# Portfolio AI Ops — Go-To-Market **Product:** Portfolio AI Ops **Buyer:** PE Operating Partner **Positioning:** Proof, not pilots. **Operator:** Mike Rodgers, solo forward-deployed engineer. --- ## 1. ICP definition ### Firmographics | Dimension | Target | |---|---| | AUM | $200M – $3B | | Fund stage | Fund II through Fund IV (has a track record, still building the platform) | | Portfolio count | 8 – 40 active companies | | Company size | $5M – $40M EBITDA, services-heavy | | Sectors | Healthcare services, business services, professional services (CPA/law/insurance), distribution, industrial services, logistics | | Geography | US, with a bias to non-coastal funds who are under-served by MBB and have no in-house technology bench | | Hold model | Buy-and-build / platform-and-add-on. Roll-ups have repeatable back offices, which is where AI actually lands | **Why services-heavy roll-ups specifically:** labour is 45–65% of COGS, the same process runs in 9 to 40 locations, and a single automation replicates across every site. In a capital-intensive or IP-heavy business the AI story is bespoke per company and the economics do not repeat. ### Buying committee | Role | Function in the deal | What they need to hear | |---|---|---| | **Operating Partner** (economic buyer) | Owns value creation, has discretionary budget in the $50K–$500K band without IC approval | "You will have an answer when the IC asks what the AI spend returned" | | **VP / Director of Value Creation** (champion) | Runs the actual initiative, is the one currently embarrassed by the lack of data | "This makes your quarterly reporting defensible" | | **CFO of the fund** (approver above ~$150K) | Cares about it being an opex line with a clear term, not a capitalised project | Fixed fee, fixed term, no seat sprawl | | **Portfolio company CEO** (blocker, not buyer) | Can kill it by refusing data access | "The forecast is co-signed by your CFO, and it measures the programme, not you" | | **Fund CTO / Head of Technology** (blocker where one exists) | Territorial about tooling and vendor selection | "I do the underwriting and deployment, your team keeps data ownership" | **Budget reality:** most lower-middle-market funds have no line item called "AI." The money comes out of the value creation budget, an existing consulting allocation, or is pushed down to the portfolio company P&L as a management fee add-back. The $18K diagnostic is deliberately priced below the level that triggers a committee. ### Disqualifiers — do not sell to these 1. **Fewer than 6 portfolio companies.** The whole thesis is that a forecast ledger produces a comparable track record. With 4 companies and 6 forecasts, n is too small for calibration to mean anything, and you would be selling a statistic that cannot be computed. Say so and walk. 2. **Funds with an existing internal AI/data platform team of 3+.** They will treat this as a turf incursion, and they are usually right to. The exception is when they explicitly want an independent scorekeeper — that is a different, smaller sale. 3. **Sub-$5M EBITDA portfolio companies.** Deployment cost does not amortise. A $50K deployment against $4M EBITDA is a 1.25% hit for an uncertain return, and the CFO will correctly refuse. 4. **Funds in active fundraise with a first close inside 90 days.** All attention is on the LP deck, nobody will authorise data access, and the initiative goes dormant for a quarter. Note the date and come back after close — the fundraise itself becomes the trigger event later. 5. **Distressed / turnaround mandates.** The operating partner is fighting a liquidity fire. A 6-month measurement cadence is irrelevant when the question is whether payroll clears. Wrong time, not wrong buyer. --- ## 2. Trigger events Seven observable signals that a fund is buying now, and exactly where to observe each one. | # | Trigger | Where to observe it | Why it means "now" | |---|---|---|---| | 1 | Fund posts a role titled **Operating Partner – Technology**, **VP Value Creation**, or **Head of Portfolio Operations** with "AI" in the body | LinkedIn Jobs, filtered to the firm; also the fund's own careers page | They have decided the capability is missing and have budget. The role takes 6–9 months to fill. The gap is the window | | 2 | **New fund close announced** | SEC Form D filings (EDGAR, full-text search on "Private Equity"), plus PE Hub / Axios Pro Rata | Fresh dry powder and a new value creation thesis to differentiate on. Budget is least constrained in the 6 months post-close | | 3 | **LP annual meeting scheduled or just held** | Fund's LinkedIn/news page, and conference calendars. Most run Q1 or Q3 | The "what is your AI strategy" question just got asked by an LP and was answered badly. Highest-intent moment in the whole cycle | | 4 | **Platform acquisition in a services roll-up** | PitchBook / Axial / GrowthCap deal announcements; press-release language containing "buy-and-build", "platform investment", "add-on strategy" | 100-day plan is being written right now, and back-office consolidation is always on it. The forecast slate writes itself | | 5 | Operating partner **posts publicly about AI** on LinkedIn, especially sceptically | LinkedIn, follow the 200 target operating partners directly | Sceptics are the best buyers for a measurement product. They have already articulated the problem you solve | | 6 | Fund publishes an **AI policy, AI