Why teams ignore AI insights
AI can now tell a team what to change. Often nobody has been given the job of deciding whether to.
By Nathan Barling Product Marketing Principal 7 min read
The scenario
On Monday morning, the model flags that paid social is past the point of diminishing returns and retail media is underspent. Moving $40,000 from one to the other looks like an easy call.
But paid social belongs to one team and retail media to another, and moving budget between them waits for Finance's quarterly review. So the insight gets logged, everyone agrees it's probably right, and the money stays where it is.
An AI insight is a request for a decision. Too often, nobody has been told they're the one who answers it. So it sits, or it gets skipped, and nothing records that it happened.
By an insight, we mean any time an AI system tells a team to change something: a price, a segment, a bid, a budget. It's often right. What happens next depends on who can act, and whether the people around them agree on what the model is for. Decision intelligence starts from that point. The decision matters as much as the data behind it.
The lead handoff already shows the problem
No survey we found measures how often teams ignore AI insights. That's part of the story. Ignoring an insight usually leaves no trace, so nobody counts it.
The closest evidence comes from a recommendation marketers have passed along for years: the marketing qualified lead. In a September 2026 survey of 192 practitioners, fielded by Heinz Marketing for Docket, 78% said sales rejects more than half of marketing qualified leads. Docket sells AI sales agents. Nearly 90% still report lead volume to leadership. Only 7% of those are confident it predicts pipeline.
Figure 1
The lead handoff, 100 respondents at a time
Each square is one organization or respondent, scaled to 100
Three 100-square grids. 78 of 100 organizations say sales rejects more than half of marketing qualified leads. About 90 of 100 respondents report lead volume to leadership. 7 of every 100 who report it are confident that volume predicts pipeline.
The survey doesn't say why sales turns those leads down. But a qualified lead is a recommendation. Marketing's system says this one is worth a call, and sales decides whether to believe it. More often than not, sales doesn't, and the volume gets reported upward anyway. A sharper score won't settle that, because the two teams have to agree on what qualified means first. AI insights now run into the same wall, on far more decisions than lead routing.
The wider numbers point the same way. In McKinsey's 2026 survey of 1,719 respondents, about 80% said AI improved their own productivity. Only 37% said it contributed to company earnings, about the same share as a year earlier. Part of the reason may be simple. A person can adopt a tool without asking anyone, but changing a price or a budget takes several functions agreeing.
Diagram. Adopting a tool, one person decides. Changing a price, segment, or budget, marketing, finance, and sales have to agree.
What a model makes teams write down
Teams can work around a missing definition for years. Marketing counts one kind of customer, sales counts another, and the difference gets settled in meetings. Nobody writes it down.
A model can't work that way. To optimize anything, it needs written definitions and one number to maximize. Writing those down is what context engineering and a semantic layer are for. For many teams, it's the first time anyone has had to agree on them.
Sometimes the definition is there and the goal is wrong. Say the model is set to maximize return on ad spend. It keeps recommending the promotion that sells the most units, while finance watches the margin on each one shrink. The recommendation is correct for the number the model was given. Finance says no anyway, and from the outside it looks like one more ignored insight.
And sometimes the definition and the goal are both fine, but the person accountable for the result can't make the change. That one is the hardest to spot, because everyone involved is doing their job.
Insights leave a record
There's a second change, and it's easy to miss. A budget argument used to happen in a meeting and leave nothing behind. A logged insight has a timestamp, a recommendation, and an outcome, even when the outcome is that nobody acted.
That makes indecision something you can count. How many insights did the team see last quarter? How many got a yes, a no, or an override with a reason? How many just aged out? If nobody can answer, that's the finding.
One count is enough to start. Of the insights your team saw, what share got a recorded decision within two weeks? Acted on, turned down, and overridden with a reason all count. Silence doesn't. We haven't found a benchmark for it, so your first count is the baseline.
First, rule out the data
Not every ignored insight is a decision problem. In Collibra's 2026 survey, 72% of AI decision-makers said a poor data foundation was the root cause when initiatives fell short. Collibra sells data governance software. In Forrester's June survey, US agencies named accuracy and bias (63%) and legal concerns (62%) as top barriers.
Volume matters too. A system that sends 200 recommendations a week is asking for 200 decisions, and no team has that many to give. Skipping most of them is triage, and it's often the right call. The fix there is fewer, better-ranked insights, and it comes before any question about who decides.
So before redrawing anyone's job, sort the cases. Pull the last 10 insights your team ignored.
The insight was wrong
Model or data problemIt was right and nothing happened
Decision problemPeople overrode it because they didn't trust the numbers
Data problem, and a reasonable overrideIf people overrode an insight because they didn't trust the numbers, they were probably right to. The fix is the data, and data governance is where it starts. If the insight was right and nothing happened, the next question is who could have said yes.
