Shadow AI is work done in AI tools your organization did not provision and cannot observe. It is usually filed as a security and compliance issue, and that framing is correct as far as it goes. It is also incomplete. The more immediate operational damage is that shadow AI quietly invalidates the AI return figure you are about to present, because every one of those figures is computed from activity inside tools you administer. Work that moves outside that perimeter does not show up as a gap. It shows up as nothing at all.
Two things are true at once. AI-assisted work is spreading fast enough that most organizations now assume it is happening everywhere. And the instruments they use to measure it are pointed at only the portion of it that stayed inside tools they bought.
What does Gartner's 2026 service data actually show?
Customer service is the first function where the migration has been measured cleanly in public, on the customer side, and the numbers are unusually blunt.
| Finding | Figure | Source and sample |
|---|---|---|
| Third-party GenAI versus company chatbots | Customers roughly 3x more likely to use third-party GenAI tools | Gartner, 3,566 B2B and B2C customers, Feb to Mar 2026 |
| Direction of travel | Third-party use nearly doubled in a year; company chatbot use statistically unchanged since 2022 | Gartner, same survey |
| AI budget share | Median 12% of 2025 budget, highest of 10 functions assessed | Gartner, 1,303 senior leaders, Jan to Apr 2026 |
| Financial return | Only 24% of service and support leaders demonstrated positive returns across AI use cases | Gartner, same leader survey |
| Human access | 87% of customers say an option to reach a human agent is essential | Gartner, 3,566 customers |
Read the third and fourth rows together. The function that put the largest share of its budget into AI is the function where roughly three quarters of leaders did not demonstrate a positive financial return. That reads like a technology failure. Gartner's own analysis says it is not.
"Customers are embracing GenAI in both life and work, but so far, that has not translated into growth in the use of company-provided customer service chatbots," said Eric Keller, Sr Director Analyst in the Gartner Customer Service & Support Practice. "Instead, GenAI is shifting some service interactions outside of company-owned channels."
The spend and the return are measuring different populations
The investment was made in a channel. The behavior moved to a different channel. The return was then calculated over the first one. Nothing in that sequence is dishonest, and every step of it is defensible on its own. The output is still wrong, because the denominator quietly stopped describing the work.
What the deflection number does not say
A chatbot containment or deflection rate answers one question: of the contacts that entered this channel, how many resolved here. It cannot answer the question finance is actually asking, which is whether total human effort per customer outcome went down. If a customer resolves an issue in a third-party assistant and never contacts you, the channel records no interaction, no deflection, and no saving. If the same customer half-resolves it and arrives at a human agent with a more complicated residual problem, the channel records one contact that looks exactly like any other.
Why the same pattern is forming inside the workforce
The customer-side data is the clean version because someone finally surveyed the customers. The employee-side version is harder to see, which is precisely why it is worth naming before it becomes the standard finding of 2027.
Firm-level metrics cannot see task-level substitution
In NBER Working Paper 34836, Yotzov, Barrero, Bloom, Bunn, Davis and colleagues surveyed nearly 6,000 senior business executives at firms in the United States, United Kingdom, Germany and Australia. 69% of firms actively use AI. Across the four surveys, 89% of executives reported no impact on labor productivity over the past three years. The paper defines productivity here as the volume of sales per employee.
Sales per employee is a real number and a poor instrument for this question. It aggregates over every task in the firm, and it moves for a dozen reasons that have nothing to do with AI. A tool can absorb a meaningful share of a team's routine work and leave that ratio untouched for a year. The absence of a signal at the firm level is not evidence of absence at the task level. It is evidence that the measurement is happening at the wrong altitude.
The work that leaves first is the work that was easiest to codify
The distributional finding in Generative AI at Work by Brynjolfsson, Li and Raymond is more useful here than its headline. Across 5,179 customer support agents, access to a generative AI assistant raised productivity, measured as issues resolved per hour, by 14% on average and 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers.
That is not a uniform lift. It is a lift concentrated where the work was most codifiable. Codifiable work is also the work most likely to migrate out of a sanctioned workflow, because it needs the least context to complete somewhere else. The residual work is denser, more exception-driven, and more dependent on the people who were already carrying the hard cases. Headcount models built on volume reduction do not survive that shift, and adoption dashboards built on seat counts cannot detect it.
