The CMO job description has changed. For thirty years, a marketing chief ran a programme — a set of campaigns with defined audiences, creative, channels, budgets, and dates. Budget was committed at the start of the cycle and reviewed at the end.
That apparatus does not fit marketing AI.
A marketing AI investment is not a campaign. It is a capital position in a production asset that generates uncertain returns over time, whose performance decays, whose risks shift, and whose correlations with other positions change as underlying models and data evolve. The right mental model is portfolio manager.
That claim is the spine of The Agentic CMO. Everything else — measurement, governance, legal liability, agency economics, brand coherence — is the apparatus required to discharge the job. This article sets out the framework in operating-manual detail.
Why a Financial Frame and Not a Brand Frame
Four structural properties define agentic marketing investments and align them with financial positions rather than campaigns.
Dispersed outcomes. A campaign’s outcome at launch is a distribution around a point estimate with relatively low variance. An agent investment — an outbound sales-development agent, a retention-intervention agent — has outcomes with far wider dispersion. You do not know whether the model outperforms the baseline, equals it, or falls short until you run it. Results shift over time.
Path dependence. A campaign completed is gone. An agent deployed persists. Its outcomes this quarter change the data it is trained on next quarter. Serial correlation of returns is the rule.
Optionality. An agent investment is also a real option to expand, contract, extend, or retire. The value of that option depends on conditions — model quality, infrastructure cost curves, regulatory regime — that change independently of the agent’s own performance.
Shared risk. Two agents built on the same foundation model share vendor concentration risk, prompt-drift risk, correlated failure modes, and, under the EU AI Act from 2 August 2026 onward, correlated regulatory exposure. These correlations must be budgeted or the portfolio will concentrate risk faster than it accumulates return.
These four properties — dispersion, path dependence, optionality, shared risk — define the conditions under which modern portfolio theory was formulated. Harry Markowitz’s 1952 paper in the Journal of Finance made the case that an investor holding uncertain positions must optimise not for the expected return of each position alone but for the joint distribution across the set. The mean-variance frontier is one expression of that logic. Every active-management tradition since — Sharpe, Treynor, Black, Litterman, Grinold, Kahn — refines the same insight: the position is not the unit of decision. The portfolio is.
The brand-coherence problem is, structurally, a correlation problem. A fleet of agents trained on a shared foundation model, fed shared brand-voice assets, prompted from shared scaffolding does not produce uncorrelated output. It produces correlated output that drifts together. Locally, each agent does what its brief asks. Globally, the agents pull the brand toward whatever cluster the underlying model finds most natural — a cluster set by training data the CMO did not choose.
The thesis of the book — that brand coherence is the dominant failure mode of agentic marketing — and the apparatus of this chapter, the portfolio framework, are the same argument stated at different levels of abstraction.
The Empirical Punch
Brinson, Hood, and Beebower’s 1986 paper in Financial Analysts Journal decomposed pension-fund returns across 91 large US plans over 1974 to 1983 and found that around 93.6 per cent of the variation in total return across plans was explained by asset-allocation policy, not by security selection or market timing.
Allocation dominates selection.
In marketing AI terms: which agents you fund, and in what proportions, determines portfolio return more than the internal optimisation of any single agent.
A CMO operating in campaign-programme mode spends roughly 90 per cent of attention on selection — which tool, which vendor, which prompt framework — and roughly 10 per cent on allocation. A portfolio-manager CMO inverts that ratio. That inversion is the first deliverable of the framework.
Position Sizing
Position sizing is the first question a fund manager asks and usually the last question a CMO answers. Given a fixed annual budget for agent investments, how much capital goes behind each position, and why?
The temptation is to size by business case. The agent with the largest projected return gets the biggest cheque. Fund managers know this is wrong. Marketing leaders are learning it the hard way.
The right answer uses four inputs and two constraints.
Expected value per quarter (EV). A central estimate of quarterly incremental gross profit attributable to the agent, net of operating cost and depreciation on model assets. Not revenue. Contribution to gross profit at the firm level.
Outcome variance (σ²). The dispersion of EV under realistic stress. An outbound agent pointed at an ICP-adjacent list may return close to nothing for two quarters and produce a windfall in the third — its variance is high. A creative-variant agent writing the sixth version of a proven ad concept has low variance. Variance is estimated empirically in pilot, not assumed on spec.
Governance load (G). The share of CMO and direct-report attention the agent consumes. A customer-contacting agent with brand exposure imposes a larger governance load than a back-office media-mix agent, even at the same EV. Governance load is the operating cost most CMOs ignore and the one that breaks the budget first.
Strategic fit (S). The extent to which the agent serves a commitment the CEO has made to the Board this year. A retention agent in a year when gross retention is the CEO’s lead metric has higher S than a demand-generation agent, even if the demand-gen agent’s standalone EV is higher.
The constraints are the total capital budget and a sub-budget for governance attention. A portfolio cannot exceed its attention budget any more than it can exceed its cash budget. The symptom of breach is the same — degraded quality, missed overrides, regulatory near-misses.
A defensible sizing rule: Capital = k × (EV/σ) × S × (1/G), subject to total capital ≤ budget and total governance load ≤ governance capacity.
The form matters more than the precision. It says: allocate more to higher risk-adjusted expected return, more to positions aligned with strategy, less to positions that consume management time, and stop when either cash or attention runs out.
