Human-in-the-Loop Marketing: Where Approval Gates Belong in an AI Workflow

Key points

• Gate by risk, not by job title: score a task on reach, undo cost and stakes, and approve before anything that reaches customers, spends money or cannot be taken back.
• In the matrix below, 11 of 17 tasks need approval before they ship, 3 a spot-check or review after, and 3 only a log. The list leans toward customer-facing work, so the counts describe the list and not marketing as a whole.
• A gate fires on a rule anyone can test, such as a budget above a set amount, and the same task can need approval at one size and a log at another.
• Track two rates: how often a person steps in and how often they reject what they are shown. Improvado says most teams aim for 2% to 10% intervention.
• All three sources are vendors, none measures results, and Kanso ran no test. The scoring is judgment, shown so you can change it.

What is human-in-the-loop marketing?

Knak says the idea comes from machine learning, where human oversight trains and evaluates a model, and that in marketing it applies wherever AI touches something a customer will see. Improvado frames it by decision point: humans stay in control where context, compliance or budget risk demands it, and machines handle the repetitive processing.

This page is about marketing workflows. It does not cover human-oversight rules in AI regulation or the data-labelling jobs that share the name, because the research behind it contains no primary source for either.

Why does AI marketing output need a human gate?

Because most teams still edit what AI produces before it ships, and the gap between using AI and trusting it is real. In its State of Marketing Production 2026 report, as Knak cites it, 70% of teams have deployed AI in marketing production, 29% call that use advanced and fully embedded in the workflow, and 88% still need moderate or substantial human editing of the output.

Knak calls that gap a governance problem, not a technology problem. Improvado reports a separate figure: about 70% of workplace AI users say AI is reliable only when paired with human review or oversight. The two 70% figures measure different things, and Improvado's page does not name its study.

Smartly adds the adoption side: roughly 60% of marketing teams report they have fully invested in AI for live campaign optimization, and 95% of marketers use generative AI for visual assets, by its own research. Its conclusion is that leading teams ask where human judgment matters most instead of how to remove it.

Read these as vendor claims. All three companies sell into this market, none of the pages reports a measured comparison of gated and ungated work, and statements such as Knak's that the approach can be faster in the long run are assertions.

Where do approval gates belong in an AI workflow?

Approval gates belong before any step that reaches customers, spends money or cannot be taken back, and nowhere else by default. Kanso's AI CMO capability map reached the same rule by combining two vendors' risk tiers, and Improvado states the test underneath it.

Improvado says to use a gate when the cost of an error exceeds the cost of human time, and to automate fully the low-stakes, high-volume tasks where an error is cheap and the fix is fast. It describes human-in-the-loop as the middle path between manual work that is too slow and automation that is too risky.

So the question for each task is what a wrong output costs and who finds out first. If the answer is a customer, a bank balance or a public thread, a person approves first. If the answer is an analyst on the same team, a log is enough.

cost of an error > cost of a person's time → approve before it ships

How was the risk-by-task matrix built?

Each task is scored from 1 to 3 on three criteria, and the three scores are added. The criteria come from the risk factors the sources name: Improvado lists budget, compliance, customer experience and executive reporting, and Knak points to anything a customer will see.

The total sets the gate. A total of 7 to 9 means approve before the output ships, 5 to 6 means a sample, spot-check or review after, and 3 to 4 means log it and look at the log in the weekly pass. One override applies: a 3 on undo cost forces approval whatever the total, which catches an irreversible internal action such as one scoring 1, 3 and 1. No row in the matrix below needs the override.

The 17 tasks are the examples the three sources give plus common channel tasks. They are not a sample of marketing work. Nothing was tested and no error rate was measured, and two careful people could score some rows a point apart, so treat the score as a way to argue about a task and change any number that your own evidence contradicts.

Which marketing tasks need approval before they ship?

In this matrix, 11 of 17 tasks need approval before they ship: anything sent to customers, published, reported upward or merged into a live site, and any spend above a threshold. Three tasks need a sample or review after, and three need only a log.

Two rows describe the same job at different sizes. A budget change inside a preset cap scores 4 and is logged, while a change above the threshold scores 7 and waits for a person, so the gate is a threshold on the task and not a label for the whole task.

Rows 6 and 11 show the other axis, volume. A batch of creative variations that are each easy to pause is sampled, while a single long article, which is public and costly to correct after indexing, is approved. The count of 11 approvals reflects a list weighted toward customer-facing work, which is the part of marketing where these risks concentrate.

The last row scores 9, and approval is not enough there. Kanso's AI CMO page keeps positioning and crisis statements in a human-only tier, so a person writes the statement and AI at most critiques it.

What kinds of approval gate can you use?

There are three gate types, matching the three score bands: approve before, sample or spot-check, and review after. Improvado separates human-in-the-loop, where approval comes before action, from human-on-the-loop, where the system acts and a person watches aggregate results and steps in when something looks wrong.

Improvado recommends the first for pre-launch checks such as budget validation and data quality gates, and the second for ongoing monitoring such as dashboard alerts and drift detection. It also names three working patterns: active learning, rule-based intervention and hybrid decision-making. The mapping in the table is our reading of how those patterns line up with the score bands.

How do you write the rule that triggers a gate?

Write the rule so anyone can test it against one item and get the same answer. Improvado says a rule must be explicit and testable and gives three examples: a campaign budget above $10,000, a tracking value that is not on the approved list, and a cost per lead up more than 50% week over week.

Those thresholds are the source's illustrations, and the right numbers depend on your budget and your tolerance for a bad week. Improvado also asks for three things around the rule: who gets the alert, what they see, and enough data to make the call in under two minutes.

