same question decides whether forward deployed engineering compounds inside a services firm.

Long Lake has the clearest named platform. Nexus is presented as proprietary transformation infrastructure used across service businesses. Travel Weekly reported chief executive Alex Taubman's claim that roughly 80% of that infrastructure can be reused across verticals, with the remainder devoted to local deployment and workflow integration. The same interview said early deployments took more than a year to show business outcomes while later ones could deliver immediate time savings (Travel Weekly). Those are executive claims reported by an independent publication, not audited cohort data.

Sequence publishes an embedded-engineer model but no named horizontal platform (Sequence). Enam describes a central technology transformation engine and roles spanning legacy integration, workflow automation, and production deployment (Enam careers).

Test reuse through six measures:

• Percentage of shared production code and infrastructure.
• Deployment time by acquisition cohort.
• Central engineering spend per holding.
• Maintenance and incident burden across the installed base.
• Domain-specific data and integration work.
• Workflows launched per engineer without quality decline.

The denominator matters. Identity, observability, evaluation, and deployment tooling can be highly reusable while the decision workflow remains bespoke. An 80% infrastructure claim can still leave most of the labor inside the final 20%.

The central team also creates a make-or-buy question. Shared identity, security, deployment, evaluation, and monitoring can justify one platform. Industry-specific workflows may be better owned locally. Pushing every decision into the center slows operators and creates one queue for the whole portfolio.

Allocate central cost to each holding, including product development, maintenance, incidents, data engineering, and senior management time. Then calculate the payback from reused assets. Code that appears in ten repositories but needs ten specialist teams to maintain it is not software-like reuse.

Evidence grade: moderate on Long Lake's named Nexus architecture, low on realized economics. No company publishes enough cost and cohort data to prove that central engineering compounds rather than becoming a permanent internal consultancy.

5. Financing and integration: Can the balance sheet survive the transformation period?

Debt is helpful only when cash arrives before interest, integration problems, and customer churn consume the equity cushion.

The pending Amex GBT acquisition is financed with equity from Long Lake's existing investors and Koch Equity Development plus committed debt from JPMorgan, Bank of America, Citi, and MUFG. The announcement does not disclose the equity contribution, debt ratio, pricing, or ownership split (Amex GBT). That is conventional buyout finance attached to a new operating thesis. The integration still has to sequence management retention, customer communication, regulatory approval, data cleanup, legacy systems, workflow changes, employee adoption, and the rollout of Nexus.

The announcement's own risk list runs through closing conditions, regulatory clearance, financing, management distraction, employees, customers, integration, and expected benefits. Standard disclosures, but they map the transformation period accurately (Amex GBT).

The category's cautionary parallel is not an AI company. Thrasio, the Amazon-brand roll-up, raised roughly $3.4 billion and briefly carried a $10 billion valuation before filing for bankruptcy protection in February 2024, once inventory, debt, and failed integration arrived together. The analyst summary is that expansion became a way of buying more problems (video explainer). The mechanism transfers to AI roll-ups without much translation. Acquisitions compound faster than integration capacity, and interest does not wait for the transformation.

Sequence's permanent-capital claim may reduce the pressure of a fixed exit clock. Public materials do not disclose debt policy, portfolio cash-flow allocation, or governance, so that benefit remains structural rather than quantified (Sequence). The Information reports that Enam raised capital at a valuation above $300 million, but its deal-level financing is private, which puts deal-level stress testing out of reach from outside (The Information).

Model at least one year of delay, customer churn, and no exit-multiple expansion, then fund the central team through that downside. Permanent ownership absorbs a longer build, but no fixed exit date does not remove the cost of capital: cash tied up in one slow transformation cannot fund the next acquisition. Ask for internal return hurdles and the rules for stopping projects.

I would phase transformation capital. Release the first tranche for data access and a production wedge, the second after quality and adoption thresholds, and the third for expansion. That does not remove uncertainty, but it stops a broad platform program from consuming cash before one workflow proves value.

Evidence grade: moderate for the pending Amex GBT financing structure and its stated risks, low for portfolio debt and integration policy across the category.

6. Recurring revenue quality: Is the service base worth owning before AI?

AI cannot rescue weak contracts, customer concentration, poor retention, or a service customers do not value.

The pre-AI underwriting checklist is familiar: contract recurrence, gross retention, concentration, pricing power, switching cost, trust, cash conversion, cyclicality, working capital, and liability. Start there before analyzing models or automation.

