When Will You Know Your Agent Has Drifted? Coherence at Velocity | October 20, 2026
When one of your agents stops producing value, or starts producing the wrong thing, how many days will pass before someone notices, and who has the authority to switch it off?
In my upcoming book, Coherence at Velocity, I describe the move from control through friction to control through observability. Friction-based governance adds approvals, reviews, and handoffs late in the process. Observability-based governance relies on thresholds, decision rights, logging, and escalation paths, so teams detect drift early enough to correct it before a customer, a service rep, or a regulator reports it first.
In Friday's edition I argued that every decision more than one AI system touches needs a named owner and written-down judgment. Analysts and authors at Gartner, Forrester, and HBR published on Monday about the next requirement: giving that owner the signals, and the advance agreement, to pause, fix, or retire an agent.
Treat Monday's reading as one instruction: write your stopping rules before you scale
Alicia Mullery of Gartner reports that CIOs who set criteria for stopping AI work in advance outperform their peers, and that few of them have done it. Lisa Gately of Forrester describes how marketing organizations govern the agents their own staff now build, using a tiered registry and three tests she calls risk, reach, and reliance. Charlie Dai, Leslie Joseph, and Indranil Bandyopadhyay of Forrester tell leaders to manage and monitor the context agents draw on as a production asset. Michelle Taite and Will Fernandez write in HBR that marketing organizations pay more and absorb more complexity before an AI transformation pays back.
Read together, these authors describe the work your shared-decision owners take on once the org chart changes. Register every agent that exists, agree on the conditions for stopping each one before it launches, monitor the inputs along with the outputs, and put the cost of that oversight in the budget.
Decide when you will stop an AI program before you approve it
Alicia Mullery, VP analyst at Gartner, presented new CIO survey results at Gartner IT Symposium/Xpo in Orlando on October 19. In Gartner's survey of 2,002 CIOs and technology executives, fielded from May through July 2026, 13 percent reported that their enterprise has realized significant or transformational value from AI. Mullery puts the success rate for individual use cases at about one in five, even as companies have launched use cases faster over the past two years. Her summary: "Volume is not value."
Gartner analysts call the stronger performers vanguards: CIOs who report better business outcomes along with high AI value as their CEOs perceive it. Organizations with guidelines for deciding when a pilot should not move to production are 2.3 times more likely to be vanguards. Organizations with guidelines for abandoning low-value AI already in production are 2.6 times more likely. Thirty-one percent of CIOs reported having the first practice, and 17 percent the second.
Skipping the retirement rule costs you more each quarter. Fifty-eight percent of CIOs report more pressure to deliver cost improvement from AI, and 51 percent expect AI to raise the total cost of ownership of their technology. Mullery points out that AI costs are probabilistic and set at run time, so you pay for a low-value agent every time it runs, whether or not anyone reads its output. Most marketing teams I work with keep a launch checklist and no retirement checklist, and their agent counts grow accordingly.
In the opening keynote the same morning, Daryl Plummer and Kristin Moyer, both distinguished VP analysts and Gartner Fellows, cited another Gartner survey of more than 2,000 CIOs and technology executives: 71 percent expect to deploy agentic AI within 12 months, and 21 percent have agentic AI governance capabilities in place. They told CIOs to set up what they call an AI Central Bank, a virtual control plane for AI decision-making across the enterprise, covering accountability, technology governance, economics, and shared intelligence. Moyer described the purpose: "It's not about centralization. It's about alignment, coordination and collaboration." Many companies already run an early version through an AI governance board that approves high-risk uses. Plummer and Moyer extend that job to managing AI economics across every program.
In Marketing Operations 3.0 I describe AI theater: experimentation visible enough to impress stakeholders and too loosely governed to matter operationally. I recommend a boundary between a laboratory, where agents can fail while teams learn, and a factory, where agents meet production standards, run under machine-readable governance, and carry a predefined kill switch. Use Mullery's first rule to decide which pilots leave the laboratory, and her second to decide which factory agents you retire.
It reminds me of what Jim Krueger, Chief Marketing Officer at Informatica, said when I interviewed him on The Agile Brand podcast: "We brought data to the table, we convinced a sales leader who was leading that area to kill those campaigns and, when we look back, it was like, wow, that was a great decision. We improved their pipeline by literally 30% by doing that."
Krueger's team had to win an argument to stop those campaigns. With stopping rules written in advance, the owner makes the call on a schedule, against a threshold everyone agreed to before anyone had a stake in the result. Writing a retirement condition into a launch plan takes the humility in Agile Brand Principle 7: acknowledge room for improvement before the results come in.
