Who Owns the AI Backlog When Every Department Has a Pilot

One person owns it, and it is the person carrying the AI mandate rather than the department with the loudest sponsor. In a company with no full-time AI executive, that is whoever holds the fractional AI officer seat. Ownership counts only when it arrives with three powers: the power to accept a request onto the list, the power to reject one outright, and the power to decide what order the survivors run in.

The departments keep what was always theirs. Support still owns first-response time. Finance still owns days payable. The person who owns the backlog does not own those numbers and should never pretend to, because the moment they do, the department stops being accountable for the outcome and starts being a customer waiting on IT. What the owner holds is narrower and harder: which of the five things goes next, and which of the five never goes at all.

The artifact that makes this real is boring. It is one list, visible to everyone who asked, with every request on it in order, every parked item carrying the condition that unparks it, and every rejection carrying its reason. Not five departmental roadmaps that meet quarterly. One list. If your company has five roadmaps, you do not have a prioritization problem yet. You have an ownership problem, and the prioritization problem is downstream of it.

Accept, reject, sequence. Hold all three and you own the backlog. Hold two and you own a shape of failure that already has a name.

Accept without reject is a queue that only grows. Every request lands, nothing leaves, and the list turns into a monument to good intentions that the department heads stop reading around month four. This is the most common failure and the most flattering one, because saying yes to everybody feels like leadership right up to the quarter where nothing shipped.

Reject without sequence is a compliance checkpoint. You can block things, so people stop bringing you things. They buy tools on the marketing card instead and you find out at renewal. KPMG's June 2026 global survey of 2,145 senior leaders found only 24% put accountability for AI outcomes with the CEO, and the ones who did were in a different business: 57% reported meaningful value against 21% for everyone else, and 14% had established ROI against 4%. Accountability parked nowhere in particular behaves exactly like the checkpoint.

Then there is sequence with neither, which is a project manager holding a spreadsheet. Useful. Not the job. Somebody upstream is still deciding what belongs on the list, and that somebody is whoever shouted last.

Gartner's April 2026 note on AI agent sprawl found that only 13% of organizations believe they have the right agent governance in place, and projected the average global Fortune 500 enterprise running more than 150,000 agents by 2028, up from fewer than 15 in 2025. Nobody decides their way to that number. You get there by never having decided anything.

Every framework for prioritizing AI work assumes a blank list. Nobody arrives to a blank list. The first ten working days are an inventory rather than a strategy, and that inventory is usually the most valuable thing produced in the whole first month.

The scale of it surprises most executives. Salesforce's 2026 Connectivity Benchmark put the average enterprise at twelve AI agents already running, with half of them operating in isolation from each other.

The buying is just as scattered. Zylo's 2026 SaaS Management Index found IT is responsible for 15% of software spend and 13% of the applications actually in use, which means the other 87% arrived through somebody's expense report. The average organization in that data carries seven generative AI apps. Nobody sat down and chose seven.

Six fields per pilot, and that is the entire inventory. Who runs it. What it costs a month. What data it touches. What number it was supposed to move. Whether anybody measured that number before it started. Whether it is still running.

Most pilots fail the fifth field. That is your first finding and it costs nothing to produce: a room full of live experiments, none with a before number, so none of them can be shown to have worked or failed. That is not a criticism of the people who built them. It is the predictable result of five departments each starting alone. Our free audit exists mostly to run this count, because the count is what turns an argument about priorities into a conversation about facts.

Every request arrives with a demo attached and a sentence about the hours it will save. Neither one is evidence, and neither one is what the four gates ask for. Fail one and it does not get argued about, it gets parked with the specific thing that would unpark it. This is deliberately not a weighted score out of a hundred. A score is a debate about weights. A gate is a yes or a no that anybody in the room can check.

One. A named owner who can change the workflow. Not a sponsor, not a stakeholder. The person who can tell their team to stop doing it the old way on a Monday morning. If the answer is a committee, the answer is no, because a committee cannot be phoned when the agent gets something wrong at 4pm.

Two. A baseline number, measured today. Not estimated. Measured. If nobody knows how many tickets came in last week or how long the invoice run takes, the project cannot succeed, because success has no definition. Measuring the baseline is usually a two-day job and it is always cheaper than the build.

Three. Data you can reach inside ten working days. An export, an API, a database view, anything. If the data lives behind a legacy system that only produces a monthly PDF, the AI request is really an integration request, and it should be funded and sequenced as one.

Four. A wrong answer that can be undone. Draft the email, do not send it. Suggest the code, do not post it. Where the action is genuinely irreversible, this gate is not a permanent no. It is a demand that the human approval step and the evaluation set get designed before the build rather than after the first bad week.

Four gates, in that order, because each one costs less to check than the one after it.

