What will your next AI initiative inherit from the last one?
Coherence at Velocity defines its North Star as the ability to improve customer-facing experiences quickly and repeatedly. Table 4.1 in chapter 4 makes the second word the test: leaders should count progress as a stable ability to improve again without treating each release as a controlled emergency, and treat a heroic launch financed by exceptions and deferred cleanup as a temporary achievement.
Three Forrester analysts published pieces between November 13 and November 15 that apply that test to AI. Most marketing and CX leaders I talk with can list their AI initiatives in production and struggle to say what the third initiative reused from the first two.
Treat reuse, cost, and clarity as one scaling problem
Frederic Giron reports that AI has reached production in most organizations and stalled there, because each new initiative rebuilds the data access, controls, and value case its predecessor already paid for. Greg Zorella argues that nobody can calculate AI ROI without first agreeing on how to categorize and allocate AI costs, and that token bills cover only a small share of the total. Gina Bhawalkar finds that most customers in six industries say brands fail to communicate in plain language, and she asks for human review of generative AI content before it reaches them.
Read together, the three pieces describe one leadership job. Before you approve the next AI initiative, decide which shared assets it must use, how its full cost will be counted, and which approved language it will put in front of customers. You pay for each of those decisions once when you build it into a shared asset that later initiatives reuse.
Fund the second initiative with what the first one leaves behind
Giron, VP and senior research director at Forrester, published AI Has A Scaling Problem on November 15. Citing Forrester's State Of AI Survey, 2026, he reports that 82 percent of AI decision-makers have at least one generative AI application in production and 77 percent have at least one agentic AI application, while only 20 percent believe their organization has the technology in place to scale AI initiatives. Organizations with high technology readiness run 2.4 times as many AI use cases in production as low-readiness organizations, yet only 42 percent can link AI initiatives to the P&L. His prescription is a flywheel in which each initiative leaves reusable infrastructure, data, models, connectors, controls for agent access, and documented value. "The harder part now is making each new and successive AI initiative faster, cheaper, and safer than the last," he writes. He closes with a funding gate: before approving the next initiative, teams state what they will reuse, what they will leave behind, and where the initiative does and does not yet deliver value.
That gate belongs in the CO2 Council described in Marketing Operations 3.0, where marketing, sales, service, and product leaders review the Commercial System as one lifecycle. Giron's "controls for agent access" map to the Identity and Permissions layer of my AI Capability Framework: on whose behalf an agent acts and how much authority it holds. A team that settles those answers once, in a shared service, saves the next three teams from settling them again with slightly different results.
It reminds me of what Kevin Lee, CTO from NiCE, said when I interviewed him on The Agile Brand podcast: "What we're seeing a lot of is organizations will pick the proverbial low-hanging fruit. Turns out the low-hanging fruit doesn't always yield high ROI."
Commentary: Giron's gate costs nothing to adopt. Add two required fields to every AI funding request, "inherits" and "leaves behind," and send back any request that leaves the second field blank.
Count the full cost of a capability before you claim the return
Zorella, a principal analyst at Forrester, published Tokenomics Needs A Taxonomy on November 13. He responds to the Tokenomics Foundation's work on a methodology for total AI cost, which argues that token and API costs make up only 10 to 25 percent of total AI spend. His point is that a defensible ROI needs a shared taxonomy in three steps: categorize AI costs (LLM, cloud, on-premises capitalization, labor), attribute them to defined capabilities, and allocate them to business outcomes. He compares the work to technology business management, which succeeds only when stakeholders agree on, own, and govern the allocation assumptions. His practical advice is to start from the outcome, build the total cost of ownership for each capability, and prove the method on one use case. "You can't wait for a standard taxonomy to emerge," he writes.
The AI Capability Framework's mechanism axis makes the same argument from the buyer's side: the "how" is where the money hides, and the lightest mechanism that does the job usually wins once you count review time, integration, and rework. Most AI business cases I review still track the token invoice in detail and estimate the labor line on a napkin, which is roughly backward given Zorella's numbers.
It reminds me of what Ben Schein, Chief AI and Analytics Officer from Domo, said when I interviewed him on The Agile Brand podcast: "We used to comfort ourselves with garbage in, garbage out. It was a harsh rule, but it was a fair one: fix the input and you fixed the output. AI broke the symmetry."
Commentary: Read Schein's point next to Zorella's taxonomy and Giron's flywheel. A team that cleans a shared data source once lowers the run cost of every initiative that reads from it, and finance will see that saving only in a capability-level cost model.
