How to Run a Lighthouse Portco AI Program: The 3-One Rule for One Quarter

TLDR: A lighthouse portco AI program proves AI value in one portfolio company, on one workflow, against one measure, and turns the result into a playbook the rest of the portfolio can copy. An operating partner with a partner meeting one quarter away needs one defensible number and a transfer test, and a portfolio-wide rollout will produce neither in time. This post gives the selection grid, the 12-week sequence, and the playbook contents.

Best For: AI operating partners and value-creation directors at mid-market PE funds whose portfolio companies are enterprise scale (1,000 to 15,000 employees), especially those new in the role who inherited a portfolio at mixed AI maturity and owe the partners an answer within a quarter.

A lighthouse portco AI program is a portfolio value-creation method that concentrates a fund's AI effort on a single portfolio company and a single workflow, measures the result against a baseline the company's own data team validated, and then packages what worked into a playbook for the other companies in the portfolio. The term borrows from the World Economic Forum's lighthouse factories: sites that prove a method at full scale before the rest of a network adopts it. The program exists because the alternative, several portfolio companies each buying AI tools on their own, produces activity without a number anyone can underwrite. For a newly appointed AI operating partner, the lighthouse is the fastest route to a result that survives the investment committee.

What a lighthouse portco AI program is, and what it is not

A lighthouse portco AI program is a deliberately narrow commitment: one company, one workflow, one measure, one quarter. It is not a portfolio-wide AI strategy, not a vendor bake-off, and not a maturity assessment. Those come later and borrow their credibility from the lighthouse result. The program's output is two things: a measured delta on one workflow, and a written playbook that says how to repeat it.

Where the lighthouse idea comes from

The World Economic Forum's Global Lighthouse Network has designated 201 production sites and value chains since 2018, and its September 2025 cohort reported an average 40% gain in labor productivity and a 44% reduction in cycle time at the designated sites, according to the WEF's 2025 announcement. The site itself was never the point. A method proven at one plant could be carried to the others, and that is what the designation certified. The same logic applies to a portfolio: one company proves the method, the playbook moves it.

The three things it is often confused with

The first confusion is with a pilot. A pilot tests whether a tool works. A lighthouse tests whether a method works and whether it transfers; the tool is incidental. The second confusion is with a center of excellence at the fund level, which builds capability before it has a result to point at. The third is with the AI maturity assessment many funds run across the portfolio. The assessment tells you where each company stands; it does not tell you what to do first. The lighthouse answers that question with evidence rather than a score.

Why funds are reaching for it now

The numbers behind the urgency are blunt. Grant Thornton's 2026 PE AI survey of 100 PE leaders found that 45% are still piloting AI, 11 points above the cross-industry rate, and only 5% have fully integrated AI into operations, against 14% across all industries. Alvarez & Marsal's May 2026 survey of 100 investors, operating partners and PE-backed executives found that 73% expect AI to increase portfolio value in the next 12 months while only 8% describe their own firm as leading in AI. That gap between expectation and capability is the operating partner's problem to close, and a quarter is not long enough to close it everywhere at once.

Why the 3-One Rule beats a portfolio-wide rollout

The 3-One Rule says a lighthouse portco AI program commits to one workflow, one measure, and one playbook, in that order, and refuses to add a second of any of them until the first has a validated number. Every added workflow doubles the baselines to build, every added measure invites the partners to argue about definitions, and every added playbook before the first transfer has been tested is a guess dressed as a method.

One workflow

The workflow has to be high-frequency and already measured, or measurable within two weeks. Invoice processing, claims intake, order entry, contract review, and customer service ticket triage all qualify at an enterprise-scale portco. Product strategy, pricing, and anything a senior manager does twice a month does not. McKinsey's 2025 State of AI survey found that only 23% of organizations have scaled AI agents in even one function, and that the roughly 6% of high performers were three times more likely than peers to have fundamentally redesigned a workflow before automating it. Pick the workflow you are willing to redesign.

One measure

The measure is the unit economics of that workflow: hours of human time per 100 completed units, the error or rework rate, and the cycle time from intake to completion, each captured before the AI touches anything and again at week 10. The formula for turning those inputs into a per-run comparison is covered in the post on how to measure AI ROI per workflow across a portfolio, so this post does not repeat it. What matters here is the discipline: the measure is chosen before the vendor, and the baseline is built by the portco's own data team, not entered by the vendor.

One playbook

The playbook is a document the next portco's COO could follow without calling you. It records the workflow boundary, the baseline method, the control design, the exception path, the measure, and the result, plus the conditions under which the result held. A playbook written before the transfer test is a hypothesis. A playbook written after one successful transfer is a method.

