Verification can’t be delegated. Martech Futurist | August 12, 2026

Verification is now the largest recurring line in what you pay to run AI, and you are paying it on meters other companies own. Decide which instruments you control before that bill compounds.

Two editions ago I argued that evidence quality belongs in procurement. On August 10 I pushed that further: verification you perform once at purchase expires, and the party holding the measurement instrument sets your decision timing.

This edition takes the next step, which is financial. Gartner's August 10 forecast puts a number on what continuous verification costs, and the number lands on a meter you rent.

Price your AI as a recurring meter, because that is what it became

‍Gartner forecast on August 10, 2026 that worldwide AI-optimized infrastructure as a service spending will reach $42.3 billion in 2026, growing 96.4%, and $66.1 billion in 2027. Inference spending of $23.3 billion will surpass training spending of $19 billion this year. Fifty-five percent of AI-optimized IaaS spending supports inference in 2026, reaching 59% in 2027. Sr Principal Research Analyst Hardeep Singh attributes the shift to organizations moving from model development to production deployment, where fine-tuned and domain-specific models run continuously inside customer-facing systems.

Read that as a marketing operations fact. Training costs sat with model builders. Inference costs sit with whoever runs the workload, which is you, every time an agent executes a step. Your AI line moved from a capital decision made once to a consumption line that tracks execution volume.‍ ‍

That maps directly to compute economics as a Force in my AI capability framework. Forces bound the system from outside. You contend with them, and this one now sets the ceiling on how much agentic work you can afford to check.

Treat the check itself as the budget line, because it is most of the meter

McKinsey's QuantumBlack researchers published a figure in July that pairs with Gartner's. They put 60% of total agentic AI spend on response refinement, meaning the iterative checking, correcting, and improving that runs before an agent delivers a usable output. In the same report, McKinsey cites its May 2026 survey finding that 93% of organizations exceeded their AI budgets.

Combine the two findings and the picture is specific. The dominant AI cost category is now inference, and the majority of inference in agentic workloads is verification. The discipline I have been arguing for across four editions carries a price tag, and that price tag is most of your run cost.

Work an example. Your team runs an agentic campaign QA workflow 500 times a day across 250 working days, so 125,000 executions a year. Primary inference costs $0.80 per execution, which totals $100,000. Apply McKinsey's split: primary inference is the remaining 40%, so your true annual run cost is $250,000, and $150,000 of that is checking. Across 125,000 executions averaging three refinement passes, each pass costs about $0.40. Remove one pass on the 70% of executions that clear on the first check and you take out roughly $35,000.

You can only find that $35,000 if you meter per step. The vendor invoice gives you a monthly total. The per-pass number requires instrumentation you build.

It reminds me of what David Funck, Chief Technology Officer at Avaya, said when I interviewed him on The Agile Brand podcast: "The contact center is a great place to put AI out there and to really measure how well it performs because it is a constrained environment that already has very sophisticated measurement tools."

Funck names the right sequencing. Deploy agents first where you already hold the instruments, then extend outward as you build new ones. Teams that reverse that order buy autonomy in environments they cannot read.

Bassem Hamdy, CEO and Co-Founder at Briq, made the cost side of this concrete on the podcast: "The president pulled me aside and goes, 'Yes, we saved two headcount. But last year we missed an insurance certificate. There was a loss on a project and that cost us $500,000 and a deductible.'"

The savings appeared in a staffing plan. The failure appeared nowhere until it arrived as a loss. Marketing has the same asymmetry, and the verification budget is what closes it.

Buy at least one instrument that sits outside the vendor relationship

Insygna launched the Insygna Agent Report Card on August 10, a free service that tests an AI agent and returns a security score before that agent touches company or client systems. The score runs out of 100 across six dimensions: Secret Exposure, Dependency Vulnerabilities, Code Security, Container Hardening, LLM Security measured against the OWASP LLM Top 10, and Image Security. Findings come back at file and line level with version control and a shareable verified badge.

The number worth carrying into your next vendor conversation: Insygna reports the median agent it has tested scores below 50 out of 100, with roughly six in ten below that mark.‍ ‍

CEO Michael Beygelman frames the gap as a hiring analogy, noting that an agent receives system access like an employee and none of the verification. He also states the market needs a reputable third party to test agents and report risk honestly. Note the commercial position: Insygna is pre-revenue and running a closed beta through 2026, so treat the free service as a market-entry move and the median score as vendor-reported.

The structural point survives the caveat. This is a verification instrument that does not belong to the company selling you the agent, which is the property my August 10 argument said you need.

