If you switched your AI systems off tomorrow, which decisions in your company would change?
In my upcoming book, Coherence at Velocity, I describe structural lag as the distance between the moment an enterprise knows it should change and the moment it changes coherently. The example I use is a Tier-1 global bank that built AI-powered personalization features several times faster than its old operating model could release them, because every customer-facing update still passed through gates designed for a slower, more centralized era. The bank could build faster than it could decide. Authors at HBR, Forrester, and Gartner published between October 30 and November 2 on the version of that lag that appears after the system goes live: the enterprise deploys an AI system that outperforms its people, keeps the people accountable for every outcome, and then watches them override it.
On Thursday I covered the agents your customers send ahead of themselves. This edition returns inside the enterprise, to the decision authority your own AI systems have on paper and the authority they have in practice.
Expect your people to override an AI system you never gave authority to
Das Narayandas of Harvard Business School writes that companies deploy AI systems faster than they redesign the governance around them, and that employees who remain personally accountable for outcomes will override a better-performing system because overriding is the rational response to their incentives. Laura Cross of Forrester writes that applying AI to a fragmented process does not make the process coherent, and that leaders should settle who owns which work and who holds decision authority before they ask whether they have enough people. Gartner's Q3 2026 survey of 308 risk, audit, and senior executives ranks AI strategy miscalculation, which Gartner defines as investing too aggressively, picking unsuitable use cases, or moving before the organization can manage the operational and governance implications, as the third most important emerging risk. Julian De Freitas of Harvard Business School writes that as chatbots become the customer's primary interface, relationship management turns into a product design and governance problem, and that leaders have to decide which AI relationship roles belong under the same brand.
Read together, these authors describe a single failure that most AI business cases never price. The system can decide, the organization still routes the decision to a person, and the person who carries the blame sends it back for review.
Transfer authority with accountability, and record every override
Narayandas, the Edsel Bryant Ford Professor of Business Administration at Harvard Business School, published Why Even Your Best Employees Override AI Systems That Work in HBR's November-December issue, live November 1. He opens with a global bank that approved an AI system in late 2022 to automate routine credit decisions. Risk and compliance signed off, and the CEO promoted it. In pilots the system processed applications faster than the credit team and with fewer errors. Six months later most files still went to manual review, the temporary safeguards had become routine steps, and the number of exceptions had grown. The bank's explanations ran to the usual list: the organization was not ready, the models needed work, people needed training. Narayandas's reading is simpler. The AI never received decision authority. The same pattern, he writes, appeared with ERP rollouts, credit scoring, and revenue management, and agentic AI raises the stakes because an agent that only recommends has nothing to act on.
He gives leaders four questions to answer before scaling: who loses discretion if the system works, who gets decision rights over which matters, who explains the outcome when people follow the system, and who bears the transition risk during the learning curve. His prescriptions follow from the questions. Make the algorithmic decision the default and require a logged reason for every override. Redesign roles toward oversight, boundary-setting, and exception handling. Reward correct use of the system, and stop grading individuals on outcomes they no longer control. His closing test is the question I opened with.
In my book, Marketing Operations 3.0, I describe the Autonomy Fit Matrix, which sorts work by reversibility and risk exposure. Low-risk, reversible work, such as routine refunds within pre-approved policy or post-contact summaries, belongs in the efficiency zone where the system acts and people handle exceptions. High-risk, irreversible work stays with people and the system supplies decision support. Narayandas's bank put a routine credit decision in the efficiency zone on the org chart and kept it in the human-led zone on the floor. Nobody wrote down which zone the decision belonged in, so the credit officers chose the zone that protected them.
Sangeet Paul Choudary made the companion argument in HBR on November 2 with The "Human-in-the-Loop" Fallacy: assigning employees to review AI output and handle exceptions shrinks their role as the system improves and erodes the attention they need to catch the errors that remain. Nobody has to answer the authority question while a review queue exists.
Take a hypothetical home services company with 400,000 service contacts a month, 12 percent of which involve a billing credit decision, or 48,000 decisions. A rep spends about 4 minutes on each one. The company deploys an AI agent at $600,000 a year, including licenses, consumption, and governance, to decide credits.
In the first version, the agent recommends and the rep decides. Reps read the recommendation, re-check the account, and approve or edit the credit. Each decision now takes 3 minutes, which frees 800 hours a month, or 9,600 hours a year. The team assigns half of those hours to named work, worth about $216,000 at a $45 loaded rate, and the rest dissolves into the queue. Return on Agent comes to ($216,000 minus $600,000) divided by $600,000, or about negative 64 percent. The usage dashboard shows 48,000 agent-assisted decisions a month, and the CFO sees a loss.
