Who gets to approve an agent's output? Martech Futurist | September 16, 2026

Written By Greg Kihlstrom

Name the person who can approve an agent's output and ship it without asking anyone else. If that takes you longer than five seconds, you have found the constraint on your AI program.

Four research teams published inside 48 hours, from four different vantage points, and they landed on the same variable. Masha Shunko and Serguei Netessine argue in HBR that companies automate tasks while leaving the workflow — and the authority inside it — untouched. Caroline Hewings at Gartner predicts that organizations treating work redesign as a standing capability will be twice as likely to sustain AI transformation by 2028. Julian De Freitas and Andrew Yan describe a discovery environment where an AI engine decides which brands a buyer hears about. Charlotte Rogers reports that B2B marketers cannot secure brand investment because the executives holding the purse understand neither the asset nor the measurement.

Authority is the thread. Who decides, at what point, with what evidence, and who can overrule it.

I have sat in enough operating-model reviews to know how this surfaces. The team has agents drafting, scoring, routing, and summarizing. Every output still queues behind the same three humans, because nobody ever wrote down which decisions those humans actually own. Velocity went up. Throughput did not.

Redesign the workflow before you license another agent

Shunko, of the University of Washington's Foster School, and Netessine, of Wharton, published Stop Automating Old Processes. Design New Ones Instead. on September 14. Their diagnosis rests on three numbers that belong in your next budget conversation: BCG's January 2026 survey has companies doubling AI spend from 0.8% to 1.7% of revenue, PwC's 2026 Global CEO Survey finds 12% of CEOs reporting both revenue and cost benefits, and McKinsey QuantumBlack's April analysis puts 60% of organizations at no enterprise-wide EBIT impact while 80% run generative AI in at least one function.

Their framework turns on three decisions. Identify what to redesign by naming the outcome a customer values and the constraint AI removes. Establish where human authority stays, using four modes — Assist, Approve, Audit, Automate — chosen by the stakes of an error and whether you can reverse it. Then design for failure: how the system recognizes uncertainty, escalates, reverses, and learns from the correction.

Those four modes are the most useful artifact published this week. Take any workflow with fifteen steps in it, assign each step one of the four, and you will find two things immediately — steps sitting in Approve that nobody can reverse anyway, and steps sitting in Automate that carry brand or compliance exposure. Both are misassignments of authority, and both are free to fix.

It reminds me of what Ben Schein, Chief AI and Analytics Officer at Domo, said when I interviewed him on The Agile Brand podcast: "Everyone knows human in the loop. I would add two more positions: human in the lead and human in the slop."

Human in the slop is what you get when you automate a step whose output nobody owns.

Commentary: Shunko and Netessine supply the missing verb in most AI roadmaps — assign. Task automation without an authority map produces faster work inside the same bottleneck.

Put work redesign on a permanent cadence

Hewings, a Sr. Director Analyst in Gartner's HR practice, published a prediction on September 15: by 2028, organizations that establish continuous work redesign as a core capability will be twice as likely to sustain AI-driven work transformation. Her framing: "Due to AI, the pace of organizational change is increasingly misaligned with the speed at which work is changing."

The mechanism she describes — dynamic redesign — covers workflows, decision rights, and talent deployment, adapted on a repeating cycle rather than a transformation calendar. It requires continuous visibility into how work actually runs, through process intelligence and task mining, plus a talent strategy flexible enough to move people as the work moves.

Pair this with the HBR framework and you get something operational. Shunko and Netessine tell you how to assign authority once. Hewings tells you that the assignment expires. Model capability shifts every quarter; a step that belonged in Approve in March belongs in Audit by September, and the only way to catch that is a standing review with a named owner and a date.

This is the discipline I keep returning to in the marketing operations work behind Marketing Operations 3.0. Humans own goals, constraints, judgment, and ethical boundaries. Agents determine and execute the path inside those limits. The contract between the two is the deliverable, and it needs a revision history.

Commentary: Hewings puts a measurable outcome on something most operators treat as hygiene — the twice-as-likely figure gives marketing leaders a number to take into a conversation about funding a redesign function rather than a redesign project.

Assume an engine briefs your buyer before you do

On September 15, HBR's Cold Call released How GEO Is Changing the Role of Brand Manager, with Harvard Business School's Julian De Freitas and AthenaHQ co-founder Andrew Yan, hosted by Brian Kenny. De Freitas frames the shift plainly: AI has an opinion about your brand, and it associates specific attributes with you. Yan describes the commercial consequence — a widening share of research happening on platforms brands neither own nor observe, and a monitoring discipline built around share of voice, factual accuracy, and sentiment across ChatGPT, Claude, and Gemini.