Google shipped the identity half of the same problem. With general availability of the Gemini Enterprise Agent Platform, every deployed agent carries its own SPIFFE-based identity, its own credentials, its own permission scope, and its own audit trail, and logs record both the agent and the user when an agent acts on someone's behalf. That trail belongs to your tenant. It is a record you can read without asking a vendor to run a report.

Hold that against Google's consumer Gemini Spark agent, which operates desktop Chrome using a person's logged-in accounts and saved passwords. One design gives the agent a name. The other lets the agent borrow yours. Every marketing leader piloting agents across Salesforce, HubSpot, or a bespoke stack picks one of those two patterns, usually without registering it as a choice. This is the Identity and Permissions layer doing load-bearing work, exactly where my framework predicts it matters most: as orchestration moves toward the unsupervised end.

For a working definition of the execution surface this governs, see Agentic CX in The Agile Brand Guide.

Audit your Google Ads account before September 1, because Google supplies both the change and the evidence for it

Google confirmed by email on August 5 that campaigns using automatically created assets or the campaign-level broad match setting will upgrade automatically to AI Max for Search on September 1, 2026. Campaigns with automatically created assets move over with search term matching and text customization enabled by default. Campaigns using campaign-level broad match receive search term matching alone. The five-month reprieve Google granted Dynamic Search Ads in June does not apply here.

The mechanical detail matters more than the deadline. Automatically created assets governed creative generation only. Search term matching governs which queries make your ad eligible. Advertisers who opted into automated headline writing inherit automated query expansion on September 1 unless they change the setting.

Now the evidence question. Google reports that AI Max campaigns running the full feature suite see an average of 7% more conversions or conversion value at similar CPA or ROAS, based on internal 2026 data excluding retail advertisers. Independent testing has repeatedly reported weaker outcomes. Text customization now requires AI Max to operate, which converts the migration into a technical dependency.

So Google changes the execution mechanism, sets the date, and publishes the number you would use to evaluate the change. Advertisers who hold no independent read enter Q4 with performance swings and one available explanation. Advertisers who baseline CPA and ROAS now, and run a holdout, hold a second one. Incrementality testing is the instrument that produces it, and August is when you install it.

This is Agile Brand Principle 4 applied to measurement. Staying true to your values under a platform migration means keeping a source of truth you can defend to your CFO.

Featured Insights

Gartner. Gartner Forecasts Worldwide AI-Optimized IaaS Spending to Grow 96% Through 2026 (August 10, 2026). Inference spending passes training spending this year, and agentic execution drives the shift. Move your AI forecast off a project line and onto a consumption line tied to execution volume. Ask your finance partner to model it the way they model cloud, then set a per-workflow ceiling.

Insygna via AIThority. Insygna Launches Free Public Service to Help Companies Stop Rogue AI Agents (August 10, 2026). Independent pre-deployment scoring across six security dimensions, with a reported median agent score below 50 out of 100. Run one third-party test on the agent you are closest to approving. The result gives your procurement team a number to negotiate against.

Google Cloud, via AI Agent Store's week of August 11 roundup. Gemini Enterprise Agent Platform reached general availability with per-agent SPIFFE identity, scoped permissions, and full operation logging. Name every agent in your marketing stack and confirm which ones run on shared credentials. Any agent you cannot identify in a log is an agent you cannot audit after an incident.

PPC Land. Google Ads Broad Match Campaigns Face AI Max Auto-Upgrade on September 1 (August 6, 2026). Automatic migration of automatically created assets and campaign-level broad match, with Google's internal 7% performance figure and weaker independent results. Baseline CPA and ROAS this week, build your exclusion lists, and start a holdout on one campaign. You want an independent read in hand before Q4 spend lands.

Key Takeaways

• Inference passed training in 2026, which converts your AI budget from a project decision into a consumption line that scales with how much agentic work you run.

• Verification consumes most of that line. McKinsey puts 60% of agentic spend on response refinement. Price checking as its own budget item and meter it per step, because the vendor invoice will only show you a monthly total.

• Deploy agents first where instruments already exist, as Funck describes with the contact center, and extend only as you build new ones. Autonomy in unreadable environments buys risk you cannot price.

• Own at least one instrument outside the vendor relationship. Third-party agent scoring, per-agent identity logs, and incrementality tests all produce evidence that does not depend on the party selling you the system.

I spent a decade running companies where the diligence question was always the same: whose numbers are these, and what would it cost me to produce my own? In three acquisitions I paid for independent verification every time, and every time it looked expensive right up until it changed the price.

Marketing has arrived at that question. Gartner's forecast tells you the meter is running. September 1 tells you the clock is too. The instruments you install this month decide whether you spend Q4 explaining a platform's numbers or defending your own.