playbook, or "AI in the portfolio" thought piece** | Fund website /insights, LinkedIn newsletter | They have committed publicly. The next question they get is "show the results", and they cannot | | 7 | **Portfolio company CFO or COO turnover** in a services platform | LinkedIn job-change alerts on the portfolio company | New finance leadership re-baselines everything and is unusually willing to instrument. Sell through the company, invoice the fund | **Highest-conviction combination:** trigger 1 + trigger 3 within the same quarter. That is a fund that just got asked the question, has no one in seat to answer it, and has authorised headcount. Lead with that. --- ## 3. Cold email sequence Five emails over 24 days. Under 120 words each. Plain text, no images, no tracking pixels, no calendar link before email three. --- ### Email 1 — Day 1 · trigger-specific observation **Subject:** the VP Value Creation role > Saw you're hiring a VP of Value Creation with AI in the scope. That role usually takes seven months to fill. > > The thing that tends to get lost in the gap is the record. Initiatives start, quarters pass, and by the time the person lands nobody can say which of them worked — not because they failed, but because no one wrote down what was supposed to happen before it happened. > > I do one narrow thing: every AI initiative in a portfolio gets filed as a dated forecast with a probability, then scored when it resolves. > > Worth ten minutes while the seat is open? > > Mike **CTA:** ten minutes, no calendar link yet. --- ### Email 2 — Day 4 · the measurement gap **Subject:** the question after the AI question > Follow-up, then I'll leave it. > > The AI question at an LP meeting is easy. It's the second question that's hard: *what did it return?* > > Most funds answer with an activity number — documents processed, seats deployed, hours saved by an estimate nobody audited. It doesn't survive a follow-up from anyone numerate, and the operating partner knows it while they're saying it. > > The fix isn't better dashboards. It's writing the claim down first, with a date and a probability, so the outcome can actually settle against something. > > Have you got a number you'd be comfortable defending? > > Mike **CTA:** a direct question, answerable in one line. --- ### Email 3 — Day 9 · a concrete number **Subject:** 0.323 → 0.131 > Those are Brier scores from a portfolio AI programme, first cohort to third. > > A Brier score is mean squared error on probabilities. 0.25 is what you get for saying fifty-fifty about everything. So the first cohort of forecasts was *worse than a coin flip* — badly overconfident, and it's on the record because that's the point. > > By the third cohort it was 0.131. Same operator, same portfolio. The only variable was that the earlier forecasts got scored. > > That's the entire product: the scoring is what makes the next underwrite better. > > Live ledger here if you want to click through it: [link] > > Mike **CTA:** the working app. Let the product do the talking. --- ### Email 4 — Day 15 · contrarian, risks disagreement **Subject:** most portfolio AI pilots should be killed > A view you may disagree with. > > The problem with portfolio AI isn't that too few pilots are running. It's that too few are ever killed. A pilot with no pre-registered claim can't fail — every outcome gets narrated into a success, so it renews, and the budget calcifies around initiatives nobody has the evidence to stop. > > Pre-registering the claim is uncomfortable for exactly that reason. It creates the possibility of a NO. In the ledger I run, roughly a third resolve NO and stay visible. > > If your AI programme has never killed anything, that's the finding. > > How many did you stop last year? > > Mike **CTA:** a question that is genuinely uncomfortable to answer. --- ### Email 5 — Day 24 · breakup **Subject:** closing this out > I'll stop here. > > If the AI programme is measured and the numbers hold up, you don't need me and I'd be glad to hear it. > > If it isn't, the useful moment is usually about six weeks before the LP meeting — early enough to have something real to say, late enough that the quarter's results are in. > > I'll leave you with the one question worth asking any AI vendor pitching your portfolio: *what was your Brier score on your last ten engagements?* None of them will have one. That answer is free and it will save you a quarter. > > Good luck with the fund. > > Mike **CTA:** none. Give a genuinely useful thing away and leave. --- ## 4. LinkedIn posts ### Post 1 — the failure story > The first cohort of AI forecasts I ran across a PE portfolio scored worse than a coin flip. > > Eight forecasts. Things like "this AP automation cuts cost per invoice from $6.40 to under $3.50 by June 30." Each one had a probability attached. Brier score came back at 0.323. > > For context: if you say "fifty-fifty" about everything you score 0.25. So I paid $500K of deployment cost to be more confident and less accurate than a shrug. > > The specific failure was interesting. My 85% and 90% forecasts were the ones that broke. Anything I was sure about, I was sure about for reasons that came from the demo rather than from the operating data. The 55% and 60% forecasts held up fine. > > Second cohort: 0.196. Third: 0.131. > > Nothing changed about the technology between those cohorts. What changed is that the first cohort