Approve the words before an AI system repeats them at scale
Bhawalkar, a principal analyst at Forrester, published Plain Language Is A Total Experience Imperative on November 13. Drawing on Forrester's total experience research, she reports that "Communicates with me using plain language" ranks among the top drivers of total experience for auto manufacturers, banks, health insurers, hotels, investment firms, and retailers. Fewer than two-thirds of customers in those industries say brands communicate in plain language, and in some industries fewer than half do. "Organizations routinely make customers work too hard to understand information," she writes. She previews six practices and details two: scale plain-language alternatives through the design system, with human review of generative AI content before publication, and apply information design practices such as descriptive headings, progressive disclosure, and accessible contrast and text scaling.
Forrester's Total Experience Score combines brand and customer experience, and Bhawalkar ranks plain wording among its top drivers. Table 6.1 of Coherence at Velocity translates the brand value of clarity into operating logic: rules, next steps, and obligations that a customer can understand at the moment of need. The fourth of the Agile Brand principles asks brands to respect customers and their time, and a return policy that takes three readings to decode spends the customer's time on the brand's behalf.
The stakes rise once you put generative AI in front of customers. A plain-language policy library that every system draws from becomes one of the reusable assets Giron describes. Without one, each AI initiative writes its own version of the same policy, and the customer gets to compare them.
It reminds me of what Kevin Lee of NiCE also told me on the podcast: "You cannot have that black box when you ask 10 questions and you get 12 different answers."
Commentary: Pick the five policies customers ask about most. Write one approved plain-language version of each, store it where every AI system can retrieve it, and require human review of any generated text that paraphrases it.
Run three initiatives two ways at a specialty retailer
Take a hypothetical home goods retailer with $1.2 billion in revenue. Its 2027 plan includes three AI initiatives: generated product descriptions for 40,000 SKUs, a service assistant for order and return questions, and personalized lifecycle email.
In the first version, three teams run the initiatives in sequence and in isolation. Each builds its own connector to product and order data, its own review process, and its own vendor contract. Each costs $900,000 and takes six months, for $2.7 million and 18 months. The product pages say returns are accepted within 30 days, the assistant says 30 days from delivery, and the email says within a month. Repeat contacts about returns climb. When the CFO asks for ROI, the only cost anyone tracked is the $310,000 token and API bill, about 11 percent of the total, and nobody can allocate the shared platform invoice to an outcome.
In the second version, the head of marketing operations requires the first initiative to leave three assets behind: a governed product-and-policy data connector with agent access controls, a plain-language library for return, shipping, and warranty terms with a human review step, and a cost model that tags every spend line as LLM, cloud, or labor and maps it to a capability. The first initiative costs $1.15 million and takes seven months. The service assistant reuses all three assets and costs $520,000 over three months. The lifecycle email program costs $430,000 over three months. The total comes to $2.1 million and 13 months, a saving of $600,000 and five months. Customers read one return policy in one wording everywhere, and the CFO receives a cost per capability tied to conversion and contact-rate outcomes.
The second team paid $250,000 more for its first initiative and recovered that amount several times over by the third.
Take these actions before the 2027 plan locks
Add a reuse gate. Require every AI funding request to name what it inherits from earlier work and what it will leave behind.
Fund shared assets as line items. Budget the data connector, access controls, and review workflow once, outside any single initiative.
Adopt a cost taxonomy on one use case. Categorize spend as LLM, cloud, capitalization, and labor, then attribute it to a capability and an outcome.
Build a plain-language policy library. Approve one version of each high-volume policy and make it the only source AI systems may paraphrase.
Review generated customer text. Put a human approval step on any AI output that states a policy, a price, or an obligation.
Budget for instability by name, then remove it
Coherence at Velocity names architecture as the first enabler: the degree to which systems, data, and dependencies allow change without destabilizing everything around them. The manuscript describes teams that budget for instability without naming it. They schedule extra coordination around a fragile dependency, narrow the scope of a release, and accept a smaller change because the cleaner one would touch too many systems. Giron's flywheel, Zorella's taxonomy, and Bhawalkar's policy library each name a piece of that hidden budget and turn it into an asset the next team can use.
Twelve years of piano lessons taught me that a teacher measures progress by what a student can still play from last month while learning this month's piece. Enterprise AI programs deserve the same standard.
Answer the opening question for every initiative in your 2027 plan. Ask each sponsor to name what the initiative inherits and what it will leave behind, fund the shared assets as their own line items, and decline any initiative whose answer to the second question is "nothing."
Featured this cycle:
AI Has A Scaling Problem. Frederic Giron, Forrester, November 15, 2026
Tokenomics Needs A Taxonomy. Greg Zorella, Forrester, November 13, 2026
Plain Language Is A Total Experience Imperative. Gina Bhawalkar, Forrester, November 13, 2026