Last year's plan holds a handful of programs you funded to see what would happen: the second webinar series, a regional event because a rep asked for it, a display flight too small to register in any test, a podcast sponsorship nobody measured. None of them is failing, and none of them has produced a number you would put in front of the CFO.
Those lines are the middle of your plan. When the CFO asks for money back in March, they go first. After a budget cut this year, a software client's paid program was spread thin across too many campaigns, and I told the team that one thing done well beats five things at a third each.
Nassim Taleb's barbell gives the plan a different shape. Put about 80% of the budget on programs with causal evidence and up to 20% on small, capped bets with a real shot at a large payoff. Fund nothing in between. I call the marketing version the Marketing Barbell, and below is the one-page sheet that sorts every program onto one end or off the plan.
Taleb's barbell, briefly
Taleb's barbell comes from bond trading, where it means holding the two ends of the yield curve and nothing in the middle. In investing, most of the capital sits in instruments that cannot hurt you. A small slice goes to positions with a large upside and a known maximum loss. In a 2012 keynote at GAIM International, Taleb put the split at 80% in zero-risk securities and 20% in maximum-risk securities, with nothing in medium-risk securities (Hedgeweek, June 2012).
For a CMO, the useful part is what the barbell leaves out. Medium-risk positions look prudent, but they carry a downside nobody priced and an upside too small to matter. The safe end limits what a bad year can cost. The speculative end is where, in Taleb's words, "error and uncertainty help," because each loss is capped and a single win can pay for the rest (Hedgeweek, June 2012).
What the middle looks like in a marketing plan
A program belongs in the middle when it fails two tests at once. It has no causal evidence that it works, so it cannot sit on the proven end. Its best realistic outcome is too small to change the year, so it does not qualify as a bet.
System1 found that 52% of UK responses to the average TV ad are neutral, meaning the ad made people feel nothing, and 47% of US responses are neutral (System1, July 2024). Those ads still work a little. System1 estimates that US brands would need an extra $189 billion of media for their dull ads to match the commercial effect of ads that are not dull. Neutral creative passes review and buys less effect per dollar than work people feel something about.
A channel pilot funded too low to read follows the same path. It never graduates and never gets killed, and three years later it is still in the plan.
Middle programs survive planning because each is small and has a sponsor. Neither fact helps in a cut. The CFO asks two questions of every line. What breaks if we stop, and what could this become? Proven programs answer the first, and bets answer the second at a known cost. A middle program answers neither.
The proven end, about 80%
The proven end holds programs with causal evidence. That means a holdout, a geo test, a matched-market read, or a media mix model checked against at least one experiment. Attribution reports do not qualify, because last-click credit and platform-reported conversions do not show what would have happened without the spend.
Inside this end, keep the brand and activation split that Les Binet and Peter Field found in the IPA Databank. Their optimum in The Long and the Short of It was 60:40 brand to activation, refined to 62:38 in Effectiveness in Context (IPA, September 2023). For B2B, their work with the LinkedIn B2B Institute puts the balance at 50/50 (LinkedIn B2B Institute). Binet and Field set the split by time horizon, and the barbell sets a second split by risk on top of it. A brand program with a tracked lift in mental availability sits on the proven end, and one with no measurement sits in the middle.
My default is 80% on this end. Move to 90% when cash is tight or the board has asked for predictability over growth. Move to 70% when the proven channels are saturating, which shows up as incremental CAC rising for two straight quarters while volume holds flat.
The bet end, up to 20%
A bet needs a small, written maximum loss and a plausible upside of at least five times its cost. Those thresholds are my own, and no published standard sets them.
Microsoft's Bing team shows why this end exists. An engineer proposed a change to how ad headlines displayed. Program managers rated it low priority, and it sat for more than six months although it needed only a few days of engineering. When someone finally tested it, revenue rose 12%, worth more than $100 million a year in the US alone (Kohavi and Thomke, HBR, 2017). It was the best revenue-generating idea in Bing's history.
The same article explains why one bet is never enough. At Google and Bing, only about 10% to 20% of experiments produce positive results. Across Microsoft, about a third of ideas work, a third are neutral and a third do harm (Kohavi and Thomke, HBR, 2017). The authors add that even experts often misjudge which ideas will pay off. Your team will misjudge too, so the bet end has to hold enough attempts for the odds to work.
My default is four or five bets, each capped at 4% of total budget. If your hit rate matches Bing's 10% to 20%, five independent bets give you between a 41% and a 67% chance of at least one winner. Before money moves, each bet gets a one-paragraph card listing the maximum loss, the upside case in dollars, a single leading indicator, and a read date no more than two quarters out. Fund each bet at the level its test needs to produce a readable result. A bet too small to read is a middle program with a better story.
Write both exits before launch. A bet that misses its indicator at the read date ends that week, and its money returns to the bet pool. Winners go through an incrementality test, then graduate to the proven end and free their slots. Allow one extension at most, and treat a second request as a sign the program has drifted into the middle.
In Partnership with Wispr Flow
Speak naturally. Send without fixing.
The Barbell Sheet
This page goes in the plan. Every program gets a row as Proven, Bet or Middle, and the Middle row should be empty by the time the plan ships.
End | Default share | Entry test | Sizing rule | Exit rule |
Proven | 80% (range 70-90%) | Holdout, geo test, or experiment-checked MMM | Fund to the incremental payback limit | Incremental CAC rises two straight quarters |
Bet | Up to 20% | Written max loss, upside at least 5x cost | Four or five bets, each capped at 4% | Misses its indicator at read date, ends that week |
Middle | 0% | Fails both tests above | Not funded | Rebuilt as a bet or cut at plan |
Each program goes on the sheet with its evidence or its bet card attached. A program with neither goes in the bottom row.
Some middle programs can be rescued. Give one a sharper hypothesis, a budget large enough to read, and a read date, and it becomes a bet. Run a proper test on another, and if it holds up it moves to the proven end. Cut the rest in the plan you submit, and list them on one page with the dollars they free.
How this sits next to the triage
In September I argued for a triage on every existing budget line, with kill, automate and defend as the verdicts, and a rule that new capability is funded only from savings already booked (Plan for the Budget You Won't Get).
The two meet at the line for new capability. The barbell says how to spend those booked savings, as bets with caps and read dates, so a new AI program does not settle into next year's middle. When the CFO asks for 10% back, start with the bet pool. Every bet there carries a written maximum loss, so ending one early costs what its card says and no more. The proven end stays at its agreed share through the whole conversation.
Before the plan locks
This week, put your top 20 programs by spend on the sheet. Bring the middle row to your next staff meeting and give each owner a choice. They can rebuild the program as a bet with a card and a read date, or let it go.
Then walk the bet cards through with the CFO before the plan goes up, so the maximum loss on every bet is agreed before the first cut conversation of 2027 begins.
James