Your Best AI Training Data Is What Reviewers Sent Back. Change Futurist | October 6, 2026
Someone on your team rejected a plausible AI draft last week. Ask where that rejection went.
In most marketing organizations it went nowhere. A senior reviewer fixed the spec, rewrote the claim, and sent the asset on. The reason for the correction stayed in a tracked change or a chat thread.
On Friday I asked what the hours AI saved your team last quarter paid for. Four pieces published since then give me an answer: spend the hours on checking, and keep a record of every check.
Log every override your team makes
Christian Catalini published AI Is Making Verification the Bottleneck for Companies in Harvard Business Review on October 2. Managers in a hierarchy do two jobs, he writes. They route information, and they verify it: they decide what a piece of information means, which outputs the firm can trust, and which are worth producing at all. Generating text, media, and code now costs very little. Checking still takes a person who knows the business, so checking is where a firm earns its margin.
Catalini opens with a controller who catches a revenue figure that counted deferred bookings. He calls that kind of rejection "the most valuable data your firm produces."
He names two assets a firm needs to verify well: ground truth that only the firm holds, and people with the expertise to apply it. Many firms, he warns, are adopting AI in ways that wear down both.
Marketing teams produce this data every day. A brand reviewer sends back a headline. A product marketer corrects a spec. A legal reviewer strikes a claim. Each one is a labeled example of what your company considers wrong, and almost nobody stores it.
In August I wrote about pricing the correction work before you price the license. Catalini takes the argument a step further. The correction is a cost and it is also an asset, and you keep the asset only when someone writes it down.
This sits in the Trust and Verification layer of the AI Capability Framework. Under augmentation, trust means calibration: the person in the authoring seat knows when to accept the contribution and when to override it. The framework's mechanism rule says you pay for a more capable mechanism twice, once in compute and once in verification. An override log tells you what the second payment bought.
Commentary: Catalini gives marketing operations leaders an economic reason to treat review as production work with its own output, and that output is the record of what reviewers rejected.
Train reviewers on your own corrections
Molly Innes published Is marketing facing a talent pipeline problem? The data says so in Marketing Week on October 2. She pulls together three datasets.
In Marketing Week's 2026 Career and Salary Survey, 66.5% of 2,350 respondents identified an AI skills gap in their team, and the figure reaches 73.5% among CMOs and marketing directors. Ingenuity+ and Pearlfinders counted at least 2,595 senior marketing appointments in the 12 months to September 22, and 62% went to external candidates. In September that share hit 73%. Gartner's survey of more than 1,300 marketing leaders found 18% have already eliminated early-stage roles because of automation, while 16% have created new ones.
Read those numbers against Catalini's second asset. A reviewer learns to verify by producing work and having someone senior correct it. Companies are cutting the roles where that happens, then hiring senior judgment from outside. The external hire arrives with judgment formed on another company's products, customers, and legal history. That person needs your ground truth, and most teams have no document to hand over.
Last week I wrote about judgment leaving the organization through eliminated junior roles and vendor-built agents. The override log is a practical response. Three months of logged corrections give a new associate several hundred worked cases of what your senior people reject and why. That is a reviewer's onboarding curriculum, and it costs you the discipline of filling it in.
I hold one principle fixed in the framework I am building on top of Coherence at Velocity: humans stay accountable for direction even when a team fully automates the analysis. Accountability takes people who can tell a right answer from a plausible one. You now have to grow those people on purpose.
Commentary: Innes documents a market where companies hire senior judgment from outside in 62% of cases. Each of those hires starts without the company-specific knowledge that makes a review worth anything.
Give buyers something they can check
Emily Manock's weekly stats roundup in Marketing Week, published October 5, carries three findings that belong next to each other.
IDHL researchers analyzed 1.5 million AI search sessions across 175 websites and 20 industries. Revenue attributable to AI search rose 321% over thirteen months, and transactions rose 553%. Organic search grew 28% over the past year. Criteo found that 51% of UK shoppers prefer to browse everything themselves over receiving personalised picks, the highest share among the six markets surveyed. LinkedIn found that 85% of UK B2B marketers believe Gen Z buyers are more likely to consider a brand that someone in their professional community has validated.
Your buyers are verifying too. A shopper who arrives from an AI answer has already seen your claims set against reviews, specifications, and third-party sources. Half of UK shoppers want to inspect the range before they accept a recommendation. Younger B2B buyers ask a peer.
Each of those behaviors rewards a company whose public information is accurate and specific. The spec error your reviewer catches internally is the same error a buyer reads in an AI answer when nobody catches it.
It reminds me of what Mark Nardone, CMO at PAN, said when I interviewed him on The Agile Brand podcast: "Gen AI has neutralized volume as a competitive advantage. It wins now with depth and proof and clarity and expertise and tight alignment across channel."
