Contents
Executive thesis
The advantage of owning twelve companies is not that you can install the same AI software twelve times. It is that you can discover what works once, make the knowledge transferable, and use each later company to make the capability more effective, cheaper to run and easier to sustain.
That advantage is not automatic. Each company has its own customers, systems, economics, management team and operating constraints. The unit worth replicating is therefore a proven business workflow, supported by reusable software, clear operating rules, local tests, accountable owners and one consistent way of measuring value.
The central finding is that portfolio-scale AI adoption is better documented than portfolio-scale AI value creation. There are credible examples of technology and methods moving between portfolio companies. The public evidence reviewed does not establish a reliable cost curve, or independently verified financial returns, for repeated deployments across separately managed lower-middle-market companies.
A successful portfolio program answers five questions: What travels? What changes? Who owns it? Who adopts it? How do we know it worked? The governing principle: scale verified operating results, not licenses, enthusiasm or forecasts. The second company is where a promising AI project starts to become a repeatable portfolio capability.
How to read this report
Seven labels. Most PE cases are self-reported by a sponsor or a vendor, so the label matters as much as the number.
- ✓Research
- An academic study with disclosed methods. Often outside PE or outside AI; used for the mechanism.
- ✓Government
- An audit, evaluation, enforcement action, regulatory notice or public filing.
- ✓Benchmark
- A measured dataset across many organizations, with a disclosed definition.
- ƒSponsor-reported
- The PE firm's own account of its portfolio. Usually without a comparison group or independent check.
- !Vendor-reported
- The supplier's account of its own work. Commercially interested.
- ƒSurvey
- Self-reported responses from executives or practitioners.
- ◇Illustrative
- Invented planning numbers. Never presented as an observed result.
- Three levels of replication
- Technology deployment (the same product is available), operational replication (the workflow is adopted to the same standard) and economic replication (the receiving company realizes a verified net benefit). Only the third proves financial value.
The six findings that matter most
Each with its evidence, the caveats that must travel with it, the operating response and the decision it changes.
Case dossiers
Five named cases and one counterexample. They describe different parts of replication; they should not be merged into one fictional success story.
Which workflows travel
Similarity is established at the workflow level, not inferred from industry, sponsor or software vendor. This is an editorial assessment of portability, not a ranking from measured PE transfer outcomes.
| Candidate workflow | What could travel | What may change substantially | Transferability judgment |
|---|
A ten-business-day check at the receiving company
A proposed operating method, not an industry standard. One hundred historical cases can expose common problems but cannot prove safety against rare, costly errors: zero failures in about 300 independent random cases gives an approximate 95% upper bound of 1% on the failure rate, and still says nothing about a different population.
| Days | Work | Decision output |
|---|
Standardize the core, localize the last mile
Share what improves reliability and reduces repeat effort. Keep local the facts, rules and decisions that determine the business result. Standardize how quality is established; never assume one company's quality score establishes another's.
| Component | Standardize across the portfolio | Preserve or configure locally |
|---|
Adoption and ownership
A management team can reasonably resist an initiative that looks attractive to the sponsor: the cost lands now, the benefit later. These are differences in cost, incentive and accountability, not necessarily culture.
What a receiving company needs. A specific business outcome chosen by its CEO or functional leader, not an instruction to "use AI". Named implementation capacity. A demonstration by the source company's practitioners, not a sponsor deck. The authority to retire the old work, so the company does not pay twice. And the authority to decline or stop.
The adoption sequence. Eligible work, then people using the workflow, then a significant share of work processed through it, then sustained operation, then verified business benefit. Each answers a different question; never collapse them into one percentage.
| Owner | Owns |
|---|
A starting team. One accountable enablement leader, a technical specialist or trusted implementation partner, and fractional finance, security and legal support. No more than two deep implementations at once for a one-to-two-person central team. This is a capacity recommendation, not an established staffing ratio. The sponsor should also know who funds common work, who pays for local work, how vendor rebates and intercompany fees are handled, and which entity owns the resulting intellectual property, before the capability is embedded.
The economics: does company twelve make the portfolio richer?