got scored, in public, in front of the people who had approved it. > > Most portfolio AI programmes never get to that first embarrassing number. Not because they're doing better — because nobody wrote the claim down before the work started, so there's nothing for reality to contradict. > > The embarrassing number is the asset. It's the only thing that makes the next one better. --- ### Post 2 — the framework > There are three numbers hiding inside every forecast score, and PE operating partners should know all three. > > Any Brier score splits, exactly, into: reliability minus resolution plus uncertainty. > > **Reliability** is calibration error. When you say 70%, does it happen 70% of the time? Lower is better. This is the one everybody assumes is the whole story. > > **Resolution** is information content — how far your forecasts move away from the base rate and are right to. Higher is better. A forecaster who says "60%" about everything can be perfectly reliable and completely useless. Perfect calibration, zero resolution, zero value. > > **Uncertainty** is fixed by the base rate of the questions themselves. You cannot improve it, and you should not get credit or blame for it. It's the reason a fund working on hard problems will always look worse than one cherry-picking easy ones. > > Why this matters commercially: when someone shows you an AI track record, ask which of the three they improved. Most "improvement" is a shift to easier questions, which shows up as lower uncertainty and flat resolution. > > That's not a better team. That's a smaller ambition, reported as a win. --- ### Post 3 — the contrarian argument > Unpopular position: the AI pilot is the problem, not the solution. > > The pilot exists to reduce risk. In practice it does the opposite, because a pilot is structurally incapable of failing. > > It has no pre-registered success criterion. It runs in a carve-out where the data is clean and the users are volunteers. It reports activity — documents processed, hours saved by estimate — instead of P&L. And at the end, whatever happened gets narrated forward into a business case. > > I've watched portfolios run eleven simultaneous pilots and be unable to name the three that worked. Not from incompetence. From the absence of a written claim, made in advance, that reality could have contradicted. > > The alternative isn't more governance. It's one sentence, signed before the work starts: > > *"We believe there is a 70% chance this cuts cost per invoice below $3.50 by June 30."* > > Now it can fail. Now the 70% can be scored. Now the next estimate is informed by something other than the vendor's enthusiasm. > > Pilots protect the people who authorised them. Forecasts don't. That's the whole difference, and it's why the second one is a harder sell. --- ## 5. Pricing and ROI math Full detail in `PRICING.md`. The version that goes in an email: **What one unmeasured pilot costs** | Line | Amount | |---|---| | Typical single-company AI pilot, 3–4 months | $150,000 – $250,000 | | Internal time to run it (0.3 FTE operating partner + 0.5 FTE company side) | ~$85,000 | | **Cost of one pilot** | **$235,000 – $335,000** | | Defensible EBITDA attribution produced | $0 | **The alternatives** | Option | Year-one cost | What you get | |---|---|---| | MBB / Big-4 digital diligence | $500K – $2M | A deck. No deployment, no measurement | | Director of Portfolio Technology (hire) | $280K loaded + 6–9 months to fill | Capability, eventually. Year one is orientation | | Three AI point vendors, 4 companies each | $180K – $400K/yr | Seats and usage dashboards. Attribution is your problem | | Portfolio AI Ops — Deployment | **$189K year one** | Systems in production in 4 companies + a scored ledger | | Do nothing | $0 | Defensible for one more cycle | **Payback, Deployment tier, deliberately conservative** ``` Year one cost $45,000 setup + (12 × $12,000) = $189,000 Realised value at 3.0x (demo ledger runs 7.3x; use 3.0x) = $567,000 Net EBITDA effect = $378,000 Enterprise value at a 9x exit multiple = $3,402,000 Payback period $189,000 / ($567,000 / 12) = 4.0 months ``` The number to say out loud: **at a 9x multiple, $378K of recurring EBITDA is $3.4M of enterprise value, for a $189K opex line.** That is an 18:1 return on the fee, and it assumes the programme performs at 40% of what the reference ledger shows. --- ## 6. Top three objections ### "We already have a data and analytics team." > Good, you'll need them for data access. But those are different jobs. Your analytics team reports what happened — they're a rear-view function and they're good at it. What I do is commit to what *will* happen, in advance, with a probability attached, and then get scored on whether I was right. > > Almost no analytics team is asked to do that, because it's an uncomfortable thing to ask of someone whose performance review you own. It's much easier to have an outsider carry the forecast risk. > > And genuinely: if your team already files dated probabilistic forecasts and computes calibration on them, you don't need me. I've never met one that does. If yours does, I'd like to see it. --- ### "We tried AI consultants and got a deck." > I believe you, that's the default outcome. Two structural things are different here, and they're both about who carries the risk. > > First, the deployment cost gets booked against the forecast whether or not it lands. If