Your team spends 87 minutes a day inside AI tools. Of that, 42% goes to troubleshooting errors and iterating on prompts, against 35% on work that actually advances the job (BambooHR, 2026). AI troubleshooting time is now the largest single category of AI time in the average workday — larger than the productive category it was bought to expand.
Annualized, the 87 minutes come to roughly 47 eight-hour workdays a year. About 20 of those days are spent getting the tool to behave.
That is not an adoption problem. Adoption is working. It is a measurement problem: almost every dashboard tracking this rollout counts the 87 minutes and none of them split it.
The Study Your Renewal Conversation Should Start With
BambooHR's Redesigning Work: AI's Performance Review surveyed 1,608 full-time salaried US desk workers, including a 520-person subgroup of HR professionals, fielded 26 June–15 July 2026 by Method Research and released 1 September (BambooHR, 2026). Respondents spanned more than a dozen industries, evenly split by gender, with a spread of age groups and geographies.
The headline split is the finding:
- 42% of AI time — troubleshooting errors and iterating on prompts
- 35% of AI time — productive work that furthers the employee's workload
- 23% — everything else
Hold the ratio rather than the percentages. For every three minutes your team spends getting value out of an AI tool, it spends four minutes getting the tool to produce something usable. That ratio has a cost, it has a distribution, and neither one appears on a seat-count report.
Meanwhile, 63% of organizations have already increased AI tool budgets. The spend is scaling against a denominator nobody has decomposed.
The Load Inverts Your Org Chart
Here is the part that should change where you look.
Daily minutes in AI tools, by level (BambooHR, 2026):
- VPs and C-suite: 101 minutes
- Managers and directors: 89 minutes
- Individual contributors: 54 minutes
Executives spend nearly twice as long in these tools as the people who report to them. If the troubleshooting share holds even roughly across bands, the largest pool of unproductive AI time in your company is being generated by its most expensive labour.
Run it at mid-market scale. A 200-FTE company with 15 people at VP level and above: 101 minutes a day, 42% of it troubleshooting, is about 42 minutes per executive per day. Across 15 people that is over 10 executive-hours a day — more than a full additional senior FTE, spent entirely on prompt iteration and error correction, funded out of the most expensive line in your payroll.
Nobody approved that headcount. It arrived through a tool budget.
Why Seniority Concentrates the Cost
The direction is not surprising once you look at what senior people use AI for. Their work is less templated, more judgment-dense, and more context-dependent — which is precisely the profile that produces more iteration, more correction, and more abandoned output. The tasks with the cleanest AI yield tend to be the structured ones further down the org chart.
There is supporting shape here from the euro area. The ECB's Consumer Expectations Survey found that the tasks with the highest AI time-yield are narrow and under-used — code generation and debugging return nearly eight hours a week but only about 8% of workers apply AI there — while the most common uses, research, information gathering, writing and editing, save considerably less (ECB, 2026). Broad, unstructured, high-context work is where the hours go in and the smallest yield comes out. That is a fair description of an executive's day.
Why Nobody Has Complained About It
Because it does not feel like waste. This is the part that keeps the problem invisible through an entire budget cycle.
BambooHR found 65% of workers feel confident and enthusiastic about using AI at work, and 58% name time savings as their main motivation — while sitting inside that same 42/35 split.
Culture Amp's first AI at Work benchmark shows the same detachment at scale. Across roughly 112,000 employees and 343,000 responses collected in the 12 months to July 2026, 71% of employees said AI tools help them feel more productive, rising to 93% among the heaviest users. Yet when asked whether their workload was reasonable, power users came in at 72% against 69% of non-users — a difference of three points (Culture Amp, 2026).
Felt productivity climbs 22 points with usage intensity. Felt workload relief moves three. Those two curves should travel together, and they do not.
