Your team got faster at the desk. Drafts appear in minutes. Research that used to eat a morning now fits in a coffee break. Someone will tell you AI saved them half a day last week, and they are not wrong.
Then look at the rest of the week.
Someone is pasting the same brief into three chat windows. Someone is reconciling three summaries of the same numbers. Someone is babysitting a "smart" agent that still needs a human to carry context between systems that never agreed on a source of truth.
The hours are real. The handoff tax takes them back.
The personal win and the company miss
Glean's Work AI Index 2026 surveyed 6,000 digital workers in the US, UK, and Australia. People reported that AI automation saved about 11 hours a week. They also spent 6.4 hours a week botsitting: feeding context into tools, checking outputs, debugging failures, and cleanup.
Of total AI time in that study, 37% went to botsitting, 36% to actual production work with AI, and 27% to learning or building agents. Almost four in ten AI sessions failed hard enough to need a full restart or serious rework.
Here is the gap that should bother any owner or ops lead. About 87% of workers use AI. About 75% say it makes them more productive. Only 13% say the organization is performing significantly better because of it.
Individual speed. Organizational shrug. That is not a model problem. That is a system problem.
Six tools that still do not talk
Tool sprawl is how the handoff tax scales.
In the same Glean work, 77% of AI users juggle multiple AI tools each week. A third run four or more. Multi-tool users were 35% more likely to report frequent botsitting. Sixty percent re-ran the same prompt across tools because the first answer was not good enough.
Small businesses are not immune. The SBE Council's March 2026 tech survey of 517 employers (2 to 99 employees) found 82% had adopted at least one AI tool. The median stack was five AI tools, with plans to add more.
Five tools sounds modern. Five tools with no shared memory is a part-time job.
I have watched this pattern for years in client ops and in our own shop. The stack grows by good intentions. Each department adds a tool for a real pain. Nobody owns the inventory, so nobody owns the handoffs.
A professional services firm of about 110 people learned that the expensive way. Field consultants counted AI charges across four cards: competing chat assistants, a writing tool, a meeting note-taker, a forgotten sales-coaching beta, a personal image generator on expenses, a legal review tool under evaluation. Roughly $1,840 a month. Zero-page AI policy. Nobody had a full inventory until someone counted it.
That is not laziness. That is ownership failure dressed up as innovation.
Botsitting is not always waste
The wrong takeaway is "stop checking the AI."
Glean found high AI performers spent more of their AI time on supervision (about 40% versus 33% for low performers). They caught more errors.
Productive botsitting is judgment: high-stakes verify, reject garbage, decide when the model should stay out of the room. Unproductive botsitting is reloading the same brief six times because none of your tools share a system of record.
One story from the research captures the second kind. A consultant bought a scheduling agent expecting calendar freedom. What he got was a second job: rephrasing requests, blocking lunch and travel, reintroducing the bot in every thread, apologizing after three duplicate client invites. After a year, the bot kept only low-stakes scraps.
When the human becomes the integration layer, you did not automate work. You renamed it.
The tax was already there
Before the AI wave, teams already paid a toggle tax. HBR reported workers averaging about 1,200 app and website switches a day, with roughly four hours a week lost reorienting after each jump.
AI did not invent context switching. It stacked agent silos on top of app silos.
Salesforce and MuleSoft's 2026 Connectivity Benchmark put the next stage in plain numbers. Organizations averaged about 12 AI agents. Half of those agents ran in isolated silos. Eighty-six percent of IT leaders feared agents would add more complexity than value without proper integration.
More agents without a shared record is not progress. It is a bigger room full of people who cannot hear each other.
Quality has a tax too. In Glean's work, 69% of AI users admitted shipping AI work they had not verified or could not defend. Fluency hides risk. A polished wrong answer is more dangerous than a clumsy right one.
Workflow automation ROI is not "hours saved"
If your ROI slide only shows desk speed, you are measuring the easy half.
Real workflow automation ROI includes supervision cost, integration cost (people acting as the API), and rework cost. Glean's split is the cleanest version: 11 hours claimed, 6.4 hours clawed back in botsitting, and only 13% of people seeing the org get markedly better. That is not "AI failed." That is automation without a system of record.
A system of record is not another dashboard. It is authority. Which number is final. Which SOP is current. What "Q3" means here. Who owns the tool list. Where the brief lives so five tools do not get five slightly different stories.
Glean also reported that 53% of workers said critical job information was not accessible through their AI systems. In context-rich orgs, people spent less AI time botsitting, shipped less unverified work, and reported less exhaustion from AI. Integration pipes move data. A system of record moves authority.
Start with ownership, not another subscription
You do not need a new platform to start. You need ownership.
List every AI tool your company pays for, tries, or reimburses. If you cannot list them in one sitting, that is the finding.
Pick one owner. One person who can say "this is in, this is out, this is the source of truth for X." Committees invent more tools. Owners kill quiet duplicates.
For each workflow that touches AI, name the input, the tool, the output, and the human who still carries context to the next step. Those humans are your current integration layer. That labor is the automation handoff tax.
Keep judgment on high-stakes work. Kill any process where the main job is retyping the same brief into tools that do not share memory. Measure net time: hours generated minus hours spent feeding, reconciling, and redoing. If the org still feels the same, the stack is theater.
This is the work we do with clients under LTFI: not "add more bots," but build the operating layer so people stop being the API. The point is fewer handoffs with clearer ownership.
If your team is faster at the keyboard and still drowning in reconciling tabs, the problem is not effort. It is the missing system of record.
We help small businesses and agencies clean that up. First conversation is free. No commitment. kief.studio/contact