I'm optimistic and wrong, that shows up in my number, not in a footnote. A consultant's engagement ends at the recommendation, which is exactly why the recommendation is always confident. > > Second, there's a score at the end that I can't spin. It's arithmetic on a public record and it follows me into the next quarter and the quarter after. > > Here's a test that costs you nothing. Go back to the firm that gave you the deck and ask what their Brier score was across their last ten engagements. They won't have one. Not because they're bad at their jobs — because the business model doesn't require them to be scored, and mine does. --- ### "$18K for three weeks is a lot for a diagnostic." > It is, and I'd push back on the comparison. The alternative you're implicitly pricing it against is a free assessment, and free assessments have a specific failure mode: they get treated exactly as seriously as they cost. The CFO doesn't clear time, the data access request sits for three weeks, and you get a document nobody acts on. > > The real comparison is the $150K–$250K pilot you'd otherwise run to find the same thing out, or the $500K MBB engagement that ends in a deck. > > Two things to make it easy. It credits in full against the Deployment setup fee if you continue, so if we work together it's effectively free. And if it doesn't surface at least $500K of measurable annualised upside across the three companies, you don't pay for it. I'll have spent three weeks and told you not to buy the thing, which is a finding worth having. --- ## 7. First ten prospect segments Ranked by expected conversion. "Where to find them" names a real source. | # | Segment | Archetype | Why they buy | Where to build the list | |---|---|---|---|---| | 1 | **Healthcare services roll-ups** (dental, vet, derm, ophthalmology, behavioural health) | $400M–$1.5B fund, 15–35 clinic platforms | Identical back office in 40 sites. One automation replicates everywhere. Mike already has dental domain depth and reference language | PitchBook sector screen on Healthcare Services + "platform"; DSO-specific: Group Dentistry Now deal tracker; ADSO member list | | 2 | **CPA and accounting firm consolidators** | Newer vehicles (2021+) rolling up regional CPA firms | Q1 capacity is a hard physical constraint. Preparer leverage is *the* KPI, and it is directly forecastable | Accounting Today "Top 100 Firms" cross-referenced with PE ownership; INSIDE Public Accounting M&A tracker | | 3 | **Funds that just hired a VP Value Creation** | Any size, actively building the ops function | Trigger 1. The seat is open and the mandate is fresh | LinkedIn Jobs saved search on the title, filtered to PE firms; Sales Navigator alert on the 200 target funds | | 4 | **Lower-middle-market industrial and field services** | $200M–$800M fund, route-based or dispatch-based businesses | Labour is 55–65% of COGS and scheduling density is directly measurable. Cleanest forecast slate in the whole list | Axial deal-flow network; ACG chapter member directories (Chicago, Dallas, Atlanta) | | 5 | **Funds 6 months post-close on a new vehicle** | Fund III/IV, dry powder, differentiation pressure | Trigger 2. Budget least constrained, thesis still being written | SEC EDGAR full-text search on Form D, filtered by state and fund size; PE Hub weekly close roundup | | 6 | **Insurance brokerage and TPA roll-ups** | Heavy consolidation, document-intensive | Submissions, certs, and claims intake are enormous, boring, and highly automatable | Insurance Journal M&A reports; Reagan Consulting deal tracker | | 7 | **Specialty distribution platforms** | $150M–$400M revenue, 100K+ SKUs | Quote-to-order cycle time and rebate leakage are both large and directly countable | Modern Distribution Management (MDM) top-40 lists cross-referenced with PE ownership | | 8 | **Independent sponsors and family offices with 6+ holdings** | No institutional ops bench at all | Nobody in seat, so no internal turf. Fastest sales cycle in the list; smaller cheque | Independent Sponsor Summit attendee lists; McGuireWoods independent sponsor directory | | 9 | **Funds with a public AI position** | Published a playbook, newsletter, or LP letter section on AI | Trigger 6. Publicly committed, now needs results to show | Fund websites /insights; LinkedIn newsletter search; Google `site:` search on fund domains for "artificial intelligence" | | 10 | **Legal services and ALSP roll-ups** | eDiscovery, IP prosecution, immigration, claims | Highest per-hour labour cost of any segment here, and the review work is well-suited to automation | Legaltech News; Thomson Reuters ALSP report; ILTA member directory cross-referenced with PE ownership | **Sequencing recommendation:** work segments 1, 2, and 4 first. Mike has genuine domain language in dental and CPA (existing business lines), which lets email 1 reference a specific operating metric rather than a generic AI observation. Segment 3 is the highest-intent list but the smallest; run it as an always-on alert rather than a campaign. **List build, week one:** Apollo for the operating partner contact data at named funds (key already live in Keychain), LinkedIn Sales Navigator for the trigger alerts, and manual enrichment on the top 40. Do not buy a generic "PE contacts" list. The whole sequence depends on email 1 naming something real and specific about that fund, and a bought list cannot support that.