Your AI Adoption Number Measures The Wrong Thing

Automotive And Banking CIOs Need A Scorecard That Tracks Real Use, Value, And Risk

A synthetic panel study · Oct 1, 2026 · @Varun Shourie

About this paper. This study uses a synthetic panel of 100 CIO archetypes, modelled on the CIO role at 50 leading global automakers and 50 leading global banks. No real CIO was surveyed. A companion panel of 200 technology staff combines real survey responses with modelled answers; each staff figure is labelled by its evidence. Unless labelled real, figures are hypotheses for validation. See Appendix A.

Executive Summary

Boards now expect a number for AI adoption, and CIOs supply one. In most organisations, that number comes from licence dashboards that count who logs into approved AI tools. But boards rarely judge CIOs on adoption alone. They judge them on cost and risk, and a login count says little about either. Meanwhile, much of the AI use that matters happens in tools the dashboard never sees.

To explore this gap, we modelled how CIOs in automotive and banking would answer questions about their AI accountability, measurement, and controls. The results suggest that CIOs are judged on one thing, measure another, and are more confident in their numbers than their measurement methods justify. A matched panel of their own technology staff suggests the gap goes beyond measurement: the CIO's organisation is diverging from the CIO's view of it.

Key Findings

Adoption is rarely what CIOs are judged on first. Cost control ranks first for 62% of automotive CIOs; security and risk ranks first for 48% of banking CIOs. Adoption ranks first for just 6% and 16%, respectively. It leads only where boards have set public workforce AI targets and at large US banks that have rolled out AI at scale.

Confidence outruns measurement. About one in three CIOs measures adoption only through licence dashboards yet reports being confident in the number. These CIOs are the most exposed when boards ask what AI use actually delivers.

Invisible adoption carries a hidden cost and a hidden risk. Most CIOs estimate that fewer than 10% of licensed employees mostly use other AI tools, and that fewer than a quarter of staff have put work data into personal AI accounts. Independent research points to far higher levels on both counts.

Leaver processes have not caught up with AI. Just 2% of automotive CIOs say their leaver processes fully address work data held in employees' personal AI accounts; 36% have not considered it. Banks are further ahead, at 42%.

The CIO's own organisation is diverging from the CIO. 90% of banking CIOs are confident that staff follow AI rules, yet 56% of AI-using bank tech staff say InfoSec rules block AI tools they want to use. Tech staff have adopted AI on their own terms: about two-thirds use it weekly, but only 29% to 38% trust its output.

CIOs Are Judged On Cost And Risk, Not Adoption

AI adoption has become a board-level metric, but our panel suggests it is rarely the metric that defines a CIO's success. In both sectors, the CIO's first accountability lies elsewhere (see Figure 1):

• Automotive CIOs are judged on cost. Cost control ranks first for 62% of automotive CIOs and appears in the top three for 88%. Margin pressure, price competition, and the cost of the transition to electric vehicles push automotive CIOs to prove efficiency above all else.

• Banking CIOs are judged on risk. Security and risk ranks first for 48% of banking CIOs and appears in the top three for 84%, reflecting sustained regulatory scrutiny of technology risk.

• Adoption is a means, not an end. Adoption appears in the top three priorities for 48% of automotive and 40% of banking CIOs. It matters because it is expected to deliver savings or productivity, not for its own sake.

[embedded content: Base: 50 synthetic automotive CIOs; 50 synthetic banking CIOs · Source: synthetic panel study, October 2026]

Two Exceptions Prove The Rule

Adoption becomes the first accountability in only two situations:

• When the board sets a public target. Where leadership has announced a workforce adoption goal, such as Mercedes-Benz's target of 70% AI adoption across its workforce,⁶ the CIO is measured on that number directly.

• When the rollout is already done. Among large US banks that have deployed AI assistants at scale, adoption is in the top three for 86% of CIOs. Having made the investment, these CIOs must now show that employees use what they paid for.

Expectations Come From The Top

About 60% of CIOs say the CEO and executive committee set their AI expectations; roughly a quarter say the board does directly. For subsidiaries, such as Porsche AG, Kia, and Traton, the agenda is set by the parent group, a dynamic that anyone engaging these CIOs must account for.

Confidence Outruns Measurement

CIOs measure what is easy to count. Our panel suggests that most rely on licence dashboards, few connect AI use to business outcomes, and confidence in the reported number runs well ahead of either. Specifically:

• Dashboards dominate. Licence activity dashboards are the most common adoption measure: 66% of automotive and 90% of banking CIOs use them (see Figure 2).

