Omnicom's third-party service costs hit $2.9B in the first half of 2026, up from $1.7B a year earlier, while an IAB Europe demo burned roughly €50 in tokens on a single three-agent task that finished in 25 minutes. Nobody, holdcos included, has a defensible number for what a token should cost a client.
Why it matters: When a client asks what agentic actually costs, the answer should be a metered number you can show them. Cost-per-decision belongs in the build spec, not the sales deck.
Nils Rooijmans' Quality Score Agent moved conversions up 67.7% and cost-per-conversion down 11.9% on a live account, built on a script-then-automation-then-agent progression where autonomy is earned in stages. Separately, across roughly 115 million retail clicks, AI Max advertisers saw invalid traffic climb from 2.46% to 5.28% in six months while their standard search campaigns got slightly cleaner, and Google begins auto-migrating many retailers in September.
Why it matters: This is the shape that survives a client account: one narrow agent with a measurable test arm, not a platform migration. If you run retail search, the September auto-migration is a budget conversation to have now, not in November.
Kevin Lee's correction is to define a money-query set of a few dozen real decision prompts, track citation share only for that set, and tag AI-referred sessions through to CRM outcomes rather than counting mentions. Two things to internalize: ranking first in Google no longer predicts appearing in the AI answer for the same query, and there is a several-week lag between publishing and showing up.
Why it matters: A money-query set is a cleaner deliverable than a visibility score, and it is the version a CFO will actually fund. Judge it by the quarter, not by week two.
Metrifyr federates GA4, Search Console, AdSense, Tag Manager and Google Ads behind a single MCP connection, and its sharpest tool ranks organic pages by real AdSense earnings instead of clicks, so a 200-click page at a $12 RPM outranks a 2,000-click page at $0.40. Read-only by default, with the tool surface scoped by which accounts you actually connect.
Why it matters: The report you rebuild by hand every month is the first thing that should become a tool call. The value is the cross-product join, not the raw number.
ECD Digital Strategy's Claude-built Email Brief Writer took a 10-brief planning session from an hour to 15 minutes, and a single approval now fires four systems at once: the send-schedule sheet, an eight-subtask Asana project, a draft Klaviyo campaign with audience and send time populated, and a Google Doc copy file. Monthly Klaviyo performance flows back in over MCP, so each cycle's briefs are shaped by what actually worked three months ago.
Why it matters: The speed gain is the headline, but the per-client voice-and-rules profile held in version control is the part that compounds. Six months in, the system holds more context on each account than any one person can.