Lighthouse portco AI program vs. portfolio-wide AI rollout

A lighthouse portco AI program and a portfolio-wide AI rollout differ in where the risk sits. The rollout spreads a thin layer of AI spend across every company and asks each management team to find value; the lighthouse concentrates attention on one company until a number exists, then spreads the method rather than the spend. The rollout feels like progress at the partner meeting. The lighthouse produces something the partners can underwrite.

The rollout's failure mode shows up in every large survey. S&P Global Market Intelligence's 2025 survey of more than 1,000 companies found that 42% had abandoned most of their AI initiatives, up from 17% a year earlier, and that an average of 46% of proofs of concept were scrapped before production. MIT's 2025 GenAI Divide report put the share of pilots that never reach scaled adoption at 95%, and attributed the failure to a gap between the tools and the enterprise workflow rather than to the models. A rollout multiplies that gap by the number of companies in the portfolio.

How to choose the lighthouse: the Lighthouse Selection Grid

The lighthouse portco is chosen by scoring every candidate company and workflow pair on six inputs, each from 1 to 5, and picking the highest total rather than the most enthusiastic CEO. The grid takes an afternoon with the deal team's operating data and removes the most common failure, which is choosing the company that asked loudest instead of the one that can prove something.

The six inputs and their weights

A candidate scoring below 3 on baseline availability or transferability is disqualified regardless of total. The first disqualifier protects the number; the second protects the playbook. Baseline availability and transferability carry the heaviest weight because they are the two things a quarter cannot fix.

Why the loudest portco is usually the wrong one

The company already buying AI tools is often the one with the weakest baseline, because the spending started before the measure was defined. Before accepting that company's workflow as the lighthouse, run its existing initiative through the decision grid for evaluating an AI proposal from a portco. If the proposal cannot name a baseline, it is not a lighthouse candidate; it is a spend line to govern. Bain and StepStone's 2026 GP outlook, drawn from more than 100 investment professionals surveyed between December 2025 and January 2026, found that nearly 40% of GPs expect no material financial impact from AI in 2026. The companies behind that expectation are the ones that bought before they measured.

Scoring with operating data, not a workshop vote

The grid is filled from data the fund already holds: monthly volumes from the portco's systems, open projects from the hundred-day plans, and the data team's own account of which reports they trust. A workshop vote on the same six inputs produces a different answer, usually the one the room's most senior person preferred. Gartner's June 2025 forecast that more than 40% of agentic AI projects will be canceled by the end of 2027, mostly for unclear business value, describes what happens when the selection was a vote.

The 12-week sequence from selection to partner meeting

The lighthouse portco AI program runs in four blocks over twelve weeks: select in weeks 1 and 2, baseline in weeks 3 and 4, build and run in weeks 5 through 10, and measure and write the playbook in weeks 11 and 12. The sequence is rigid on purpose. The quickest way to lose the quarter is to let the build start before the baseline is signed off, because the week-12 number is then unverifiable.

Weeks 1 to 2: select and contract the measure

Score the grid, pick the company and workflow, and get the COO's written agreement on the measure and the baseline method. The written agreement sounds like a formality. It is not. PwC's 2026 Global CEO Survey of 4,454 CEOs found that 56% have seen no significant financial benefit from AI to date; the usual reason, when you ask, is that nobody agreed what benefit would be measured before the work started.

Weeks 3 to 4: baseline with the portco's own data team

The data team pulls the three measures for the previous 90 days and signs them. No vendor touches this step. If the data team cannot produce the baseline in two weeks, the grid score on baseline availability was wrong and the lighthouse should move to the next candidate rather than proceed on estimates.

Weeks 5 to 10: build, run, and keep a human on every exception

For a portfolio where two or three companies are already spending on AI with nothing to show and the partner meeting is a quarter away, a generic approach is not enough; the answer is a dedicated lighthouse program that does three things: it redesigns the workflow's intake and exception path before placing any AI in it, it runs the AI alongside the existing process for at least four weeks so the delta is measured against live volume, and it logs every exception with the reason it needed a person. MIT's 2025 research found tools bought from vendors succeeded about two-thirds of the time against one-third for internally built systems, which argues for buying the component and owning the workflow design, not the reverse.

Weeks 11 to 12: measure, write the playbook, and run the transfer test

Pull the same three measures for the last four weeks of the run, have the data team sign them, and write the delta into the playbook. Then run the transfer test: hand the playbook to the COO of the highest-scoring second portco and ask whether they could run it without a meeting. If the answer is no, the playbook is missing something, and the partner meeting should hear that honestly. BCG's September 2026 study of 1,330 senior leaders found the 7.5% of companies it calls future-built posted 2.8 times the EBITDA growth of laggards; the step that separates them is the one most programs skip, which is proving the result transfers.

