Principal AI Operator

Principal AI Operator

Plan of Attack

How I would approach the role: get the portfolio moving on internal efficiency, software-production speed, and AI-enabled product growth, so more companies reach maximum valuation at exit.

The mandate, in PSG's terms

This role is framed as an office-of-the-CEO capability under the AI-transformation umbrella. The job is to get roughly 130 current portfolio companies moving to the right on three vectors:

1

Internal operating efficiency

AI across GTM, finance, customer success, and operations.

2

Speed of software production

Engineering maturity, moving companies up the L0 to L5 curve.

3

AI-enabled product & revenue

Move products from dashboards to prescriptive, action-oriented AI (recommendations, automated workflows) that generates net-new revenue.

The scoreboard is the exit, driven by two engines: cost-out efficiency that lifts margin, and AI-native product that generates net-new revenue and retention. The product engine is the higher-multiple prize, and buyers now score the percentage of revenue directly attributable to AI. My job is to move as many companies as possible into that category, and to keep doing it as PSG acquires new (behind) companies.

Operating thesis

PSG already has the thesis and the network: the "rising tide lifts all boats" philosophy, the CEO forums, the twice-yearly survey. What it lacks is someone who goes into a company, gets a skeptical CEO and team to actually move, ships production AI, and then codifies what worked so the next company skips the learning curve.

I run this as a repeatable program, not a set of heroics:

Diagnose→Prove→Codify→Disseminate

Diligence: AI defensibility vs. disruption

Half the role sits before and around the deal: advising deal teams on where AI is a moat versus a threat, on both prospective and current investments. The market moves in "fruit fly years," so the underwriting question has shifted from "is it growing?" to "does this survive the next 24 months of model progress?"

What defends

Proprietary data and workflow depth, distribution and switching costs, and AI that takes action inside the system of record, not a thin wrapper any model commoditizes.

What gets disrupted

Features a frontier model or coding agent absorbs, screenshot-and-rebuild UIs, and businesses whose value was the manual work AI now automates.

I would run a fast product-and-tech read early (a 60-to-90-minute pass, in the spirit of how PSG already diligences), score AI defensibility, and flag the disruption risks worth a deeper look before an LOI. Post-close, the same lens becomes the value-creation starting point.

How I would sequence the portfolio

The lens: prioritize by capital at risk and plan status. I would build a simple two-by-two and sequence against it.

Lean in first

High capital at risk AND off-plan or at competitive risk. The swing factors for the fund, and where AI can change the outcome.

Fast wins next

High capital at risk, on-plan, but low AI maturity. Quick efficiency and margin gains for proof, paired with an AI-native product bet for the bigger multiple.

Self-serve / playbook

Smaller checks or already AI-forward. Give them the playbook and light-touch support.

Overlaid on the portfolio map (see the Portfolio tab): roughly 130 current companies, heavily Cloud Business Apps, then Big Data & Infrastructure and Security. The survey signal is the opening: product-side, ~70 percent are in decent shape; internal-side, only ~15 percent have built anything real and ~70 percent have done nothing. Internal efficiency is the fastest, most winnable proof point, straight to burn and EBITDA, but the higher-multiple prize is AI-native product and net-new revenue. I sequence both from day one.

First 90 days

0-30

Diagnose and earn trust

Read the commitment schedule and the two most recent portfolio surveys. Build the prioritization map with the deal teams. Meet the two ex-CTOs and the GTM and finance operating leads so I complement rather than collide with them. Pick three to five lighthouse companies from the "lean in first" quadrant. Run a one-page AI diagnostic per company: where AI defends the moat, cuts cost, lifts the product, and how ready the CEO and team actually are.

30-60

Prove it on lighthouses

With each lighthouse CEO, pick two or three high-leverage use cases: one internal-efficiency win for fast proof in burn, and one AI-native product or revenue bet for the bigger multiple. Stand up a production pilot in weeks, not quarters. Insist on leadership-by-example: the CEO and C-suite personally using the tools, because the portfolio data shows that is the single biggest predictor of whether a company moves. Start drafting the playbook from what actually works.

60-90

Codify and set the rhythm

Turn lighthouse wins into v1 of a reusable playbook: diagnostic template, use-case library, vendor-selection guidance, hire-vs-promote and team-assessment frameworks, and a governance and evaluation baseline. Define the metrics I will report and the cadence. Bring first results to a CEO forum so the network sees peers winning, not a mandate pushed down.

