AI Portfolio Management: How to Allocate Resources Across Multiple AI Bets
TL;DR
Companies that win in AI are not running the most projects. They are managing a deliberate portfolio: roughly 60 percent in Horizon 1 (AI enhancements to existing products), 25 to 30 percent in Horizon 2 (emerging AI product lines), and 10 to 15 percent in Horizon 3 (long-shot transformational bets). Most companies are over-indexed on H1 and will be disrupted from below. This guide gives you the decision framework, resource allocation model, portfolio review cadence, and kill criteria needed to manage AI as a portfolio rather than a project backlog.
The AI PM Minute
One tactic to make you a sharper AI PM, twice a week. 60 seconds to read. Free.
No fluff. Unsubscribe anytime.
Why AI Needs Portfolio Thinking (Not Project Management)
Project management is the wrong mental model for AI investments. Projects have defined scope, defined timelines, and defined success criteria. AI bets have probabilistic outcomes, rapidly shifting capability ceilings, and competitive dynamics that can invalidate an investment thesis in six months. Managing AI like a project backlog produces a predictable failure mode: 100 percent of resources go to the highest-confidence near-term wins, and the company wakes up three years later having been disrupted by a competitor who ran a very different portfolio.
Portfolio thinking reframes the question. Instead of "which AI project should we prioritize this quarter," the question becomes "what mix of AI bets gives us the best combination of near-term returns, medium-term positioning, and long-term optionality?" The difference sounds subtle. The strategic implications are not.
The H1 trap
Spending 95 percent of AI resources on copilot features, efficiency gains, and chatbot additions. These are real returns, but they are visible to competitors, easy to replicate, and create no durable advantage. Companies in the H1 trap mistake activity for strategy.
The moonshot mistake
The opposite failure: announcing an AI transformation initiative that consumes budget without delivering near-term value. Without H1 wins generating organizational credibility and cash flow, H3 bets die in budget cycles. The portfolio needs all three horizons.
Treating every AI bet as equal
Distributing resources evenly across 20 AI initiatives means none of them gets enough to succeed. A portfolio approach concentrates resources, kills experiments early, and doubles down on bets that show signal. Peanut butter spreading is a symptom of missing governance.
The Horizon Framework Applied to AI Products
McKinsey's Horizon model (H1, H2, H3) was designed for product portfolio management, and it maps cleanly onto AI investment decisions. The key is translating the abstract horizons into concrete AI product categories with specific time horizons, risk profiles, and success metrics.
Horizon 1: AI enhancements to existing products
Examples: AI-powered search within your SaaS product, copilot features for existing workflows, intelligent routing for support tickets, predictive analytics for existing dashboards. These leverage your existing distribution, customer relationships, and brand. Success is measured in retention improvement, NPS, and feature adoption rate.
Watch out: Commoditization is fast. H1 AI features are easy for competitors to copy because they rely on foundation models that everyone can access. H1 alone is not a strategy.
Horizon 2: Emerging AI product lines
Examples: An AI agent that automates a workflow your product currently just logs, a new AI-native product serving an adjacent persona, a data product built on the proprietary data your core product generates. These require new capabilities but can leverage existing distribution. Success is measured in new revenue and new customer segment penetration.
Watch out: H2 bets require longer investment horizons than most quarterly planning cycles support. Organizations without explicit H2 budget protection kill these investments before they reach product-market fit.
Horizon 3: Transformational AI bets
Examples: Building a foundational AI capability that does not exist in the market, acquiring a startup with a genuinely differentiated model or dataset, betting on a new AI interaction paradigm (ambient computing, spatial AI, physical AI) before it has demonstrated market fit. Success is measured in technical milestones and option value, not revenue.
Watch out: Most incumbents have zero H3 allocation. This is why startups reliably disrupt them. The purpose of H3 is not to guarantee returns, it is to maintain optionality against futures the business cannot yet fully articulate.
Resource Allocation: What the Math Actually Looks Like
Allocation recommendations are only useful when they have concrete definitions. Here is how to translate percentages into headcount, budget, and decision rights for a mid-sized product organization with 10 to 20 engineers working on AI.
H1 allocation (60 percent)
This is your core team working on shipped features. For a 15-person AI engineering team, this means 9 engineers. Governed by normal sprint cycles and quarterly product planning. Success metrics are product KPIs: retention, NPS, feature adoption, support deflection rate.
H2 allocation (25 to 30 percent)
Dedicated team, not shared bandwidth. Shared bandwidth kills H2 projects because H1 always wins in a crisis. Four to five engineers working exclusively on an H2 bet for 6 to 12 months. Governed by stage gates rather than sprint cycles. Success metrics are leading indicators: prototype quality, early customer interviews, beta conversion rate.