Long Lake is not only buying Amex GBT's workflows and data. The announcement highlights customer relationships, a global marketplace, brand, and service capabilities as strategic assets (Amex GBT). Those assets create the distribution and trust through which a new operating system can reach customers. The best HoldCo target is already a defensible service company. AI can add capacity, quality, and speed. It cannot manufacture customer trust overnight or fix a structurally bad market.

Operators building this model from scratch watch a short list. Brennan Pothetes, chief executive of Infinity Constellation, described judging portfolio health on contracted annual recurring revenue (ARR) and new ARR first, then whether that revenue genuinely recurs or is one-time, then gross margin above 50% and preferably above 60%, then net revenue retention above 100% (video interview). His holdings are built rather than bought, but that is the right shape of dashboard for an acquired service base, and none of Sequence, Long Lake, or Enam publishes it.

Sequence and Enam do not publish enough portfolio detail to assess revenue quality across holdings (Sequence, Enam). An investor needs contract cohorts, retention, concentration, price changes, customer acquisition cost, and gross margin before and after the workflow changes.

Recurring revenue also needs classification. I would split it into contracted human service, transaction-based service, managed operation, implementation, software, and pass-through model cost. For each line, measure gross retention, incremental gross margin, working capital, and customer concentration. Contracted service revenue can be durable while remaining labor-intensive, and managed operations can retain customers while carrying heavy exception costs. Ask what recurs, why it recurs, and what labor follows every additional dollar.

Customer trust is an economic asset in its own right. A bank, a travel manager, or a specialist adviser reaches decisions a new software vendor cannot. Automation that damages that trust destroys part of the acquisition thesis, so track complaints, escalations, and renewals during the transformation, not only cost per transaction.

Evidence grade: moderate for Amex GBT's disclosed strategic assets, low for Sequence and Enam portfolio quality.

7. Exit multiple and time: Who pays for a transformed service business?

The phrase software-like returns often enters the forecast through the terminal multiple rather than the operating plan.

There are three endgames:

1. Hold permanently and collect a growing cash yield.
2. Sell a faster-growing, higher-quality service business.
3. Re-rate the company as a technology-enabled platform.

Sequence explicitly presents itself as a long-term owner without a fixed fund clock (Sequence). Long Lake and Enam do not publish complete exit policies (Long Lake, Enam).

A future buyer will price revenue mix, gross margin, recurrence, organic growth, customer concentration, central intellectual property, capital needs, and management dependence. The AI label itself earns no multiple.

My base case uses the same entry and exit multiple. Improvement comes from organic growth, margins, cash conversion, and debt paydown. Then I run sensitivities for transformation delay, operating gain, financing, and exit valuation.

Time can destroy an attractive annual improvement. Central cost and interest accrue while benefits are delayed, and the present value of the exit falls. The model should work on cash flow before assuming that a buyer pays a software multiple later.

A re-rating should require evidence that survives a sale process: recurring growth, customer retention, stable service, a central platform used across holdings, lower deployment cost by cohort, and a management team that operates without a few founding engineers. A pitch about AI adoption will not substitute for those attributes.

For permanent ownership, calculate distributable cash after continued central investment and acquisition needs. A holding company can report rising EBITDA while retaining all cash for transformation. That may create value, but it is not the return profile of a mature cash-yield asset.

Evidence grade: low. Sequence states its long-term intent, while the category lacks disclosed exits or portfolio cash yields that isolate the AI contribution.

The Bottom Line

Ownership genuinely solves several barriers that slow enterprise AI: workflow access, data access, procurement, and incentive alignment. It also introduces acquisition price, financing, integration, and operating risk.

Long Lake supplies the strongest public transaction and platform evidence through the signed but pending Amex GBT agreement and Nexus. Sequence has the clearest permanent-owner design and a named BankSouth deployment partnership. Enam has an explicit acquisition and central-engineering thesis, but keeps its portfolio and operating outcomes private (Amex GBT, Sequence, Enam).

Public data does not yet prove software-like returns. Five disclosures would change the verdict: acquisition cohorts, entry multiples, organic growth, central engineering cost and deployment time, and a reconciliation of return to operating improvement, debt, and terminal multiple.

Underwrite these companies as service-business acquirers with an AI upside case. Require the operating gains to outrun price, engineering, integration, financing, and time before paying for software economics. Show that later acquisitions deploy faster, reuse more infrastructure, deliver equal or better service, and produce more cash after central cost.

Until that appears, use a flat exit multiple, delay the benefits, and fund the human review. If the deal still works, AI is upside rather than a requirement for survival.

For related analysis, see AI-powered roll-ups of managed service providers (MSPs), why the MSP roll-up era is changing, and MSP valuations in the AI era.

FAQ