The verb is proactivity, and you only earn its return where the lift actually is.

Issue 01 was about a noun. You cannot manage a level you cannot name, so we named six. This issue is about a verb. Proactivity is the act of moving first: saying the thing before the buyer asks, touching the account before the quarter forces it, spending before the return is proven. The trouble with verbs is that they cost something, and the three papers on the desk this week agree on the bill. The return on acting first is real. It is also uneven, and it is only collected by spending where the incremental lift actually sits. That last clause is the whole issue.

Probe the concern, do not broadcast louder

Start with the message, because that is where proactivity gets confused with volume. The instinct under pressure is to pitch harder. The feature this week says the opposite. In a food-delivery outbound-call simulator, the agents that persuaded were the ones that surfaced the user’s hidden concern first, then steered [1]. The method conditions a small open model on those concerns and distills that privileged view into the deployable agent. On the merchant task, simulated acceptance rose from 26.2 to 48.5 percent for one 8B model, rivaling Claude-Sonnet-4.5 at 50.8 and far above a reward-shaping baseline (GRPO) at 30.5 [1]. A 4B model went from 20.2 to 46.3.

The result that matters most to an operator is the ablation. Strip the concern-probing component out, and performance falls back toward the GRPO regime [1]. The authors’ reading is blunt: reward shaping alone cannot manufacture proactivity. You cannot bonus a rep, or an agent, into asking the right question. The question has to be built in.

Read this with the limits on. It is a simulator, it is one domain, and it is not yet peer reviewed. The transfer to a B2B sales floor is an analogy, not a finding. But the shape is the lesson for the Message pillar: the outbound that converts is the one that probes the concern, not the one that broadcasts the pitch. The feature, “Probe the Concern,” carries the full table.

Score the lift, not the lead

The Signal and Measurement pillars get the harder correction. Most lead scoring asks the wrong question. Propensity ranks the accounts most likely to convert. But many of those accounts would have converted without you, so a touch spent on them buys nothing. The honest question is uplift: how much does your acting first change the outcome for this account. This is the direct sequel to Issue 01’s lead-scoring piece, and the ai-research desk this week, “Score the Lift,” makes the case in field data.

A treatment-gated, value-weighted uplift model ranked persuadable high-value accounts about 17 percent better by Qini than the strongest prior method [2]. That figure is derived: 0.3049 divided by 0.2596 is 1.17. The number that lands harder came from a four-month A/B test at one cloud provider.

One field test, one company, a preprint. Hold it loosely. But the structural point survives the caveats: a propensity list and an uplift list are not the same list, and proactivity spent on the propensity list is mostly spent on accounts that did not need you.

Spend with discipline, because the dividend is a probability

The strategy desk pulls the camera back to the whole AI bet, and the lesson rhymes. Reading the capex ramp as a rational bet (381 billion dollars in 2025 across the five largest U.S. tech firms, a forecast 755 billion in 2026), the macro model infers that each AI boom must lift AI-sector productivity about 2.7x [3]. That assumption, run through a two-sector model, fans out into a range, not a forecast: 5 to 58 points of extra cumulative GDP growth by 2030 [3]. The honest part is the distribution. Across the simulated paths, the modal outcome is no boom at all, 44 percent of paths [3]. The dividend is a coin-flip-wide probability, not a date on a roadmap.

Now hold the two 2.7x figures side by side, and resist the urge to connect them. The uplift paper’s 2.7x is a measured ratio of incremental revenue per account in one field test. The macro paper’s 2.7x is a calibrated parameter, a revealed-preference guess about productivity backed out of capex. They measure different things, on different evidence, at different confidence. The coincidence is a warning, not a thesis: a clean multiple is easy to repeat and easy to believe, and belief is exactly what spend discipline has to survive.

The pipeline lesson from the macro view is the same as the micro one. Treat your own AI dividend as uncertain. Hold out a control, measure the incremental lift the way the uplift desk does, and fund the pillars that compound rather than betting the budget on the boom landing on schedule.

The practice change

One verb, three moves, in the order the pillars run. Probe the concern before you pitch, so the Message earns a reply. Score the lift, not the lead, so the Signal points at accounts your action actually changes. Spend with discipline, holding out and measuring, because the dividend is a probability and not a fact. Each of the three pieces this week carries the full evidence and its limits. The pattern is the one this magazine keeps: name the level, read the evidence, change one practice.

If you want a calibrated read on which pillar your proactivity is currently wasted on, the issue closes with the instrument, and the next one lands every week. Get the Weekly Research and start with the rate you cannot yet defend.