This is an old problem wearing new clothes
In our 2024 research conversations across organizations from 400 to 156,000 employees, the constant was not a missing tool. It was that work state lived across several systems that could not see each other, so the only way to answer a question about capacity was to reconcile by hand. A CFO at a private-equity-backed IT services holding company described three days of manual reconciliation per cash flow forecast because every acquired agency kept its own stack. Shadow AI adds one more unreconciled system, except this one keeps no ledger at all.
What a defensible internal number requires
The instinct is to close the gap by policing tools. That is a race against every consumer AI release, and it loses. The alternative is to stop measuring the tool and start measuring the work.
| Tool-instrumented measurement | Work-observed measurement | |
|---|---|---|
| Unit | Seats, sessions, prompts, licenses | Work items completed, cycle time, rework, exception mix |
| Source | Vendor admin consoles | Systems where output lands: tickets, repos, CRM, documents |
| Blind to | Any activity outside provisioned tools | Activity that produces no output artifact anywhere |
| Fails when | Work migrates to unsanctioned tools | Output is genuinely unrecorded |
| Answers | Did people use what we bought | Did the work change, and for whom |
Neither column is complete. The second one degrades more gracefully, because output tends to land somewhere even when the assistance did not.
Three questions worth asking before the next AI business case
- What share of the work in scope produces an artifact in a system we already read? That share is your measurable base, and if it is small, say so before you quote a return.
- Do we know what the work looked like before the tool arrived? Without a baseline there is no comparison, only a level.
- Are we comparing adopting teams to genuinely comparable non-adopting teams, or to the company average? The company average is not a control group.
The limit worth stating out loud
None of this establishes that AI produced a change. Observational workforce data does not support that claim, and the vendors implying otherwise are writing checks their methodology cannot cash. What it supports is controlled cohort analysis with confidence scoring: compare adopting teams to comparable non-adopting teams, control for tenure, role, and tool stack, attach a confidence score, and publish the limits next to the estimate. Levos states this plainly on its measurement methodology page, because a number a CFO can interrogate is worth more than a number that cannot survive one.
That is also why the AI Impact signal family treats adoption as a signal to be derived from work rather than read off a license report, and why unsanctioned tool use and license waste show up together in SaaS Stack Intelligence. They are two readings of the same underlying fact: the map of tools you pay for is not the map of where work happens.
Frequently asked questions
What is shadow AI? Work performed in AI tools the organization did not provision, does not administer, and cannot observe. The defining property is that the activity leaves no record in the systems being measured.
Why is shadow AI a measurement problem and not just a security problem? Because AI return figures are computed from sanctioned-tool telemetry. Work that migrates outside that perimeter leaves the denominator, producing a number that is internally consistent and externally wrong.
What did Gartner find about third-party AI use in customer service? Customers are roughly three times more likely to use third-party GenAI than company chatbots, third-party use nearly doubled in a year, and only 24% of service and support leaders demonstrated positive financial returns despite the highest AI budget share of ten functions.
How do you measure AI adoption that happens outside sanctioned tools? By measuring the work rather than the tool, using the systems where output lands, and comparing adopting teams to comparable non-adopting teams.
Can any platform prove AI caused a productivity change? No. Controlled cohort analysis with confidence scoring, with limits disclosed, is what observational data honestly supports.
Where Levos fits
Levos is a Human Capital Operating System: an intelligence layer above the existing stack. Seat counts are the first instrument shadow AI defeats; see AI license utilization. For how to read vendor-supplied returns, see vendor AI ROI claims.
Ready to find out what your instruments are missing?
Most AI measurement programs are not wrong about the numbers they report. They are silent about the work those numbers never touched. The first useful exercise is establishing what share of in-scope work you can actually observe today.
See how the AI Impact signal family derives adoption from work
Levos is accepting design partner applications from US organizations of 150 employees or more with an active AI rollout. Large organizations typically start with one function or division, measured against comparable teams that have not adopted yet. That is a stronger attribution design than a company-wide before-and-after.