In the second version, the service leader and the CFO put credits of $50 and under, which cover 85 percent of the volume, in the efficiency zone. The agent issues those credits by default, every override goes into a register with a reason code, and a quality lead audits a 10 percent sample at 2 minutes each. Reps handle the 15 percent above $50 and the exceptions. The agent removes about 2,720 hours of rep time a month and the audit adds back 136, for a net of about 2,584 hours a month, roughly 31,000 a year. The capacity ledger assigns 24,000 of them to retention calls and three vacant roles the team leaves unfilled, worth about $1.08 million. Return on Agent comes to about 80 percent. The override register shows a 7 percent rate in month one and 3 percent by month four, and the reps who overrode correctly in month one wrote the two policy fixes that drove the rate down.
The service leader changed one thing between the two versions: who decided by default.
My take: Pull the override rate for every AI system in your marketing and service stack that makes recommendations. If nobody records overrides, you have a recommendation engine with a usage dashboard, and you should price it as one.
Standardize the process and name its decision owner before you automate it
Cross, a VP and principal analyst, published Stop Choosing Between Running And Transforming The Business on November 2. She argues that operations leaders treat running the business and transforming it as competing priorities, then layer transformation on top of daily work as extra effort, which blurs ownership and adds complexity. Her sequence starts with four questions: what work runs the business, what work transforms it, who owns each, and who has decision authority. She lists the decision rights that matter: what gets prioritized, standardized, automated, funded, scaled, and stopped. On AI she is direct. Applying it to fragmented processes does not make them coherent, automation accelerates whatever weakness already exists, and leaders should decide which steps to standardize before automating any of them.
Cross's "standardize before you automate" is the step Narayandas's bank skipped. The credit process had informal thresholds, dual-review habits, and manual confirmation flags that nobody had written down, and the AI system inherited all of them as exceptions. Standardizing the process means deciding, on paper, which cases the system owns outright.
Section 17.4 of Marketing Operations 3.0 describes governance-as-code: pricing limits, consent rules, approval thresholds, escalation triggers, and brand constraints written in a form the system can interpret before it acts. Cross's list of decision rights is the index for that document. Once the agent checks a threshold at runtime and logs each override, the organization has transferred that decision right, and a threshold that lives in a policy folder has transferred nothing.
It reminds me of what Kevin Yang, Director of AI at Front, said when I interviewed him on The Agile Brand podcast: "The takeaway isn't that, hey, AI doesn't work, it's that automation alone doesn't solve coordination. AI actually performs best when it's embedded in the workflows with access to full customer context, the right systems, and clear ownership across teams."
Yang's company measured the coordination cost directly and found teams spending three hours coordinating for every hour solving customer problems. The Structural Friction Index measures the same thing in a form a CMO can track quarterly: meetings, handoffs, and approvals divided by customer-impacting outputs. An AI deployment that adds a review step to every decision raises the index while the vendor's dashboard reports adoption.
My take: For each AI program in your 2027 plan, write the process it automates as a one-page standard with its thresholds and exceptions before the build starts. If the standard does not exist, the program automates your current inconsistency.
Put every AI investment in front of the people who manage enterprise risk
Gartner published its Q3 2026 Emerging Risk survey on November 2, drawing on 308 enterprise risk management leaders, auditors, and senior executives. Two AI risks entered the top five for the first time. AI strategy miscalculation ranked third, behind AI discovery of cyber vulnerabilities and geopolitical energy supply shocks, and AI-driven competitive displacement also entered the top five. Kevin Mercado, senior principal analyst in Gartner's Risk and Audit Practice, said AI risk has moved beyond technology and governance concerns into the core of enterprise strategy. Gartner's recommendations: evaluate AI investments as a portfolio of business decisions, tie each initiative to a strategic priority, validate readiness before scaling, and measure AI's impact on decision-making and the quality of decisions, with productivity as a secondary measure.
That last recommendation changes the report a CMO sends upward. On October 23 I argued that every AI program needs a one-page CFO report with outcome, total cost, and Return on Agent. Gartner's recommendation adds a second line to that page: how many decisions the system made, how many a person overrode, and what happened to the decisions it made. A program with high usage, a 60 percent override rate, and no record of why the overrides happened is the "moving ahead before the organization is prepared" case Gartner's definition names.
Marketing has spent two years asking for AI budget on the growth side of the ledger, and Gartner's survey now places the same investments on the risk register, so the two conversations should run on the same numbers. The override register supplies most of them.
My take: Before the 2027 budget locks, give your enterprise risk lead a list of every AI program that makes or recommends a customer-facing decision, with its decision owner and override rate. Expect the list to be shorter than your vendor count.
Give your customer-facing assistant a charter before it defines your brand
De Freitas, an associate professor in the marketing unit at Harvard Business School, published Your Chatbot Will Define Your Brand in the same November-December issue. He describes presenting research to executives on how many people report feeling closer to a chatbot than to their friends, and he argues that as conversational interfaces become the customer's primary interface with a company, relationship management shifts from a marketing challenge to a product design challenge. Leaders, in his account, have to design and govern the relationship life cycle: define the chatbot's role, adapt interactions to the customer's needs, calibrate intimacy, repair trust when the assistant violates expectations, and decide which AI relationship roles belong under one brand.