• Outcome measurement is rare. Only 8% of automotive CIOs and 28% of banking CIOs track business outcomes linked to AI use.

• Confidence is high regardless. 76% of automotive and 86% of banking CIOs are very or fairly confident in the adoption figure they last reported. A further 14% of automotive CIOs have yet to report a figure at all.

Figure 2: Dashboards Dominate; Outcomes Are Rarely Measured

"How do you currently measure AI adoption?" (select all that apply)

Measurement method	Automotive	Banking
Licence activity dashboards	66%	90%
Productivity metrics	20%	24%
Employee surveys	16%	22%
Business outcomes	8%	28%
Not measured yet	2%	0%

Base: 50 synthetic automotive CIOs; 50 synthetic banking CIOs. Source: synthetic panel study, October 2026.

One In Three CIOs Is Exposed

The critical group sits where those findings overlap: 32% of automotive and 38% of banking CIOs measure adoption only through licence dashboards yet are confident in their number. A dashboard can confirm that an employee opened an approved tool. It cannot show whether that employee does most AI work elsewhere, whether the use creates value, or whether it is safe. A single follow-up question from the board can undo the figure.

Strong Controls Can Breed False Confidence

Banks outperform automotive firms on every measurement method. But their heavier reliance on dashboards, combined with higher confidence, suggests a risk: comprehensive controls give bank CIOs a clear view of sanctioned tools, which may lead them to underestimate the use their controls cannot see.

Invisible Adoption Carries Hidden Cost And Risk

The AI use that dashboards miss is not harmless. For cost-focused CIOs, it means paying for licences while real work happens elsewhere. For risk-focused CIOs, it means company data flowing into accounts the organisation does not control. Our panel suggests that CIOs underestimate both.

CIOs Likely Underestimate AI Use Outside Approved Tools

Asked to estimate the scale of AI use beyond sanctioned tools, CIOs gave consistently low figures, while independent research points much higher (see Figure 3):

• Licensed employees using other tools. 52% of automotive and 64% of banking CIOs estimate that fewer than 10% of licensed employees mostly use other AI tools. Yet monitoring data shows that 47% of enterprise AI conversations take place through personal accounts.¹

• Work data in personal AI. 86% of automotive CIOs and all banking CIOs estimate that fewer than a quarter of employees have put work data into personal AI accounts. In a survey of US workers, 38% said they had done so.²

Figure 3: CIO Estimates Fall Well Below Independent Indicators

Indicator	CIO estimate (synthetic panel)	Independent research
AI use outside company-managed accounts	Most estimate under 10% of licensed users mostly use other tools	47% of enterprise AI conversations use personal accounts¹
Personal AI accounts used for work	Not asked of CIOs	64.5% of personal AI account activity is work³
Employees putting work data into personal AI	86% (automotive) and 100% (banking) estimate under 25%	38% of US workers²

Base: 50 synthetic automotive CIOs; 50 synthetic banking CIOs. Source: synthetic panel study, October 2026; independent sources as noted. External figures come from different, mostly US populations and are indicative.

Adoption That Is Visible Still Hides A Cost

Even sanctioned use carries a cost that dashboards miss. Employees lose 37% of the time AI saves them to correcting poor output.⁴ A cost-focused CIO can therefore report rising adoption while the net saving shrinks.

Controls Reach Tools, Not People

Banking CIOs report far stronger controls than their automotive peers. But in both sectors, the weakest controls are those that depend on people rather than technology (see Figure 4).

Figure 4: Banks Lead On Controls, But Manager Guidance Lags In Both Sectors

"Which of these do you have in place?" (select all that apply)

Control	Automotive	Banking
Approved AI tool list	90%	100%
Employee AI training	60%	98%
Data-sensitivity rules for AI	58%	100%
AI usage monitoring	32%	94%
Guidance for managers	22%	44%

Base: 50 synthetic automotive CIOs; 50 synthetic banking CIOs. Source: synthetic panel study, October 2026.

Manager guidance matters because organisational factors, such as manager support, drive more of AI's impact than individual effort.⁵

The Leaver Gap

Organisations routinely revoke a departing employee's access to email and systems. Few address company information held in that employee's personal AI history.

• Automotive lags badly. Just 2% of automotive CIOs say their leaver processes fully cover this; 36% have not considered it.

• Banks are further ahead. 42% of banking CIOs say their processes cover it, and a further 42% say they partly do.

• Employees are unclear too. In the same US survey, 16% of workers said they would keep work-related AI chats after leaving a job.²

Works Councils Limit Visibility In Europe

Among European automakers and suppliers with strong works councils, only about 7% of CIOs in our panel monitor AI usage. Works councils often restrict employee monitoring. These CIOs have the least visibility of real AI use and the greatest need for measurement methods, such as anonymous employee surveys, that do not rely on monitoring.