My take: For every AI program in your 2027 plan, write two lines before approval: the result that moves it from pilot to production, and the result that retires it. Seventeen percent of Mullery's respondents have the second line in place, so you will set the standard in most budget meetings.
Register every agent your marketers build, and scale oversight by risk, reach, and reliance
Lisa Gately, principal analyst at Forrester, published her account of what marketing leaders discussed at MAICON on October 19. She describes marketing organizations moving from AI literacy and personal productivity toward repeatable workflows and agents. In one organization, 76 percent of marketers earned an AI certification and the company still struggled to turn the training into working workflows, so leaders asked each team to pick one high-impact workflow and redesign it. Another organization ran builder bootcamps and team hackathons and moved the top three agents into production across marketing.
Gately tells leaders to let teams solve problems locally and standardize the foundations: approved messaging, brand guidelines, trusted knowledge sources, content quality requirements, templates, and evaluation practices. Once marketers start building, she lists the problems that follow, including duplicate agents, inconsistent quality, rising costs, and maintenance gaps when builders leave. One organization she describes uses three tiers. Individuals build personal productivity agents freely, team-wide agents require registration and an owner, and organization-wide agents receive central stewardship. Gately sets oversight by three factors: risk, meaning what goes wrong if the output is inaccurate or inappropriate; reach, meaning how broadly the outputs affect the organization and its audiences; and reliance, meaning how much other people and processes depend on the workflow. She uses the scores to set requirements for testing, approval, ownership, monitoring, and maintenance.
Take a hypothetical specialty retailer with $1.2 billion in revenue. After a quarter of hackathons, the marketing operations lead asks every team to register its agents and counts 48: 31 personal, 12 team-wide, and 5 organization-wide. The people who built four of the team agents have left the company, so nobody owns those four. Two teams built separate product-description agents, each drawing on a different version of the brand guidelines.
One organization-wide agent writes and sends promotional emails and pulls its exclusions from a promotions file. In August, merchandising excluded three premium brands from a 20 percent code, and nobody updated the file the agent reads. For 11 days customers redeemed the code on excluded items, about 170 orders a day at an average discount of $18, until a spike in service tickets prompted someone to check. The team lost about $33,700 in margin.
Score that agent on Gately's factors. Risk is high, because the agent commits the company to a price. Reach covers the full email list. Reliance runs through merchandising, service, and finance. Based on those scores, the team writes a stopping rule into the agent's charter: if redemptions on excluded SKUs pass 25 in a day, pause the send and alert the merchandising owner. With that rule in place, the team pauses the agent within the first day and loses under $3,100. Assign owners to the four orphaned agents this week or shut them off, because you still pay for every run of an agent nobody owns.
It reminds me of what Mark Abramowitz, Chief Marketing Officer at Dataiku, said when I interviewed him on The Agile Brand podcast: "Within a year, we will start seeing AI agents appear on organizational charts, treated not as software, but as a new form of labor. These agents will be 'hired' by business units, their performance will be reviewed, and they will be 'fired' if they don't deliver ROI."
The organization Gately describes applies Abramowitz's prediction at the team level by requiring registration and an owner for any agent a team depends on. In my AI Capability Framework, the Trust layer under Orchestration asks whether an agent did what it reports, and the Identity and Permissions layer asks on whose behalf and within what bounds it acts. Both questions get harder to answer as more marketers build. Use Gately's three factors to set how much evidence you require from each agent's owner, and for the highest-scoring agents keep a human in the loop at the point where the agent commits money or makes a claim.
My take: Run the registry before you run another hackathon. Count personal, team, and organization-wide agents, name an owner for every agent above the personal tier, and attach a stopping rule to each organization-wide agent.
Assign owners to the context your agents read, and monitor it
Charlie Dai and Leslie Joseph, both VP principal analysts at Forrester, and principal analyst Indranil Bandyopadhyay published their case for context engineering and ContextOps on October 19. They separate three terms: data describes reality, knowledge explains it, and context applies the relevant knowledge to a specific situation so an agent can reason and act. As agents move from answering questions to running business processes, the authors treat context as a production dependency. They define ContextOps as the operational discipline for managing, governing, monitoring, and improving context assets and the flows between them, and they tell leaders to build shared context services and common governance in place of separate retrieval, memory, and tool integrations for each project.