"Twelve companies times one company's saving" assumes similar volume, a similar sustained improvement, and a real route to money. All three can be wrong. No reliable published dataset shows the cost of deploying the same AI workflow at successive, separately managed lower-middle-market companies.
| Effect | What it establishes | How to report it |
|---|
A sanity check. APQC's cross-industry median total cost is $6.00 per invoice, across 5,846 organizations. A company processing 15,000 invoices a year therefore spends roughly $90,000 on the whole accounts-payable process. A promised $100,000 of savings exceeds the entire benchmarked cost; even $45,000 would mean removing half of it.
An illustrative model. All numbers below are invented planning assumptions. A twelve-company portfolio has six suitable companies; three adopt. Shared build $25,000 once; local setup $12,000 per adopter; central upkeep $24,000 a year; local running $10,000 a year per adopter. Full-year financially convertible benefit: $75,000, $55,000 and $35,000. Year one realizes half the benefit.
| Economic measure | Year 1 | Year 2 | Three-year cumulative |
|---|---|---|---|
| Gross converted benefit | $82,500 | $165,000 | $412,500 |
| Shared and local recurring costs | ($54,000) | ($54,000) | ($162,000) |
| One-time shared and local setup | ($61,000) | $0 | ($61,000) |
| Net economic result | ($32,500) | +$111,000 | +$189,500 |
Cumulative benefit breaks even about 15.5 months after the start. The model is positive because three companies have large enough converted benefits, not because software is cheap.
If only 40% of the benefit materializes, full-year gross benefit is $66,000 against $54,000 of running cost; year one is ($82,000) and the three-year result is ($58,000). If local support doubles to $20,000 a company, running costs reach $84,000 and the program loses money every year before counting the original investment.
A fourth company with $5,000 of annual benefit and $10,000 of local running cost reduces annual net portfolio benefit by $5,000, plus $12,000 of setup, even though the platform already exists. The economically rational portfolio is not necessarily the one with the most adopters.
What would prove company twelve costs less. Record, for each deployment, discovery hours, local data work, configuration, testing, training, management time, elapsed time, initial error rate and recurring support, and compare like-for-like scope. Falling hours from a better package is evidence; falling hours from narrower scope or unrecorded central help is not.
The first 90 days, with six gates
A hypothetical twelve-company portfolio: five distributors, four manufacturers and three business-services companies, $20M to $150M of revenue each. One distributor has a proven improvement in handling customer orders. One operating leader, an implementation partner, limited local management capacity. The day ranges are planning windows; the gates are governed by evidence.
Stop, or rethink, when
The transfer document
The practical takeaway. Before moving an AI success to a second company, require one document that answers six questions, then run the second-company test on that company's own records, management team and financial baseline.
| Question | Required evidence |
|---|
Tempting claims to correct
Claims that circulate in decks and press coverage, with the defensible version.
| Tempting claim | Problem | Defensible treatment |
|---|
The research gaps that would most change the conclusions
Working with Local AI
This report argues for testing a proven workflow at a second company before anyone forecasts twelve. That is the work we do, inside the systems and permissions each company already has.
Where we would start. The transfer document for your first AI win: the original result, why it worked, and what will change. Then a bounded test at one deliberately chosen second company, on its own historical transactions, with a local owner, a spending cap and the CFO's baseline agreed up front.
How it runs. On the cloud each company already uses, with its own credentials, data and off switch. The shared package stays reusable; each company's data and decisions stay its own, so the capability survives a sale.
Live capabilities in 21 days
Discovery to live capabilities in 21 days, on the systems you already use. We start inside one real workflow and prove it against your own numbers.
Yours from day one
Production AI agents on the cloud you already use, with code, models, integrations and deployment patterns owned by you from day one. No vendor lock-in.
Built by operators
Our founder built machine learning at scale inside C.H. Robinson. A boutique senior team, not a generalist consultancy, not an SI, not a SaaS vendor.
Bring one real workflow. We will tell you, honestly, whether AI belongs in it yet, and if it does, how fast we can ship.
Sources
The principal external sources behind this report. A finding repeated across the research was counted once when it relied on the same original source.
| Source | Publisher | Date | Evidence type | Primary? |
|---|