Proof starts as an internal habit. The fifth of the Agile Brand principles tells you to focus on the customer relationship across transactions. A buyer who checks your claim and finds that it holds has a reason to return.
Commentary: Manock's roundup puts a 321% revenue figure next to a 51% preference for self-directed browsing. Together they describe a buyer who trusts what they can confirm, which makes accuracy a growth input.
Put the check where the work happens
Sarah Jensen Clayton, Michael Welch, and Khoi Tu of Korn Ferry published There's No Such Thing as an AI-Ready Culture in Harvard Business Review on October 5. Organizations have to perform and transform at the same time, they write, and most run better at performing. Speed, agility, and action orientation rank among the attributes least characteristic of organizations today. Their prescription is a culture of adaptability, which they break into five recurring patterns of behavior. Their opening line: "Most companies are transforming for AI faster than their cultures can keep up."
Behavior is the word to hold on to. A culture program gives you posters. A logged override is a behavior a manager can observe, count, and coach.
It reminds me of what Nayaki Nayyar, CEO at Siteimprove, said when I interviewed her on The Agile Brand podcast: "We live in an AI world with infinite content, so Shift Left—common in DevOps—means shifting quality and compliance checks early in the content creation process to detect and remediate issues right there instead of after publication."
Apply that to verification. A check at the end of the workflow catches errors after the team has built on them. A check at the drafting step, run against live product data and an approved claims list, catches them while the fix costs minutes.
The first and third Agile Brand principles ask you to build for change and to make adaptivity part of how the team operates. An override log does both at small scale. Each time the log shows a pattern, the team changes the prompt, the data source, or the template.
Commentary: Clayton, Welch, and Tu move the culture conversation from readiness to repeatable behavior. Logging a correction is about the smallest repeatable behavior a marketing team can adopt this month.
Run the numbers on one content workflow
This example is illustrative. A $700M industrial distributor runs a marketing team of 34. The team produces 1,400 AI-drafted assets a month: product descriptions, lifecycle emails, and comparison pages. Six senior reviewers approve them.
The reviewers reject or materially edit 22% of drafts, which comes to 308 overrides a month. Each override takes about 25 minutes, so the team spends roughly 128 senior hours a month on corrections. Nobody records why.
The marketing operations lead starts a log with five fields: asset type, what was wrong, error category, the fix, and the source the reviewer checked against. Reviewers fill it in from a dropdown in under a minute.
After 90 days the log holds 924 entries, and two categories account for most of them. Spec mismatches make up 41%, because the drafting step pulls from a product data export that someone last refreshed in the spring. Claims without evidence make up 23%.
The repairs are small. The team points the drafting step at the live product feed and adds a rule-based check that compares every spec in a draft to the product record. That check uses the lightest mechanism that clears the accuracy bar, and it costs almost nothing to run. The team also builds an approved claims library from the 213 logged claim corrections.
The override rate falls from 22% to 9%. Monthly overrides drop from 308 to 126, and senior correction time drops from 128 hours to about 53. Two associates now train on the log, with a senior reviewer grading their calls. The corrected specs also flow to the product pages that answer engines read.
Seventy-five senior hours a month came back, and the team bought nothing.
Take these four steps this month
Start an override log on one workflow. Pick your highest-volume AI-assisted workflow and capture five fields per correction. Keep entry time under a minute or reviewers will skip it.
Review the log every month and fix the top category. Change the data source, the prompt, or the template. Reach for a rule-based check first and a more capable mechanism only when the rule fails.
Make the log your reviewer curriculum. Give new hires and associates the logged cases, have them make the call, and have a senior reviewer grade it. Do the same for senior people you hire from outside.
Audit what buyers can verify. Check that the specs, prices, and claims on your public pages match the product record, and that each claim has a source a buyer or an answer engine can find.
In my advisory work with Fortune 500 marketing teams, I ask early to see the last ten things a senior reviewer sent back. Most teams cannot produce the list. The corrections live in the reviewers' heads.
Find last week's rejection. Write it down, and make it the first entry in the log.
Featured this cycle:
AI Is Making Verification the Bottleneck for Companies. Christian Catalini, Harvard Business Review, October 2, 2026
Is marketing facing a talent pipeline problem? The data says so. Molly Innes, Marketing Week, October 2, 2026
AI search revenue, Reddit, Gen Z buyers: 5 interesting stats to start your week. Emily Manock, Marketing Week, October 5, 2026
There's No Such Thing as an AI-Ready Culture. Sarah Jensen Clayton, Michael Welch, and Khoi Tu, Harvard Business Review, October 5, 2026