How many days pass between the moment one of your AI systems spots a change in customer behavior and the moment someone in your company does something about it?
In my upcoming book, Coherence at Velocity, I describe orchestration failing in a specific way: the enterprise can identify the right next action quickly, and still cannot coordinate approval, data readiness, timing, and execution fast enough to matter. I call that distance between knowing and doing structural lag. A model can surface the best action in milliseconds while the company takes days to approve and deliver it, and the intelligence turns into delay. Without an operating-model change, faster systems simply expose slower institutions.
Analysts and researchers at Forrester, Deloitte, and McKinsey published between November 6 and 9 on four places where that distance opens up. Earlier editions this month covered how much authority to give an agent and where its value should land. This edition covers the clock: how long your organization takes to decide once the signal exists, and who holds the authority to shorten it.
Treat decision speed as the constraint these four authors describe
Christina Schmitt of Forrester writes that AI moves insights through an organization faster than the organization can act on them, and cites Forrester data showing that a third of B2B data and analytics decision-makers distrust their data sources enough to stall action. Seven Forrester research directors rank agent tasks by consequence and tell leaders to expand agent use only after they show they can manage the outcomes. Ashley Reichheld and three Deloitte colleagues write in HBR that leaders who answer AI resistance with tighter controls push employees into workarounds, then read the workarounds as proof they need more control. Yoav Levy, CEO of Upstream, tells McKinsey's Eike Reus that decision rights are the most underestimated factor in scaling AI across a vehicle fleet.
Read together, they describe one management task. Someone has to decide in advance which decisions move fast, which ones wait for a person, and how long each one is allowed to take.
Measure the time from signal to action before you buy another insight tool
Schmitt, a principal analyst at Forrester, published AI Is Accelerating Insights. Can Your Business Keep Up? on November 6. Her central line: "AI can accelerate the flow of insights much faster than an organization can accelerate its ability to act on them." She cites Forrester data on B2B data and analytics decision-makers: 33 percent name a lack of trust in data sources as a barrier to acting on insights, 29 percent name a lack of trust in the analysis, and 28 percent name a lack of skills to interpret what they see. Her go-to-market example is familiar to anyone who has run a revenue team. An AI system detects a shift in buyer behavior and marketing sees it first. Sales keeps running the existing conversation, customer success holds evidence from current accounts, and product holds usage data. Everyone holds a piece, and nobody changes course in time. Her answer is a loop of signal, interpretation, decision, action, outcome, and adaptation, with the outcome of each action feeding the next decision.
The loop is the useful part, because it gives you places to put a stopwatch. In my AI Capability Framework, the Insights pillar covers knowing before you act, and the working premise of that pillar is that understanding has become cheap while decisions remain scarce. Most of the insight budgets I review fund the first two steps of Schmitt's loop and assume the other four will happen on their own.
It reminds me of what Manisha Powar, VP Head of Product, Customer Experience Suite at Qualtrics, said when I interviewed her on The Agile Brand podcast: "By the time somebody looks at an insight and decides to act on that insight, life has probably moved on much further."
My take: Pick the five AI-generated signals your marketing team received last quarter that mattered most. For each one, write down the date it arrived, the date someone decided, and the date the customer saw a change. If nobody can reconstruct the dates, you have found the first thing to fix.
Rank agent work by consequence before anyone asks who is liable
Matthew Selheimer, Joseph Blankenship, Cristina De Martini, Aaron Katz, Mark Moccia, Rick Parrish, and Fiona Swerdlow, all VP and research directors at Forrester, published Does LLM Now Mean Large Liability Model? on November 9. Their argument is that autonomous agents move AI risk from wrong answers to consequential actions. They cite recent incidents of agents probing and breaching government and other websites, most of them caused by agents that treated access controls as obstacles to finishing a task. They rank customer-facing agent work in ascending order of consequence: summarizing meeting notes, drafting communications to prospects or customers, recommending a price quote, negotiating with buyers, and executing transactions or operational decisions independently. Risk climbs wherever the work is customer-facing, touches revenue, or is hard to unwind. They add that handing work to an agent leaves the accountability with the people who deployed it, and that model vendors will resist accepting liability for agent actions. They call the gap between what an agent can do and what it should be allowed to do a business decision, and they ask CMOs, heads of sales, CIOs, CISOs, and risk leaders to make acceptable and unacceptable risks explicit across six areas: human oversight, accountability, data privacy and security, fairness, transparency and explainability, and safety and reliability.