Self-Report Is the Weak Instrument Here
The gap between perceived and measured effect is documented, and it runs in one direction. METR's randomized trial on experienced open-source developers found early-2025 AI tools made them 19% slower while participants believed they had been sped up; METR's later survey of 349 technical workers reports that participants in that study overestimated AI's effect on task time by about 40 percentage points (METR, 2026).
The UK government's own evaluations show the same compression under better method. The cross-government Microsoft 365 Copilot experiment — 20,000 civil servants — reported average savings of 26 minutes a day, a figure participants estimated themselves from a set of ranges (GDS, 2025). When DWP evaluated its own slice of that trial against a comparison group of non-users, controlling for demographics, occupation and AI-keenness, the estimate came in at 19 minutes a day across eight routine tasks (DWP, 2025).
Different populations and different task scope, so this is not a clean before-and-after. But it is two measurements of the same rollout, and the one with a comparison group is the smaller one. The GDS report is explicit in its conclusions that "due to experimental constraints it was not possible to identify how time saved was spent."
Your team's enthusiasm is real. It is not evidence.
AI Troubleshooting Time Breaks Your Adoption Metrics
Seat counts, licence utilisation and login rates all measure the 87 minutes. None of them distinguish the 42 from the 35. An adoption dashboard showing strong usage is, on this evidence, equally consistent with a team getting substantial leverage and a team fighting its tooling for 20 workdays a year.
Worse, the metric rewards the wrong movement. Rising minutes-in-tool reads as success. If the troubleshooting share is stable, rising minutes are mostly rising rework.
The ECB result adds the second correction. The median euro-area AI user saves about three hours a week, roughly 7.7% of working time — but because only 48.8% of workers both use AI and save time with it, the economy-wide gain is about 3.8% (ECB, 2026). Per-user figures multiplied by headcount overstate return roughly twofold. A business case built on "X minutes saved × everyone" is wrong twice: once on participation, once on the troubleshooting share inside the minutes it does count.
The Honest Counter
Three limits, stated plainly.
This is a vendor study. BambooHR sells HR software to small and mid-sized businesses and has a commercial interest in the argument that AI rollouts need people-side management. The fieldwork was run by an independent research firm, the methodology is disclosed, and the sample is reasonable — but this is a survey release, not a peer-reviewed paper. Weight it accordingly.
It is self-report, and so is most of what supports it. The 42/35 split is workers estimating their own time allocation, which is exactly the instrument METR showed to be unreliable in this domain. The honest reading: the direction is well-corroborated across BambooHR, Culture Amp and the UK trials, while the magnitudes are soft. Do not quote 42% as a measured constant. Quote it as the reason to measure your own.
The seniority split is a correlation, not a mechanism. BambooHR reports minutes by band; it does not report the troubleshooting share by band. My scaling above assumes the 42% holds roughly constant across levels. That assumption is plausible and untested — senior work may generate more iteration, or senior users may be more skilled at avoiding it. Treat the executive-hours figure as an order of magnitude to go verify, not a number to put in a board deck.
What to Decide This Quarter
Four moves. None require new software.
- Split the minute count before the next renewal. Your current metric is minutes-in-tool. Add one dimension: what share of that time is producing output versus producing a usable prompt. A two-week self-log across 20 people gets you a defensible internal number, and it is the only figure in this article that is about your company.
- Sample by role band, not by average. Pull the senior band separately. If the BambooHR shape holds in your organization, the expensive problem is concentrated in fifteen calendars, not two hundred — which also makes it tractable.
- Make troubleshooting share the adoption KPI. Replace login rate with the ratio. A rollout where minutes rise and troubleshooting share falls is working. One where both rise is a rework pipeline wearing an adoption dashboard.
- Name the destination for the hours before you count them. Both UK evaluations could measure time saved and neither could say where it went. A time-savings number without a named destination is not an ROI input. Decide what stops before you bank what starts.
The 63% who raised AI budgets this year did so against a number that counts the fighting and the working as the same minute.
Before you approve the next increment, find out which one you have been buying. Your most senior people have been paying for the answer all year — in AI troubleshooting time nobody put on the budget.