The repeatable playbook

The unit of work per company:

  1. Diagnose (1 page): moat defensibility, cost-out opportunities, product lift, and an honest team-readiness score.
  2. Select 2 to 3 use cases with clear owners and a measurable target.
  3. Ship a production pilot fast, with a human-in-the-loop verification step from day one.
  4. Prove the number: margin and cycle time on the efficiency side; net-new revenue, retention/NRR, and % of revenue attributable to AI on the product side. Get the CEO using it personally.
  5. Codify into the shared library, including what failed.
  6. Disseminate across the portfolio through forums, webinars, hackathons, and the playbook (the Blitzy-webinar model works). The voice-AI customer-service cohort is a ready-made first cross-pollination play.

The product through-line is moving each function and product from analytics and dashboards to prescriptive, action-oriented AI: recommendations and automated workflows that take the action, not just surface the insight. Engineering has its own maturity ladder (see the Metrics tab); everywhere else, the same idea applies function by function.

Company AI maturity: scoring the business with the CEO

The heart of the role is sitting with a CEO and honestly scoring the whole company, and each function, on how well it actually uses AI, then setting the next-rung target together. The same L0 to L5 spine applies to any function.

L0
L0 · AbsentNo meaningful AI use; the function runs on manual workflows.
L1
L1 · Ad hocIndividuals use chatbots on their own to draft and summarize. No standard, no process change.
L2
L2 · EmbeddedAI is built into the function's core tools and used daily. Real time savings, still fully human-run.
L3
L3 · SystematizedConnected data and agents run recurring workflows (forecasts, board updates, outreach, ticket triage) with a human in the loop. The operating cadence changes.
L4
L4 · Agent-runAgents own core processes end to end. People set goals, handle exceptions, and govern. Leverage and headcount shift.
L5
L5 · ReinventedThe function is redesigned around AI. A fraction of the former team delivers more, with leadership using it by example.

What I would score

Rate the business overall and each function against the ladder:

Executive / Office of the CEOProduct & EngineeringSales / GTMMarketingCustomer SuccessFinance / Office of the CFOOperations & People
Overlay dimension: leadership by example. Whether the CEO and C-suite personally use AI is the single biggest predictor of whether a company moves, so I score it on its own.
The through-line: moving a function up the ladder shows up as a specific operating metric (efficiency and margin, ARR per FTE, net-new AI revenue, NRR, Rule of 40), and those improved metrics are exactly what earns a higher multiple at exit. Maturity is the input, the metric is the gauge, the exit is the payoff.

Engineering drill-down: the coding maturity ladder

Product & Engineering gets its own deeper rubric. PSG runs the agentic-coding maturity model created by Dan Shapiro ("The Five Levels: from Spicy Autocomplete to the Dark Factory"), with Steve Yegge's eight-stage progression as the engineer's-eye view of the same curve. The levels below follow Dan's definitions. The real insight is that your role changes at each level: coder, then pair, then manager, then product manager, then nothing. Most developers cluster at L2 and most teams top out at L3, so the AI Operator earns their keep by breaking the L2-to-L3 and L3-to-L4 ceilings across the portfolio.

L0
L0 · ManualNothing hits disk without your approval. AI is a search engine on steroids or an occasional tab-complete. The code is unmistakably yours.
L1
L1 · Delegated tasksYou still write the important code but hand off discrete jobs ("write this unit test," "add a docstring"). Copilot or copy-paste chat. You still move at the speed you type.
L2
L2 · AI-native pairingYou pair with the AI like a colleague in an AI-native tool, in a flow state and more productive than ever. About 90% of AI-forward developers plateau here: it feels done, but it is not.
L3
L3 · Human in the loopYou are no longer the senior developer; the AI is. You manage it, run multiple agents, and review a constant stream of diffs. Most teams top out here.
L4
L4 · Spec-driven (a PM)You write and argue over specs, craft reusable skills, plan and review, then let agents run for hours and come back to check whether the tests pass.
L5
L5 · Dark factoryA black box that turns specs into working, tested software, with humans neither needed nor welcome. Only small, elite teams operate here today.

How I would measure success

Two layers. First, the AI-transformation scoreboard I would roll up for the firm. Second, the operating KPIs I would hold each portfolio company to: the health dashboard I run, fed by the week-4 / week-10 board cadence.