The retailer in the example lost margin to a context problem. The agent applied its rules correctly to a file that described July. In Chapter 11 of Coherence at Velocity I call this policy lag: one system still applies yesterday's logic while another has moved on, and the customer finds the gap first. Gately makes the same recommendation from the content side. She tells leaders to assign owners to product information, approved claims, and customer insights, because every agent that reads outdated source material repeats it.
MarketingProfs' October 16 AI Update relays figures from Kore.ai's 2026 Agent Productivity Index: 79 percent of enterprises have reversed an action an AI agent took, and 70 percent have encountered failures they could not trace. Kore.ai sells an agent platform, so weigh the source. In the engagements I see, teams that cannot trace a failure usually never logged which version of a file the agent read.
My take: List the five files or feeds your organization-wide agents read most often. Give each one a named owner, a review date, and a change log that records which version every agent used.
Budget for oversight costs to rise before the savings arrive
Michelle Taite, most recently CMO of Intuit Mailchimp, and Will Fernandez, cofounder of the AI-native marketing consultancy Defyner, published Your AI Marketing Transformation Will Get Harder Before It Gets Easier in HBR on October 19. It follows their earlier HBR piece, which describes an agentic marketing organization built on a machine-readable brand code, execution agents, and an orchestration layer that escalates decisions requiring human judgment. In the new piece they argue that marketing organizations often see costs and complexity rise before they see AI's speed and efficiency gains.
They describe three stages: codifying individual workstreams, connecting skills and workflows so the organization can reuse them, and compounding returns as information flows across the marketing system. They tell leaders to codify the judgment, rules, and expertise experienced marketers apply by instinct. The payoff they describe is the ability to scale demand without adding headcount, outside spend, or coordination for each new launch, segment, channel, or market.
Put their first stage next to Gately's registry and the Forrester ContextOps definition, and you can name the costs that arrive early: the registry, the owners, the charters, the monitoring, and the context maintenance. In Mullery's survey, 51 percent of CIOs expect AI to raise total cost of ownership, so your CFO already expects the increase. Show it as named line items in the 2027 plan, so finance approves the oversight before the first incident.
In my book Marketing Operations 3.0 I recommend a 90-day review rhythm for each AI system, frequent enough to catch weak design, hidden cost, and drift before they harden into habit. Hold your stopping rules to that calendar, and have each owner check every threshold at least once a quarter.
My take: Ask your finance partner to model oversight costs for your top five agents through 2027, and attach each agent's stopping rule to its budget line.
Take these actions before you approve another agent
Register every agent. Sort them into personal, team-wide, and organization-wide, and name an owner for everything above the personal tier.
Write two stopping rules per program. Define the result that moves a pilot into production and the result that retires a production agent.
Score each agent on risk, reach, and reliance. Set testing, approval, and monitoring requirements from the scores.
Assign owners to context. Give every file and feed your agents read an owner, a review date, and a version log.
Budget for oversight. Show registry, monitoring, and context maintenance as line items in the 2027 plan.
Review every quarter. Check each threshold on a 90-day calendar and retire the agents that miss.
Answer the opening question with a threshold and a name
I have been a CEO twice, and in both companies I approved more launches than shutdowns. The shutdowns I put off longest were the ones nobody had defined in advance, because every project had a sponsor who could explain why one more quarter would turn it around.
In my upcoming book, Coherence at Velocit, I write that the human role changes as an enterprise automates. Someone still has to define behavioral boundaries, monitor drift, interpret edge cases, and decide when a system should escalate. Mullery describes that job from the CIO's office, and Gately describes it from inside marketing.
Answer the opening question in writing for every organization-wide agent you run. Record the threshold that triggers a pause, the number of days you expect between drift and detection, and the name of the person who switches the agent off. Send any agent missing one of the three back to the laboratory until it has all three.
Featured this cycle:
Gartner Survey Finds Only 13% of CIOs Say Their Enterprise Has Realized Significant or Transformational Value from AI. Alicia Mullery, Gartner, October 19, 2026
Marketing's Next AI Challenge: Scale What Works Without Losing Control. Lisa Gately, Forrester, October 19, 2026
Context Engineering And ContextOps: The Missing Foundation For Scaling Agentic AI. Charlie Dai, Leslie Joseph, and Indranil Bandyopadhyay, Forrester, October 19, 2026
Your AI Marketing Transformation Will Get Harder Before It Gets Easier. Michelle Taite and Will Fernandez, Harvard Business Review, October 19, 2026
Also referenced: Gartner Says AI Risk is Rising Faster Than AI Value, Daryl Plummer and Kristin Moyer, Gartner, October 19, 2026; AI Update, October 16, 2026, MarketingProfs, October 16, 2026.