• Trigger: the testable condition, such as a daily budget above the approved cap.
• Owner: one named person, not a channel.
• What they see: the item, the rule that fired and the numbers needed to decide.
• Default if nobody answers: the item waits and nothing ships. This field is our addition, not Improvado's.

Take Improvado's $10,000 example. Changes under it go ahead and are logged (row 2), and anything above it stops for the named owner (row 10).

How much work does a gate add, and how do you know it works?

A gate adds work only for the items it flags, and two rates show whether it flags the right ones: the intervention rate and the override rate. Improvado defines both and gives the worked figures in the table.

The arithmetic answers the speed question for that example. At 5%, 950 of 1,000 campaigns proceed with no human step (1,000 minus 50), and the 2% to 10% band means 20 to 100 items a month reach a reviewer at that volume.

Improvado reads a high override rate as a sign the model needs retraining. An 80% override could equally mean the trigger is too sensitive, so read the rejected items before changing anything. A rate near zero is ambiguous too: it can mean a well-tuned rule or a reviewer who approves by habit, which the next section covers.

Can a human gate turn into a rubber stamp?

Yes, a gate protects nothing when the reviewer cannot read what is put in front of them or approves from habit. The objection is fair when AI output grows faster than the people checking it, and none of the three sources measures reviewer quality, so what follows is analysis.

• Gate fewer steps. Limit approvals to the tasks scoring 7 or more, so each approval is rare enough to be read.
• Flag by rule. Let the trigger bring exceptions to the reviewer instead of asking them to scan everything.
• Show the deciding facts. Improvado's two-minute test is a good limit: if the call takes longer, the screen is missing something.
• Sample the approved items. Check some approvals after the fact, and look hard at any gate where nothing is ever edited or rejected.

Knak argues that the share of work that stays human is often the most valuable part of a workflow, and its 88% editing figure suggests review is doing real work in the cases it measures. That is an inference from one vendor's number, not proof that any given gate works.

How do you roll out approval gates in an existing workflow?

Roll them out one workflow at a time, in six steps that combine Knak's five-step build order with Improvado's three parts: trigger, notification and feedback. The sequence below is our synthesis, not either company's published method.

1. Map one workflow end to end, from brief to publish, and mark every point where output leaves the team (Knak's second step).
2. Score each step on reach, undo cost and stakes using the criteria above.
3. Gate every step that scores 7 or more. Write its trigger, its owner, what the owner sees and what happens if nobody answers.
4. Pilot on a narrow, easily reviewed task, not the whole calendar (Knak's third step).
5. Record every approval, edit and rejection, and feed the edits back into the brief and knowledge files the AI works from (Knak's fourth step and Improvado's feedback loop).
6. Review the rates at your next planning pass. Move a step down a band when its sample stays clean, and up a band after a miss.

Knak's first step, keeping personas, positioning and competitive documents current, belongs before all six, because the AI works from them. If a task is one you should hold back from AI entirely, Kanso's guide to when not to use AI in marketing covers the break-even test, and the weekly growth checklist is a place to run the review in step 6.

How do Kanso's agents handle approval?

Kanso's agents stop at a person: they draft, recommend or open a pull request, and you review, edit, post or merge. The table maps each live agent to a row of the matrix and states where it stops, as described on Kanso's own pages in October 2026.

None of the eight agents covers email sends, budget changes, translation, attribution models or executive reports, so those rows of the matrix need a gate in whatever tool you use for them. The Agent Hub lists every live agent and what each does.

What is documented and what is our analysis?

The documented items are claims the three vendor guides make, attributed to each, and the analysis items are Kanso's own reasoning built on them.

Documented, as published by the sources (read 19 October 2026)

• Knak cites its State of Marketing Production 2026 report for 70% of teams deploying AI in marketing production, 29% calling it advanced and fully embedded, and 88% needing moderate or substantial human editing.
• Improvado states that about 70% of workplace AI users say AI is reliable only with human review, and defines human-in-the-loop, human-on-the-loop and the three patterns.
• Improvado defines intervention rate and override rate, gives the 5% and 80% examples, and says most teams aim for 2% to 10% intervention.
• Improvado gives three example trigger rules: a budget above $10,000, a tracking value off the approved list, and cost per lead up more than 50% week over week.
• Smartly reports roughly 60% of teams fully invested in AI for live campaign optimization and 95% of marketers using generative AI for visual assets.
• Kanso's agent descriptions and limits are as published on its agent pages in October 2026.

Our analysis, not stated by any source

• The three criteria, the 1 to 3 scales, every score in the matrix, the score bands and the undo override.
• The counts of 11, 3 and 3 rows, which describe the 17 tasks chosen and not marketing as a whole.
• The mapping of Improvado's patterns to score bands, and the default-if-nobody-answers field.
• The readings of an 80% override rate and a near-zero one, and the four rubber-stamp safeguards.
• The six-step rollout, which combines Knak's and Improvado's steps.

Human-in-the-loop AI marketing: frequently asked questions

Sources

• What is Human-In-The-Loop and Why It Matters to Marketers, Knak, a marketing production platform vendor (no page date visible in the text read)
• Human-in-the-Loop AI for Marketing Teams | Guide 2026, Improvado, a marketing data platform vendor (no page date visible in the text read)
• Human-in-the-Loop AI: The New Marketing Playbook, Smartly, an advertising platform vendor (no page date visible in the text read)
• What an AI CMO Can and Can't Do: An Honest Capability Map, Kanso, 17 October 2026