A flat no is the most expensive sentence in this job. The department that hears it does not stop wanting the thing. It stops telling you about it and buys the thing itself, and now you have a shadow pilot you cannot see, on data you did not approve, that renews automatically.

The same Zylo index measures the size of that reflex. Sixty percent of IT leaders admitted they cannot see every generative AI tool already in use, and 77% have found AI features running somewhere in their stack without IT knowing about it. People route around a closed door. Every time.

So there are three shapes of no, and each one hands something back.

Not now. The request clears every gate. The slot is taken by something with a shorter path to evidence. What they leave with is a position on the one list, the date it gets reviewed, and the single number that would move it up. That last part matters more than the position, because a department that knows the rule can play by it.

Not this way. The outcome is worth having and the build is the wrong instrument. A good share of what reaches an AI backlog is really a report nobody has asked finance to run, a permission somebody could change in an afternoon, or a seat in a tool the company already pays for. Naming the cheaper route out loud is the highest-trust move available here, because it costs you a project and buys you a department that believes you.

Not ever, as written. It fails a gate that will not move. No owner, no measurable outcome, or a blast radius nobody will sign for. What they leave with is the reason in writing and the exact condition that reopens it.

The rule that makes all three survivable: every no gets written down with its reason, and anything parked or rejected can come back in ninety days with the missing evidence attached. A no that cannot be appealed is just a grudge with a calendar invite on it.

Order by time to first evidence, not by the size of the prize. This is the rule that most disappoints executives and most reliably works.

The reason is arithmetic rather than philosophy. A six-month project with a large number attached produces nothing to show at the board meeting in November. A three-week project with a smaller number produces a measured before and after, and that measured result is what funds the six-month project in January. Evidence compounds. Ambition does not. McKinsey's state of AI research has 88% of organizations using AI in at least one function while only 39% can point to any EBIT impact. The variable that correlates hardest with landing on the right side of that gap is workflow redesign, and only 21% of the organizations using gen AI have redesigned even some of their workflows.

Three rules do the actual sequencing.

One department at the front at a time. Running one project in each of five departments looks fair and delivers nothing, because the constraint is never the builder. It is how much attention a department head can give to changing how their team works while also running their team. Two of those at once is already a stretch. Five is a promise nobody keeps.

Protect one slot a quarter for the thing nobody asked for. Data access, the evaluation harness, the audit log, the integration that unblocks four other requests. No department will ever request these and everything else waits on them. If the list only contains what was asked for, the list will stall in month five and nobody will be able to say why.

Cap the front. Count how many live builds and discoveries the team can actually carry at once, then say that number out loud to every department in the first week, so that accepting one more request visibly costs something. A cap that nobody states is a cap that gets quietly broken, and a backlog owner who breaks their own cap has stopped owning anything.

None of this is a scoring rubric, deliberately. If your ordering needs three decimal places to defend it, the problem was never the ordering.

Let them, and write down the price. Fighting it burns the mandate you need next quarter. Losing it quietly is worse.

The sentence is short. "Understood, this goes to the front. That displaces the invoice coding work, which now lands in March instead of January. I will update the list today." Then update the list today, in public, where every department that asked can see it.

That turns an override from a political win into a trade with a visible cost, which is all you actually need. Sometimes the executive looks at the trade and takes it, and they were probably right, because they can see something you cannot. Sometimes they look at the trade and quietly put it back. Either way, the list stayed honest, and the people whose work just moved found out from the list rather than from a rumor.

A backlog that silently reorders itself after every hallway conversation is not a backlog. It is a record of who spoke to the CEO most recently, and every department will work out within two quarters that speaking to the CEO is the fastest route to a slot. At that point you have taught the building to route around you, which is the same failure as the flat no, arriving by a more polite road.

This is also the clearest signal in the room about whether the seat was real. If overrides never come with a written trade, the mandate was decorative from the start, and that is one of the situations where a fractional hire was the wrong purchase rather than a fixable problem.

Accepting things is only possible if things also leave. The review that kills work is what makes the rest of this affordable, and it runs on a schedule so that nobody has to be brave on a Tuesday.

Four stop rules, checked every quarter, any one of which ends the item.

1. The named owner is gone or has moved on. Not replaced, gone. The owner was the whole basis for gate one.
2. The baseline never got measured. Three months in with no before number means there will never be an after number.
3. Adoption is below the floor the owner agreed to. The floor gets written down at the start, in the owner's own words, so this is not a judgment call later.
4. There is no credible path to the number. It works, people use it, and it is not going to move anything anybody cares about. This is the hardest one and the most important.

Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, on escalating costs, unclear value or weak risk controls. Those cancellations are coming either way. The only question is whether they happen on a schedule, with the lesson written down and the budget recovered, or eighteen months late when somebody notices the renewal.

Killing a pilot with the reason written down is the cheapest line on the whole list, and it is the only way the next request gets a slot.