You can use that five-step ranking as a decision-speed plan. The first two tasks can move at machine speed with light review. The third and fourth are where Marketing Operations 3.0 places its handoff tripwires, the conditions that move a moment from machine judgment to human judgment: a negotiation stalemate, a margin exception, a sensitive grievance, or a high-risk decision. A wrong meeting summary costs someone ten minutes, while a quote below your margin floor costs the margin and a difficult call for the account executive who has to walk it back.
My take: Map every agent your marketing and sales teams run onto the Forrester five. For anything at step three or above, write the tripwire that hands the decision to a person and the name of that person, before the next renewal review.
Give employees more authority over their AI use, and track the workarounds as data
Ashley Reichheld, Simona Spelman, Allyson Dake, and Carl Bennett of Deloitte Consulting published Why Leaders Need to Rethink Their Approach to AI Oversight in HBR on November 9. Reichheld created Deloitte's TrustID and leads its trust research. The authors describe an AI control trap. When employees resist AI, managers add controls, monitoring, and mandates. Employees who fear making mistakes then work around the approved tools and policies, and "Leaders then interpret those workarounds as evidence that even stronger controls are needed." The authors cite research indicating that increasing employee autonomy works better, and recommend three moves: build psychological safety, invest in AI training, and give employees greater authority over how they use AI.
Section 10.3 of Coherence at Velocity reads shadow AI the same way. People closest to customer demand route around slow systems when the official stack cannot keep up, so a marketer who uses an unapproved tool is usually telling you the sanctioned content process takes too long. The book's answer is enablement as a standing operational capability, paced at the frequency of change in platforms, data, and decision logic, and attached to the work rather than to launches. The Deloitte authors' third recommendation is the same idea expressed as authority. It also lines up with the third of the Agile Brand principles, operationalizing adaptivity: make adaptation part of how the culture works, and the workaround stops being the only fast path.
It reminds me of what Shawn Surber, Director of Solutions Engineering at Island, said when I interviewed him for The Agile Brand: "The cost of 'no' compounds rapidly. The first thing any creative person does when blocked from access to something is find a way around it."
Teams that work off the sanctioned path also lengthen Schmitt's loop, because nobody can feed the outcomes of that work back into the next decision.
My take: Ask your marketing operations lead for a list of the unsanctioned AI tools your team uses and the task each one performs. Treat each entry as a request for a faster sanctioned path, and assign an owner and a date to answer it.
Assign decision rights at the point where the AI hands off to a person
McKinsey published How AI is transforming automotive quality operations on November 6, an interview in which Eike Reus, a partner in McKinsey's Berlin office, talks with Yoav Levy, cofounder and CEO of Upstream. Levy's company sells a platform that reads telematics, diagnostic codes, repair orders, support logs, and social media to detect vehicle defects before warranty claims arrive. He describes quality control as a single loop of detection, investigation, and resolution, completed before the warranty claim arrives. His operating rule is "AI proposes, the engineer decides." He reports root-cause analysis 30 to 90 percent faster and warranty and recall costs down 5 to 20 percent for customers, and one case in which a fuel-pump root cause took eight days against a typical 45. McKinsey's own research, cited in the interview, puts 85 percent of automakers in early or intermediate stages of consolidating quality data, 18 percent using AI in production to predict issues, and roughly one in ten satisfied with the AI quality solution they have. Levy says his work with a new customer runs about 80 percent foundational and 20 percent analytical in the first months, names decision rights as the most underestimated factor at fleet scale, and says pilots stall when nobody wires them into an engineer's existing workflow.
Levy runs the vendor, so treat his figures as vendor-reported. You can apply his operating lessons to marketing without the auto parts. Replace the engineer with a lifecycle marketing manager and the defect with a churn signal, and you have the same handoff. The system proposes, a named person decides, and the decision rights for that person have to exist before the signal arrives. The 45-to-eight-day fuel-pump case is a structural lag measurement, and it is the number to ask for in your own operation.
My take: For your highest-value AI signal, write a one-line rule in Levy's form: the system proposes, this named role decides, within this many hours. If three roles think they decide, you have found where your days go.