Yan's term for where this ends up is semi-autonomous marketing: humans making the high-leverage calls, agents running the routine ones. Which returns you to the four modes. Generative engine optimization is a workflow like any other, and most teams currently run it in the least defensible configuration — fully automated content production feeding a channel where a factual error about pricing or availability propagates into thousands of answers before anyone reads one.

It reminds me of what Imri Marcus, CEO and Co-Founder of Brandlight, said when I interviewed him on The Agile Brand podcast: "It's the first time in history where as a brand you can add a piece of content that no person ever sees, but the AI engine saw it and it's now completely affecting entire customer journeys. The measurement really shifts to influence — how often are you included, when you are included, with what type of sentiment, how are you described. How accurate is the model about you?"

Accuracy monitoring belongs in Audit, weekly, with a named owner. Publishing belongs in Approve until your error rate earns something looser.

Commentary: De Freitas and Yan describe a decision that has already moved outside the company — an engine now performs the consideration-set filtering a brand manager used to influence directly, and the response runs through instrumentation across the platforms where that filtering happens.

Settle the brand case with your CFO before you touch anything else

Rogers reported on September 14 that B2B marketers are losing the internal argument for brand. Marketing Week's State of Brand in B2B survey, drawing on 300 respondents, finds 44.9% who find it difficult or very difficult to secure brand investment, 38.9% who say better CEO and CFO understanding of brand building would ease it, and 47.7% reporting brand building is not a budget priority at all. Two figures point at the same repair: 58.3% say the conversation would improve if their business valued marketing beyond short-term results, and 55.1% want access to more robust effectiveness measurement.

Read that alongside the GEO discussion and the stakes sharpen. Brand attributes are now training signal. The consistency of how you describe your products across owned content, third-party coverage, and structured data determines what a model says about you when a buyer asks. An organization that defunds brand for four quarters is degrading an input to machine-mediated discovery, and the effect surfaces with a lag long enough that nobody connects it to the budget decision that caused it.

A CFO will fund that argument. A CFO will not fund awareness in the abstract.

Commentary: Rogers documents an evidence gap — over half the sample is asking for measurement they do not have, which puts effectiveness instrumentation ahead of narrative work on the path to funded brand investment.

Run the math on a workflow you recognize

A $400M B2B software company runs lifecycle content with nine people. Output is roughly 180 assets a quarter across email, web, product pages, and partner collateral. Last year the team added four agents: draft generation, channel adaptation, metadata and schema tagging, and a competitive summarizer. Cycle time on a first draft dropped from nine days to two. Quarterly output moved from 180 to 195.

Fifteen assets. That is the entire return on four agents, because every asset still routes through one director for brand review and one legal reviewer with a three-day queue, and both queues grew.

Now map the authority. Draft generation sits in Assist — a human owns the output and always did. Channel adaptation moves to Automate for email and web, where an error is visible within an hour and reversible with a resend. Metadata and schema tagging runs on deterministic rules rather than a model, because the accuracy bar is exact-match and the lightest mechanism that clears the bar is the right one. Brand review splits: templated derivative assets move to Audit with a weekly 10% sample, while net-new positioning claims stay in Approve. Legal review stays in Approve for anything touching pricing, security, or regulated claims, and moves to Audit for everything else.

The constraint moves. Review capacity that previously absorbed 195 assets now absorbs roughly 400, because only net-new claims and regulated content consume the two scarce humans. Add the GEO instrumentation on top — accuracy monitoring across three engines, owned by the same director in Audit mode — and the schema-tagged product pages start earning inclusion in answers the team can now observe.

Same nine people. Same four agents. The difference is a one-page document naming who decides what.

Do these four things before October

Map authority on your three highest-volume workflows. Every step gets Assist, Approve, Audit, or Automate. Rate each by stakes and reversibility, and expect to find steps sitting in the wrong mode in both directions.

Give the map an expiration date. Put a quarterly review on the calendar with a named owner. Hewings' finding is about cadence, and the teams that skip the cadence spend next year operating a map built for last year's model capability.

Instrument brand accuracy across three engines. Track inclusion, sentiment, and factual correctness monthly. This is the effectiveness measurement 55.1% of B2B marketers say they lack, and it produces the kind of evidence that survives a CFO conversation.

Take the authority map into your next budget meeting. Fifteen extra assets from four agents is a story about queues, and a CFO who sees the queue diagram will fund the redesign that clears it.

Name the person who can approve and ship. Then write it down, review it in ninety days, and stop buying capability you have not given anyone permission to use.

Featured insights:

Greg Kihlstrom
Next

The distance between perceiving and deciding. Change Futurist | September 14, 2026