In its risk-adjusted form it echoes Grinold’s fundamental law of active management — information ratio equals information coefficient times the square root of breadth — which states that return scales with the product of skill and the number of independent bets. A portfolio earns by placing more positions whose returns are not perfectly correlated, not by doubling down on one favourite.
The Seventeen Per Cent Reserve
A portfolio of agent positions sized at one hundred per cent of available capital is the marketing equivalent of a margin call waiting to happen.
The discipline of running a roughly seventeen per cent cash reserve is non-negotiable. It is not padding. It is dry powder for mid-year opportunities — a new foundation model, a new data source, a regulatory carve-out, an acquired competitor’s stack to integrate — and the single most useful line item in the whole allocation.
Fund managers call it cash. CMOs should call it cash. The agents that compound the most over a five-year horizon are rarely the ones budgeted at the start of year one. They are the ones available to be sized into mid-cycle when the opportunity appears.
A portfolio that is one hundred per cent invested is a portfolio that has surrendered the option to act on new information. That is the most expensive surrender in the framework.
Lifecycle Management
Agents have three phases: incubation, production, retirement. Campaigns have two: alive or complete. The missing phase is the one that gets CMOs into trouble.
Incubation. A pilot with defined scope, a protected budget envelope, and explicit decision gates. The goal is not return but the estimation of EV and σ under realistic operating conditions. Success criterion: the pilot produces a defensible forecast another executive would stake their budget on. Failure criterion: after the scope period, neither EV nor σ can be estimated with a standard error below thirty per cent of the mean.
Production. The agent is resourced, governed, and measured as an ongoing asset. Capital is sized per the rule above. Measurement runs on a quarterly cycle. Performance is reported against the benchmark set, not against the pilot forecast.
Retirement. The agent is wound down when it crosses one of four gates: performance decay below the do-nothing baseline for two consecutive quarters; strategic-fit collapse; cost-curve inversion (a cheaper alternative delivers within ninety per cent of its EV); or regulatory inadmissibility. Retirement is treated as a normal event, not a failure. A portfolio that never retires a position is either very lucky or not measuring.
The decision gates between phases are non-negotiable.
Gate 1: incubate → produce. EV standard error below thirty per cent of the mean. Governance playbook documented with a minimum of four override rules and a named human owner for each. Brand, legal, and data-protection review signed in writing. Attribution method specified before production, not inferred after.
Gate 2: produce → retire. One of the four retirement triggers has fired and been verified against two independent measurement methods. A wind-down plan exists with dates, reassignment of human owners, and data-retention handling compliant with applicable regulation. A post-mortem document is scheduled within thirty days of retirement and circulated at the next portfolio review.
The CMO must not run agents in perpetual pilot. The most expensive agents in large enterprises today are the ones that have sat in pilot for eighteen months, resourced like production assets but measured like experiments — exposed to the discipline of neither stage. They are marketing AI’s equivalent of the zombie fund.
The Risk Budget
Risk in a marketing AI portfolio is not a single number. It is a set of exposures, each measured, each limited.
Brand risk. Exposure to the probability that agent output damages brand equity. Limit: no single agent may carry more than a stated fraction of portfolio brand risk. A typical cap is twenty per cent.
Legal risk. Exposure to regulatory, liability, and contractual breach. Under the EU AI Act, high-risk systems carry direct deployer obligations. Concentration in high-risk Annex III categories is itself an exposure to be budgeted.
Customer-experience risk. Exposure to autonomous decisions that materially degrade the customer journey. The cap here is qualitative and tracked through complaint volume, NPS deltas, and journey-completion rates.
Operational risk. Vendor concentration, model concentration, and integration fragility. If three of four agents run on the same foundation model, one provider outage shuts down most of the portfolio.
Vendor risk. Procurement risk, financial stability risk, and contractual leverage risk. Agentic AI vendor markets are consolidating fast. The vendor sized at fifty per cent of the portfolio today may be a different company by year-end.
Model risk. Drift, degradation, and bias risks specific to the underlying models. Measured continuously, reviewed quarterly.
The risk budget is not a separate document. It is a constraint that sits underneath the position-sizing formula and constrains the portfolio composition.
The Portfolio Audit
Three questions for the portfolio review. None of them is rhetorical.
Run the position-sizing formula on every active AI investment. Most portfolios will reveal at least one position that is consuming twenty to thirty per cent of governance attention against a wide-variance EV — the kind of position the formula refuses to over-concentrate on, but which the business case alone would size aggressively.
Identify the three positions most at risk of failing a retirement gate in the next two quarters. Begin the wind-down conversation now. Retirement is not failure. Perpetual pilot is.
Audit your portfolio for shared risk. How many positions run on the same foundation model? Same vendor? Same training data lineage? Concentration that does not show up in any individual business case shows up in the risk budget. Cap it before the regulator does.
The correct governance unit is the fund. The sooner the marketing function moves from selection-dominated to allocation-dominated thinking, the sooner it earns the right to be governed as a strategic asset rather than an expense line.
By the end of 2027, the firms that have retired agents will outperform the firms that have not, and the gap will be visible in the contribution line. The CMOs still carrying zombie positions in perpetual pilot — because retirement is the one call nobody wants to make — will be looking for new jobs.
The distance between portfolio manager and zombie programme is a decision taken in the next fourteen days.
Keep Reading
That’s all for this week book chapter summary, come back next Monday for the next chapter summary.
The Agentic CMO - Second Edition is available today in hardcover, paperback and ebook.
Disclaimer: The views and opinions expressed in The Agentic CMO, Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.