The last item is the one most brands have already got wrong. A finance app whose assistant coaches the customer on spending in the morning and sells a credit product in the afternoon has two roles under one name, and the customer experiences the contradiction before the brand team hears about it. Chapter 6 of Coherence at Velocity makes the case that brand has become behavior: decision policies, eligibility logic, and the consistency of an answer across channels. An assistant with authority to offer, decline, and escalate is the most concentrated expression of that behavior a company has ever shipped.
De Freitas's charter and Narayandas's decision rights are the same document seen from two sides. The decision-rights side lists what the assistant may decide, and the charter side lists what it exists to do. On October 29 I described customers who send an agent ahead of themselves. The owned-site assistant is the counterpart your company sends to meet them, and it needs the same authority definition, the same override record, and the same owner.
Alla Valente of Forrester added the contractual version on November 2 in "It Wasn't Me" Is Not An AI Risk Management Strategy, prompted by Meta's updated commercial terms, under which continued use after November 20 means accepting sole responsibility for its AI features and their output. Her four steps apply to every vendor assistant in your stack: bound the agent's autonomy in policy or code, demand logs and incident reporting, match accountability to control, and set remedies in advance.
It reminds me of what Albert Castellana, Co-Founder and CEO of GenLayer, said when I interviewed him on The Agile Brand podcast: "You're saying, okay, I want this agent to be representing my brand in front of the public, but I don't even know how it's thinking, what it's doing, what promises it's making, right? Did it hallucinate something, right? Did it, you know, deny something, let's say, deny a refund that really should have been refunded, right? Can I get sued because of that? Am I responsible because of that, right?"
Castellana's questions and Narayandas's four questions share one answer: a named person. A RACI matrix for an assistant looks odd to people who are used to building them for campaigns, and it is the fastest way I know to find the decision nobody owns.
My take: Write a one-page charter for your customer-facing assistant: its role, the decisions it owns outright, the decisions it escalates, the roles it will never play, and the person accountable for all four. If a vendor's terms shift responsibility to you, that charter is also your audit file.
Take these actions before the 2027 budget locks
Run the switch-off test. For each AI system, ask which decisions would change tomorrow if it went dark. Where the answer is "none," authority never moved.
Make the system the default in the efficiency zone. Low-risk, reversible, in-policy decisions go to the agent; a person overrides with a logged reason.
Open an override register. Rate, reason codes, and outcomes, reviewed monthly by the decision owner.
Standardize before you automate. One-page process standard with thresholds and exceptions, written before the build.
Put AI programs on the risk register. Decision owner, override rate, and decision outcomes beside the growth case.
Charter the assistant. Role, owned decisions, escalations, prohibited roles, and one accountable name.
Move the signature authority on day one
Each of the three companies I acquired came with a decision that had an owner on paper and a different owner in practice. The day after close, the question on every desk was who signs, and until we answered it, the former owners kept signing, because signing was the safe thing to do. Integration started the week the new authority matrix went out, and nothing we announced before that week changed a single decision.
Coherence at Velocity describes governance designed for throughput: low-risk changes move with far less manual escalation, higher-risk changes get scrutiny that is explicit, proportionate, and tied to reversibility and impact. Table 4.1 in the same chapter lists what leaders count as progress against what counts, and the row that applies here reads "more experimentation or automation" on one side and "change that can be repeated without destabilizing the surrounding system" on the other. An AI system that recommends while people re-decide produces the first column. Narayandas measured the gap at a bank over six months, Cross located its cause in processes nobody standardized, Gartner's respondents put it third on the enterprise risk list, and De Freitas found it in the assistant that now speaks for the brand.
Answer the opening question this week. Run the switch-off test on every AI system in your stack, move the in-policy decisions to the system by default, open the override register, and name the person who answers for both the decisions it makes and the ones your people take back.
Featured this cycle:
Why Even Your Best Employees Override AI Systems That Work. Das Narayandas, Harvard Business Review, November-December 2026 issue (online November 1, 2026)
Stop Choosing Between Running And Transforming The Business. Laura Cross, Forrester, November 2, 2026
Gartner Survey Finds Two New AI-Related Risks Emerge in Third-Quarter 2026 Report. Gartner, November 2, 2026
Your Chatbot Will Define Your Brand. Julian De Freitas, Harvard Business Review, November-December 2026 issue (online November 1, 2026)
Also cited: The "Human-in-the-Loop" Fallacy. Sangeet Paul Choudary, Harvard Business Review, November 2, 2026. "It Wasn't Me" Is Not An AI Risk Management Strategy. Alla Valente, Forrester, November 2, 2026