AI-transformation scoreboard

Company operating KPIs (the health dashboard)

The standard growth-software metric set, with the healthy ranges I manage toward. AI should move the efficiency and productivity lines (Rule of 40, ARR per FTE, burn multiple, CAC payback) the fastest.

MetricWhat it measuresHealthy range (2026)
Growth & retention
ARR growthYear-over-year growth in annual recurring revenue>40% early stage; >30% at scale
NRRNet revenue retention: expansion minus churn and contraction on the existing base110%+ good, 120%+ excellent
GRRGross revenue retention (churn only, no expansion credit)90%+ (85%+ acceptable)
ChurnLogo or revenue lost per year<5-7%; lower for enterprise
Efficiency
CAC paybackMonths to recover fully-loaded customer-acquisition cost<12 mo great, <18 mo ok
LTV / CACLifetime value vs. cost to acquire a customer3x+ healthy, 5x+ excellent
Magic numberNet new ARR per $1 of prior-period sales & marketing>0.75 efficient
Burn multipleNet cash burned per $1 of net new ARR<1.0x great, <1.5x good
Profitability & margin
Gross margin(Revenue minus COGS) / revenue75-80%+
FCF marginFree cash flow / revenueImproving toward positive; rolled into Rule of 40
Rule of 40Revenue growth % + FCF (or EBITDA) margin %>40%
Productivity (where AI shows up first)
ARR per FTEAnnual recurring revenue per full-time employee>$200k baseline; $300-500k+ top-tier
ACVAverage contract value (ARR / number of customers)Segment-dependent; watch the trend

Ranges are 2026 growth-software benchmarks, not hard rules: the right target flexes by stage, segment, and go-to-market. Engineering maturity (L0-L5) sits alongside these as the "how efficiently is software produced" axis.

Building the function

This is a program to build, not a solo act. Where surface area is large, I would use a small partner ecosystem and selective internal hires rather than trying to touch 130 companies personally.

I would solve the two structural gaps the survey already surfaced:

Company-owned agents

So an individual's agent does not walk out the door when they leave the company.

Agent registry / shared library

So wins compound across the portfolio instead of being rebuilt each time.

I would stay tightly coordinated with the ex-CTOs (engineering and product) and the GTM and finance operating teams, owning the office-of-the-CEO layer that today has no owner.

Risks and guardrails

A core thesis here is that the biggest AI risk is ignoring it, and that value comes from taking risk off the table. So I treat guardrails as a value-creation lever, not a compliance tax: clean AI governance is now something buyers diligence, so getting it right lifts exit multiples while it protects the fund.

Tiered guardrails, not blanket rules

Match the strength of the control to the blast radius, so low-risk work moves fast and only high-risk work gets heavy gates.

1

Internal / low

Drafting, summarizing, research. A human reviews before use. Light governance, high speed. Where most of the portfolio should move now.

2

Operational / medium

Writes to a system of record, touches customer data, or informs a decision. Approval gates, verification against source, audit logging.

3

Autonomous / high

Acts without a human, or is customer-facing in the product. Formal evals, monitoring, rollback, and sign-off before scaling.

Seven risk domains, each with a control

The operating model that makes it real

Buyers now underwrite AI. A company with a defensible AI product story plus clean governance presents as lower-risk and earns a higher multiple. Guardrails are not a brake on the transformation: they are what make it bankable at exit.

Where I want the outcome in 12 months

Portfolio moved

A meaningful share of the target cohort advanced at least one maturity level, and more companies now carry both a defensible AI product story and rising efficiency.

Metrics, not slideware

Lighthouse case studies each with a named result, for example ARR per FTE up 25%+, a Rule-of-40 gain, or net-new AI revenue as a growing share of ARR.

A reusable playbook

Diagnostic, use-case library, and benchmarks adopted portfolio-wide, so each new (behind) company starts ahead.

Function + diligence lens

Office-of-the-CEO capability live (partners, agent registry), plus an AI-defensibility read the deal teams rely on pre-investment.

PSG portfolio

All companies from PSG's public site. Industry is a best-fit estimate from each company's description, not PSG's official taxonomy. A small number of recent investments may not yet appear on the public site.

CompanyStatusRegionIndustry (est.)MaturityDescription