Who can say yes
Write down who is held accountable for each result and who can actually change it. Here's one common way it splits.
| Decision | Accountable for the result | Authority to change it |
|---|---|---|
| Pricing and promotions | MarketingRevenue target | Finance Sales |
| Who counts as a customer | MarketingAcquisition target | Sales Product |
| What makes a lead qualified | MarketingPipeline target | Sales |
| Moving budget mid-quarter | MarketingGrowth target | FinanceQuarterly approval |
| What counts as marketing-influenced revenue | Marketing | Finance |
Marketing is accountable in every row and holds the authority in none. Your table will look different. Fill it in anyway, because the rows where the two columns name different teams are where insights stall.
Keep the overrides
Back to that Monday. The same week the budget insight stalls, a planner overrides a different recommendation because she doesn't trust the number behind it. She might be right. Nobody writes down that she did it, so nobody follows up, and nobody finds out. Next month the model says the same thing, and it goes the same way.
Overrides can be the most useful entries in the log. A 2026 study of human-in-the-loop pricing by Charles River Associates found that people and a model working together can beat either one alone, but only when the person adds something the model can't see. An override with a recorded reason tells you which kind it was. One without a reason tells you nothing.
So ask one thing about your own process. When someone overrides an insight, are the override and the reason recorded anywhere?
What it looks like when it works
Picture the same Monday, and the same $40,000, at a team that has done this work.
Last quarter the two channel leads agreed with Finance that either of them can move up to $50,000 a month between channels without waiting for the review. So the insight arrives with a named owner. She checks it against retail media's stock position, which the model can't see, approves $25,000, and writes one line on why she held back the rest. It takes two days. At the quarterly review, Finance reads the log.
Nothing about the model changed. The team decided ahead of time who answers, within what limits, and where the answer goes.
The survey data fits. McKinsey's high performers are the roughly 6% of respondents who attribute 5% or more of their earnings to AI. About three-quarters of them say they've fundamentally redesigned workflows, against about a quarter of everyone else. A survey can't show cause, but it points the same direction.
We've seen a version of it in our own work. For Kohler's Sterling brand in 2022, PMG ran strategy, creative, and media together and built a dashboard to guide optimizations and budget. The team reading the dashboard was the team that could act on it. Within weeks of launch, the work delivered 34% higher cost effectiveness than benchmark.
Where to start
Days 1 to 30
Pick one decision where insights get ignored. Get both sides in a room and agree on one definition and one number to optimize.
Leave with: a written definition and one number
Days 31 to 60
Name an owner for that decision. Set the override rules: who can override, how fast, and where the reason gets logged. Then start counting the share of insights that get a recorded decision within two weeks.
Leave with: a named owner, override rules, and a first count
Days 61 to 90
Review the log. Where overrides were wrong, make the insight the default. Where they were right, fix the model or the data. Then count again.
Leave with: a reviewed log and a second count
Start with the insight your team ignores most often, and find the person who could have approved it. Ask what would have had to be true for them to say yes.
Frequently asked questions
Why do teams ignore AI insights?
Often because the person who sees the insight can't make the change it asks for, or because teams disagree on what the model should maximize. Data quality and trust matter too.
Are ignored AI insights usually wrong?
Not necessarily. Pull the last 10 ignored insights. If they were wrong, it's a model or data problem. If they were right and nothing happened, it's a decision problem.
How do you get a team to act on AI insights?
Agree on one definition and one number to optimize, name an owner for the decision, and set override rules, including where the reason is logged. Then track the share of insights that get a recorded decision.
Should people override AI recommendations?
Sometimes. A 2026 Charles River Associates study of pricing found people and a model together can beat either one alone, but only when the person adds something the model can't see. Log the reason for every override so you can tell which kind it was.
Sources
- McKinsey, The state of AI in 2026: On the road to ROI, August 2026.
- Forrester and 4As, Nine in 10 US marketing agencies use AI to cut costs at the expense of creativity, June 2026.
- Docket and Heinz Marketing, The Trust Gap, Part 1: Sales and marketing alignment, October 2026. Survey fielded in September 2026. Docket sells AI sales agents.
- Collibra and Harris Poll, AI failures often trace back to poor data foundation, September 2026. Reported by CIO Dive. Collibra sells data governance software.
- Charles River Associates, Human-in-the-loop dynamic pricing, July 2026.
- PMG, Plumbers Getting It Done.
Last updated October 31, 2026. Statistics are quoted from the sources above and were checked on that date. Survey results are self-reported unless noted.