H3 allocation (10 to 15 percent)
One to two engineers or a research budget equivalent. This is exploration, not delivery. Output is assessed quarterly on technical milestones and market signal, not on revenue or adoption. Most H3 bets will fail or migrate to H2. That is expected and healthy.
The protected budget principle
H2 and H3 budget must be explicitly protected in the annual planning process. If it is not line-itemized separately from H1, it will be reallocated to H1 within two quarters when the first cost pressure hits. Governance, not intention, is what preserves the portfolio balance.
Build an AI Strategy That Compounds Over Time
The AI PM Masterclass covers portfolio strategy, resource allocation frameworks, and how to build organizational capability for AI, taught live by a Salesforce Sr. Director PM.
Portfolio Reviews: Stage Gates and Kill Criteria
A portfolio without governance is just a list of projects. The governance mechanism that distinguishes portfolio management from project management is the stage gate: a predefined decision point where a bet is evaluated on predefined criteria and either advances, is restructured, or is killed. Kill decisions are the hardest and most important part of portfolio management. Organizations that cannot kill failing AI bets lose the compounding advantage that portfolio management is supposed to provide.
Gate 0: Problem-solution fit (month 1 to 2)
Is there a real problem? Do customers describe this pain in their own words without prompting? Is the AI solution plausibly better than the non-AI alternative by a factor that justifies the complexity? Failure here means the bet is killed before any engineering investment.
Gate 1: Technical feasibility (month 2 to 4)
Can we build a working prototype on current AI capabilities? What is the quality floor we need to achieve for this to be useful? Do we have or can we get the data required? Failure here means restructure the bet (different approach) or kill it.
Gate 2: Product-market fit signal (month 4 to 9)
Do beta users return without prompting? What is the week-2 retention rate? Are users describing the product as significantly better than alternatives? Failure here is the most common and most expensive kill. Sunk cost pressure is high. Kill criteria must be defined before the gate is reached.
Gate 3: Business model validation (month 9 to 18)
Is there a path to unit economics that make sense? What does the cost structure look like at scale? Is customer acquisition cost sustainable? Can this generate returns that justify continued H2 investment or make it worth moving to H1 resourcing?
Organizational Structure for AI Portfolio Management
The right governance model depends on company size and AI maturity. There is no one-size-fits-all org structure, but there are common failure modes to avoid.
Who owns portfolio decisions
Portfolio decisions (which bets advance, which are killed, how allocation shifts) should sit with a small group that includes the CPO, CTO or Head of AI, and a finance sponsor. Keeping it small prevents the committee dynamics that turn portfolio reviews into lobbying exercises.
Avoiding internal competition
When multiple product teams compete for the same AI infrastructure, data, or model access, portfolio efficiency collapses. Designate a platform team that serves all three horizons and is not owned by any single horizon team. Internal competition for shared resources is a portfolio anti-pattern.
The H2 leadership problem
H2 bets fail when they are led by someone whose primary accountability is H1. H1 always wins when priorities conflict. H2 bets need dedicated leadership with a charter that is explicitly separate from H1 metrics. This is often a structural change organizations resist until a competitor forces it.
Reporting cadence
H1 bets should be reviewed in monthly product reviews. H2 bets should be reviewed at each stage gate (not monthly, not never). H3 bets should be reviewed quarterly on technical milestones. The cadences are different because the maturity of signal is different. Applying H1 cadence to H3 produces premature kill decisions.
The Portfolio Scorecard: Tracking Your AI Bets
A portfolio scorecard makes the abstract concrete. It is a single view of all active AI bets across horizons, their current stage, their key metric, and their kill threshold. Here is a lightweight template adapted for AI product portfolios.
Bet name and one-line hypothesis
What are we betting on and why do we believe it will work? Should fit in a single sentence.
Horizon classification (H1, H2, H3)
Which horizon does this bet belong to? This drives resource expectations, review cadence, and success criteria.
Current stage gate
Which gate are we at? When is the next gate decision? Who owns it?
Leading indicator (the signal we are tracking)
One metric that is predictive of success at this stage. Not revenue. Not user count. The specific signal that tells us this bet is working at its current maturity.
Current value and target
Where is the metric now? What does it need to reach to pass the next gate?
Kill threshold
If this metric reaches this value by this date, the bet is killed. Non-negotiable. Defined before the gate, not at it.
Resource consumption (headcount and compute cost)
Monthly spend including engineering time and inference costs. This becomes a portfolio-level signal: bets that consume more than their horizon allocation are crowding out other bets.
Build Your AI Strategy From Portfolio to Execution
The AI PM Masterclass covers how to design AI strategy, structure portfolio decisions, and build organizational capability to execute across all three horizons, taught live by a Salesforce Sr. Director PM.
Related Articles
Before you go: get the AI PM Minute
One tactic to make you a sharper AI PM, twice a week. 60 seconds to read. Free.
No fluff. Unsubscribe anytime.