What the partner meeting needs to see

The partner meeting needs a one-page lighthouse result with five lines: the workflow and its boundary, the baseline and who signed it, the delta on each of the three measures, the exception rate and what the exceptions were, and the transfer test outcome for the next two portcos. Anything beyond that page is supporting material. Partners who distrust strategy theater will read the five lines and ask one question: what would it take to run this at the next company?

The five lines, and what each one defends against

The workflow boundary defends against the objection that the result is cherry-picked. The signed baseline defends against the objection that the vendor entered the before-number. The three-measure delta defends against a single flattering metric. The exception rate defends against the belief that AI handled everything, which no partner believes anyway. The transfer test defends against the objection that this was a one-off. L.E.K.'s 2026 PE Pulse, surveying 100 US buyout professionals in April 2026, found respondents reporting an average 28% productivity improvement from their own AI use; the same partners will expect a comparable number from a portco, measured the same way.

What the page leaves out on purpose

The page does not report tool adoption, licenses assigned, or prompts written. Those are the usage metrics that hide a decline in efficiency, and the post on governing AI spend across a portfolio explains why they mislead. KPMG's Q4 2025 pulse survey of more than 100 asset management and PE executives found 68% piloting AI agents and 40% expecting measurable return within 12 months. The lighthouse page is how an operating partner shows a measured return rather than an expected one.

Common objections from the partner room, and what to say

The objections an operating partner hears about a lighthouse portco AI program are predictable, and each has a direct answer. The three below come up in nearly every partner discussion, usually in this order: the first questions whether the method is worth paying for, the second questions whether it is new, and the third questions whether one company is enough. Each deserves a straight reply rather than a slide.

"Couldn't an AI assistant build this grid in two seconds?"

Yes, the grid template. Not the scores. The scores come from the portco's monthly volumes, the hundred-day plan, and the data team's judgment about which reports are trustworthy, and none of that is in an assistant's context. The method is kept visible so that the partners can check the inputs; the work is in filling them honestly, and in getting a COO to sign a measure before a build starts.

"This is just RPA with a new name"

Some of it is, and that is fine. If the lighthouse workflow turns out to be rules-based end to end, the playbook should say so and the cheaper tool should win. The lighthouse differs in the exception path: the steps where documents are unstructured, where the right action depends on reading context, and where a person has to stay in the loop. The exception log from weeks 5 to 10 is what shows which parts needed judgment and which did not.

"Why only one company when the whole portfolio needs this?"

Because the whole portfolio needs a method, not a tool, and a method has to be proven somewhere first. Grant Thornton's 2026 survey found only 9% of PE leaders were very confident they could pass an AI governance audit within 90 days, the lowest of any sector. A fund that cannot yet govern one AI workflow should not be deploying twelve. The post on why portfolio company AI pilots fail covers what the untested operating model does to the second and third company.

How the playbook moves from one portco to the portfolio

The lighthouse playbook moves to the next portfolio company through the transfer test, then a second run under the same measure, and only then a portfolio-wide decision on which companies qualify. The sequencing rule is that no company enters the program until the previous transfer has produced a signed delta. Two signed deltas on the same workflow are the point at which the partners can underwrite the method for the portfolio.

The transfer test, defined

The transfer test asks three questions of the candidate company: does it run a workflow that matches the playbook's boundary, can its data team produce the same three-measure baseline within two weeks, and does its COO accept the measure in writing. Three yeses means the company runs the playbook as written. One no means the playbook is adapted and the adaptation is recorded. Two or more means the company is not next.

What changes at the second company

The second run is usually faster but rarely identical. Volumes differ, the system of record differs, and the exception categories shift. The playbook's second edition records those differences, which is what turns it from a case study into a method. The PEI operating partner survey of 2026, fielded with 145 value-creation professionals in April and May 2026, exists because funds are asking how to turn portfolio-company AI activity into something repeatable. The second-edition playbook is the answer.

When to stop

A lighthouse that fails the transfer test twice, with two different candidate companies, is telling you the workflow is specific to the first portco. That is a useful result. Record it, keep the first company's gain, and run the selection grid again with transferability weighted higher. The fund wanted a reusable method, and a clean negative result on transfer is worth more than a flattering number that cannot leave the building.

This analysis was developed using methodologies and operating experience from Assembly.

Frequently Asked Questions

What is a lighthouse portco AI program?

A lighthouse portco AI program is a portfolio value-creation method that concentrates a fund's AI effort on one portco and one workflow, measures the result against a baseline the company's data team validated, and packages what worked into a playbook for the other portcos. The name comes from the WEF's lighthouse factories, which prove a method at one site first.

What is the 3-One Rule for a lighthouse program?

