MIT Gave CMOs a 101-Task AI Scorecard — Use It

For the first time, there is a public, citable, task-by-task scorecard for agentic AI work. On June 29, 2026, MIT Technology Review Insights, in partnership with Microsoft, published the 2026 Agent Confidence Index — 300 global technology leaders ranking 101 tasks across AI, data, and cloud workflows by the confidence executives have in agents acting on their behalf. The average confidence score is 64 of 100. Automated business report generation scores 83.5; service mesh configuration scores 37.5. Fifty-nine percent cite "keeping humans in the loop" as a top priority for adoption.

The Index is the first public artifact in the cycle that lets a buyer respond to AI services claims with a number, not a slogan.

The scorecard, by band

The 101 tasks cluster into three bands the Index makes explicit. The high-confidence band (80-plus) is execution work the consulting-tier AI services layer is now visibly packaging. Automated report generation at 83.5, boilerplate code at 82.5, certificate expiration monitoring at 81.5. These tasks a hyperscaler, a holdco, or a consulting firm can credibly claim a layer for, because the agent's output is observable, verifiable, and bounded.

The medium-confidence band (50 to 80) is synthesis work. Multi-step summarization, cross-source intelligence synthesis, audience signal interpretation, competitive landscape mapping, message testing at scale. This is the band an AI-native strategy agency is built to operate in, with the agent doing the work and a named human signing the call.

The low-confidence band (below 50) is board-defining work. Service mesh configuration at 37.5, database schema migration scripting at 46.5, disaster recovery testing at 43. The named human is the whole answer in this band — board reviews, regulatory disclosures, brand-defining moments. There is no agent-led version of the work the buyer is actually procuring.

Why this matters in the current AI services market

The week of July 2, 2026 saw Microsoft launch Microsoft Frontier Company — $2.5B and 6,000 embedded engineers, named-marquee clients Unilever and Novo Nordisk — on the same day WPP detailed its Enterprise Solutions five-service portfolio and David Droga told The Drum he is "assembling Avenger teams from across Accenture." Three named consulting-tier positions, all claimed in 72 hours. None cited the Index. All cited headcount and dollar amounts. The Index is the public benchmark a buyer can now use to ask each one which of the 101 tasks their layer is actually being measured against.

Three procurement questions, anchored on the Index

For any CMO with an AI services procurement decision in front of them in the next 90 days, three questions turn the Index from a benchmark into a contract.

Question 1 — Which of the 101 tasks is your agent stack scoring 80-plus on, and which is it scoring below 50? The answer has to be task-by-task, not portfolio-level. If the partner answers "our stack is industry-leading," the partner has not read the Index. If the partner answers with named task numbers, the partner is measuring their own capability against the public benchmark the buyer can cite. The named task list is the first wedge.

Question 2 — Which named human is accountable for the work in the medium and low bands? The Index has already told the partner that the medium and low bands are where named-human accountability matters. If the partner cannot name the human, the partner is selling the wrong band. If the partner names the human but that human is not the one signing the decision memo, the partner is selling the band but not the counsel. The signature on the memo is the second wedge.

Question 3 — Which subscription ends the tool sprawl, and which adds a layer to it? A subscription that delivers one named human, one decision memo, one chain of accountability ends the sprawl. A subscription that adds another dashboard, another license, another team to operate, is adding to it. The tool-sprawl test is the third wedge.

Three questions. Three wedges. Each one anchored on a named number in a public Index. The CMO who walks into the next AI services pitch with these three questions owns the conversation.

What the Index does to "AI pitch-maxxing"

The Index converts the seller's "AI pitch-maxxing" — to use Publicis CEO Arthur Sadoun's Cannes framing — into the buyer's task-by-task scorecard. The buyer now has a published benchmark to score against. The seller no longer controls the framing.

The model layer is also visibly fragmenting in the same week. BleepingComputer reported Claude Fable 5's global re-access disappointed users with "nerfed performance" — capped at 50 percent of weekly usage limits and routed to Opus 4.8 even when the task does not appear to be a safety risk. The buyer-side question is no longer "who has the best model." It is "who is accountable when the model is throttled, when the agent is rerouted, or when the consulting-tier offer rotates its portfolio."

The scorecard an AI-native strategy agency is built to deliver

Autostrat operates on the medium-confidence band and the human-in-the-loop band of the Index. The agent stack does the synthesis — audience segmentation, competitive signal summarization, market brief generation, content tagging, message testing at scale. The named human on the agency's side signs the call, writes the decision memo, and owns the recommendation when the agent's work has to be defended in a board review.

The subscription is one. The named human is one. The decision memo is one. The chain of accountability is one. The tool sprawl the buyer's team is currently running to do the same work ends when the subscription lands. The Index is the buyer's first public scorecard for telling which AI services partner actually delivers the named-counsel layer, and which one is selling the band below it.