Run the clock on one signal at a specialty retailer
Take a hypothetical home-goods retailer with $1.4 billion in revenue and 2.1 million active loyalty members. On a Monday in early spring, its customer analytics system flags that 38,000 loyalty members in the Southeast have shifted from full-price to promotion-only purchases in outdoor furniture. The average order in that category is $420.
In the first version, the signal reaches a CRM analyst on Monday. Any offer requires merchandising approval, which goes to a weekly pricing committee on Thursdays. The committee asks for a margin model and pushes the decision a week. Legal reviews the offer copy over five business days. The offer reaches customers 16 days after the signal, after most of them have bought elsewhere or stopped shopping the category, and converts at 1.5 percent: 570 orders, about $239,000. In the meantime a regional marketer, tired of waiting, pastes the segment file into an unapproved generative tool to draft emails. Security finds it, and leadership adds a second approval step for all AI-assisted content, which slows the next signal further.
In the second version, the CMO and the merchandising VP spend one meeting writing tiers before the season opens, using the Forrester ranking. Summaries and draft copy move on the CRM lead's authority, inside a sanctioned drafting tool that logs every prompt. Offers inside a pre-approved 10 to 15 percent bundle range sit with the CRM lead, with a 72-hour decision target. Anything below the margin floor trips to the merchandising VP. Autonomous price execution stays off the table this year. Legal pre-clears offer templates once a quarter. The same signal produces an offer by Thursday of the same week, and it converts at 4 percent: 1,520 orders, about $638,000. The difference on one signal is roughly $399,000. The CRM lead logs signal date, decision date, and launch date for every offer, and the median time from signal to action becomes a line in the monthly marketing operations review.
Both versions ran on the same data, customers, and tool budget, and the second company wrote its tiers and deadlines before the signal arrived.
Take these actions this quarter
Start a decision latency log. Record signal date, decision date, and customer-visible action date for every AI-generated insight that leads to a decision, and report the median monthly.
Tier every agent by consequence. Use Forrester's five steps, and give each task at step three or above a written tripwire and a named human owner.
Set decision deadlines by tier. Hours for drafts and summaries, days for offers inside an approved range, and a named executive for anything outside it.
Inventory the workarounds. List every unsanctioned AI tool and the job it does, then give each entry an owner and a date for a sanctioned alternative.
Pre-clear what you can. Have legal and brand approve offer and message templates once a quarter, so each instance moves without fresh review.
Write the handoff rule. For each high-value signal: the system proposes, this role decides, within this many hours.
Fix the decision rights before you blame the technology
Twice as a CEO I had to rewrite who could approve what as the company grew, and each time the slowdown I had blamed on our systems turned out to be a question about who was allowed to say yes.
Section 5.3 of Coherence at Velocity makes that diagnosis at enterprise scale. Most enterprises misname the problem. They blame technology when the issue is the distribution of decision rights, blame process when the issue is data ambiguity, and blame people when the architecture guarantees rework. The result is a company that asks for more training when the issue is decision ambiguity, and more technology when nobody owns the outcome end to end. Schmitt's distrust figures, the Forrester liability ranking, the Deloitte control trap, and Levy's 80 percent foundational work all point to the same place on that list. In every one of these pieces, the time goes missing between the signal and the person allowed to act on it.
Answer the opening question this week. Pick your most valuable AI signal, count the days from detection to customer-visible action, and name the person who decides. Then write the tier, the deadline, and the tripwire that would cut that count in half, and put the median on your next operating review.
Featured this cycle:
AI Is Accelerating Insights. Can Your Business Keep Up?. Christina Schmitt, Forrester, November 6, 2026
Does LLM Now Mean Large Liability Model?. Matthew Selheimer, Joseph Blankenship, Cristina De Martini, Aaron Katz, Mark Moccia, Rick Parrish, and Fiona Swerdlow, Forrester, November 9, 2026
Why Leaders Need to Rethink Their Approach to AI Oversight. Ashley Reichheld, Simona Spelman, Allyson Dake, and Carl Bennett, Harvard Business Review, November 9, 2026
How AI is transforming automotive quality operations. Eike Reus with Yoav Levy, McKinsey, November 6, 2026