The 3-One Rule commits a lighthouse program to one workflow, one measure, and one playbook, in that order, and refuses to add a second of any until the first has a validated number. Every added workflow doubles the baselines to build, every added measure invites argument over definitions, and a playbook written before a transfer test is only a guess.

How do you choose which portfolio company becomes the lighthouse?

The lighthouse portco is chosen by scoring each candidate company and workflow on six inputs from 1 to 5: baseline availability, volume and frequency, process stability, data and system access, sponsorship, and transferability. Baseline availability and transferability carry triple weight, and any candidate scoring below 3 on either is disqualified regardless of total, because a quarter cannot fix those two.

Which workflow should a lighthouse AI program start with?

A lighthouse AI program should start with a high-frequency workflow that is already measured or measurable within two weeks: invoice processing, claims intake, order entry, contract review, or ticket triage at an enterprise-scale portco. McKinsey's 2025 survey found high performers three times more likely to redesign a workflow before automating it, so choose one you are willing to redesign.

How is a lighthouse portco AI program measured?

A lighthouse program is measured on three unit-economics inputs captured before and after: hours of human time per 100 completed units, the error or rework rate, and the cycle time from intake to completion. The portco's data team builds the baseline from the previous 90 days and signs both readings. A vendor-entered baseline disqualifies the result.

How long does a lighthouse portco AI program take?

A lighthouse portco AI program takes twelve weeks in four blocks: selection in weeks 1 and 2, baseline in weeks 3 and 4, build and run in weeks 5 through 10, and measurement plus playbook in weeks 11 and 12. The sequence is rigid because a build that starts before the baseline is signed makes the week-12 number unverifiable.

What is the difference between a lighthouse program and a portfolio-wide AI rollout?

A lighthouse program concentrates attention on one company until a validated number exists, then spreads the method; a rollout spreads spend across every company and hopes each finds value. S&P Global's 2025 survey found 42% of companies abandoned most AI initiatives and scrapped 46% of proofs of concept, which is the rollout's failure mode multiplied by the portfolio.

What is the difference between a lighthouse and an AI pilot?

A pilot tests whether a tool works; a lighthouse tests whether a method works and whether it transfers to another company. A pilot ends with a demo and a decision about the tool. A lighthouse ends with a signed delta on three measures, an exception log, a written playbook, and a transfer test result for the next two portcos.

Who should own a lighthouse portco AI program?

The operating partner and the lighthouse portco's COO jointly own the program, with the COO owning the workflow and the written measure and the operating partner owning the selection grid, the sequence, and the playbook. The portco's data team owns the baseline and the final readings. Vendors supply components and never own the number or the workflow design.

What goes in the lighthouse playbook?

The lighthouse playbook records the workflow boundary, the baseline method, the control design, the exception path, the three measures, the signed result, and the conditions under which the result held. The test of a finished playbook is whether the next portco's COO could run it without calling the operating partner. Before a successful transfer, it is only a hypothesis.

What is the transfer test in a lighthouse program?

The transfer test asks three questions of the next candidate company: does it run a workflow matching the playbook's boundary, can its data team produce the same baseline within two weeks, and does its COO accept the measure in writing. Three yeses means the playbook runs as written; one no means a recorded adaptation; two means the company waits.

What should the partner meeting see from a lighthouse program?

The partner meeting should see a single page with five lines: the workflow and its boundary, the baseline and who signed it, the delta on each measure, the exception rate and causes, and the transfer test outcome for the next two portcos. Tool adoption and license counts are left off because they hide efficiency declines behind usage growth.

Why do lighthouse portco AI programs fail?

Lighthouse programs fail most often because the wrong company was chosen, usually the one already spending on AI with the weakest baseline, or because the build started before the baseline was signed. Gartner's June 2025 forecast that over 40% of agentic AI projects will be canceled by 2027 for unclear business value describes what a selection by workshop vote produces.

Is a lighthouse AI program just RPA under another name?

Parts of a lighthouse workflow may be rules-based, and the playbook should say so and let the cheaper tool win. The lighthouse differs in the exception path, where documents are unstructured, the right action depends on context, and a person stays in the loop. The exception log from weeks 5 to 10 shows which steps needed judgment.

How does a lighthouse program scale across a portfolio?

A lighthouse program scales through the transfer test, a second run under the same measure, and only then a portfolio-wide qualification decision. No company enters until the previous transfer has produced a signed delta. BCG's September 2026 study found the 7.5% of future-built companies posted 2.8 times the EBITDA growth of laggards; proving transfer is the step that separates them.

When should an operating partner stop a lighthouse program?

An operating partner should stop a lighthouse program after it fails the transfer test twice with two different candidate companies, because that result says the workflow is specific to the first portco. Record the negative result, keep the first company's gain, and rerun the selection grid with transferability weighted higher. A clean negative on transfer is a useful result.