The last piece made the case that AI spending fails on the foundation it lands on, and ended with the audit, three honest counts of what’s in use, where the data breaks, and how much AI the company already owns. This one is about what to do with the map.
The appetite is not the problem. 89% of martech leaders with AI agent initiatives still expect them to deliver significant business benefit, per Gartner’s late-2025 survey, even as nearly half report the tools underdelivering today. The gap between expecting benefit and collecting it is an operating strategy, and the strategy is assignment. Every AI in the building gets a job, an owner, and a number, or it gets switched off. The foundation work, connecting the data seams and writing down the workflows the audit exposed, runs in parallel, and every seam that closes makes every assignment below perform better.
One layer per job
The audit almost always finds the same capability owned three ways. Drafting lives in the assistant seats, in the marketing platform’s AI, and in a point tool somebody bought last spring. Scoring lives in the CRM and in an add-on. Pick one owner per job and switch the duplicates off, with their seats flagged for the next renewal.
Choosing the owner is rarely a coin flip. The job goes to the layer closest to the data it needs and the person who checks it. Scoring belongs in the CRM because the records live there. Send-time optimization belongs in the platform that sends. General drafting belongs in the assistant seats, where a human is in the loop by design. A point tool keeps its job only when it does something the layers you already pay for genuinely can’t.
Deduplication also means reading vendor labels skeptically. Gartner estimates that of the thousands of vendors selling “AI agents,” only about 130 offer real agentic capability, the rest rebranding existing assistants, chatbots, and automation, a practice the firm calls agent washing. A rebranded chatbot doesn’t earn a second subscription to a job the stack already covers, whatever the label says.
Every AI gets a job description and an owner
For each layer that keeps a job, write down what it drafts, what it scores, what it may send, and which human checks its work before the work counts. The document can be a page. The point is that it exists, because an AI without an owner drifts into work nobody assigned it, two owned AIs never end up fighting over the same customer record, and when something goes wrong, and in marketing something eventually publishes wrong, the question of who checks what has an answer that predates the incident. In a regulated industry that page is also the first thing legal and compliance ask for, and having it in writing before they ask is the difference between a review and an investigation.
The job descriptions matter double because the unassigned AI is already in the building, in tools nobody approved and workflows nobody wrote down. Assignment beats prohibition. People stop improvising with unapproved tools when the approved one has a clear job and actually works, and a written policy makes the boundary legible to everyone.
A number for every assignment
Every assignment ties to a process and a number. Time from brief to approved draft. Lead response time. The share of records complete enough to score. Baseline the number the week before the AI takes the job, and read it again in ninety days.
Optimism is the industry default, the 89% above proves it, and a baseline is what turns optimism into a finding. Sometimes the finding is that the tool moved the number and earns a bigger job. Sometimes the finding is that it underperforms the process it replaced, which is worth knowing at renewal, when the finding converts directly into budget. Either way the board stops hearing adjectives.
A gate for the next purchase
New AI spending passes one written test before procurement sees it. The process it fixes, the number that should move, and the clean, connected data it needs, named in advance by the team that will use it. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, on escalating costs, unclear business value, and inadequate risk controls, and those are precisely the three failure modes a written test screens for. A tool that can’t name its process, its number, and its data before purchase was going to become one of the 40%.
One more sequencing rule keeps the whole thing honest. While the foundation work runs, at most one new tool enters per quarter, because every addition mid-repair is another layer to audit later.
The next board conversation
It opens with the roster. Which AI holds which job, who owns each one, which numbers moved this quarter and from what baseline, what the gate turned away, and which seam the foundation work closes next. That’s an AI strategy a board can actually govern, and none of it requires a new subscription to start.
An AI without a job description is just a subscription.