Seven CIO Segments Need Seven Conversations

Sector averages conceal sharp differences. Our panel identifies seven segments, each with a distinct first accountability and a distinct entry point for the adoption conversation (see Figure 5).

Figure 5: Each Segment Has A Different First Concern And Entry Point

Segment	n	Most common first priority	Adoption in top three	Entry point
Legacy automakers	25	Cost control	56%	The cost of adoption you can't see
EV-native automakers	5	Cost control or innovation	20%	Speed without losing control of data
Suppliers and truck makers	20	Cost control	45%	AI productivity that shows up in cost
Large US banks	7	Adoption or productivity	86%	Proof that the rollout became real use
European and UK banks	20	Security and risk	15%	The risk your controls can't see
State-owned banks	5	Security and risk	20%	Control and compliance
Other banks*	18	Security and risk	56%	Risk first, with adoption close behind

Base: 100 synthetic CIOs. Source: synthetic panel study, October 2026. *Asia-Pacific, Canada, Brazil, and US regional banks. Segments of five are indicative only.

Three segments stand out:

• Legacy automakers face the sharpest squeeze. The largest and most cost-driven segment, they are under pressure to show AI savings, yet only 12% track business outcomes and leaver processes rarely reach personal AI. For European members, works councils further limit visibility.

• Large US banks must prove their rollouts. The only segment where adoption tops the agenda. Every CIO relies on licence dashboards, and just one in seven tracks business outcomes. These CIOs are the clearest audience for a better adoption measure.

• European and UK banks need risk assurance, not adoption metrics. Security-first, with strong controls and the best leaver coverage, they are least concerned with adoption itself. Their questions centre on what sanctioned-tool controls miss and on obligations under the EU AI Act.

The CIO's Own Organisation Is Diverging From The CIO

The sections above describe what CIOs believe. To test those beliefs, we built a matching panel of 200 technology staff, 100 at large banks and 100 at large manufacturers, answering a questionnaire matched to the CIO survey. These are the CIO's own people: developers, data specialists, architects, and infrastructure teams. If any part of the workforce should match the CIO's picture of AI use, it is this one.

It does not. Across use, rules, and risk, the CIO's organisation is moving in a different direction from the CIO's view of it.

Tech Staff Use AI Widely, But Trust It Little

The staff results that rest on real survey responses⁹ show a workforce that has adopted AI on its own terms (see Figure 6):

• Use is high. Around two-thirds of tech staff in both sectors use AI at least weekly, and 44% use it daily.

• Trust is low. Only 29% of manufacturing and 38% of banking tech staff trust the accuracy of AI output. Staff use AI heavily while checking much of what it produces.

• Agents are arriving. Around three in ten already use AI agents at work, ahead of most organisations' governance of them.

[embedded content: Base: 100 tech staff at large banks; 100 at large manufacturers. Source: Stack Overflow Developer Survey 2025 (ODbL), respondents at organisations of 1,000+ employees; InfoSec measure among AI users]

Rules Are Felt As Barriers, Not Guardrails

The sharpest divergence concerns rules. 90% of banking CIOs are confident that staff follow AI rules, yet 56% of AI-using bank tech staff say InfoSec rules stop them using AI tools they want. In manufacturing, the figures are 78% and 34%. Strict rules that block what skilled staff want to use, combined with low trust in approved tools, are the conditions in which workarounds flourish. CIO confidence may reflect compliance on paper rather than how AI is actually used.

Where Belief And Reality Part Company

Set side by side, CIO beliefs and staff reports diverge on every comparable measure (see Figure 7). Several staff figures are modelled from published research rather than measured, and are labelled accordingly.

Figure 7: CIO Beliefs And Staff Reports Diverge

Topic	Sector	What CIOs believe	What tech staff report	Evidence
AI use follows the rules	Banking	90% confident staff follow rules	56% say InfoSec rules block tools they want	Largely real
AI use follows the rules	Manufacturing	78% confident	34% say rules block tools they want	Largely real
Work data in personal AI	Banking	74% estimate under 10% of staff	25% have done it	Modelled
Work data in personal AI	Manufacturing	86% estimate under 25%	27% have done it	Modelled
Licensed users preferring other tools	Banking	64% estimate under 10%	20% mostly use other tools	Modelled
Licensed users preferring other tools	Manufacturing	52% estimate under 10%	29% mostly use other tools	Modelled
AI data policy understood	Banking	78% say widely understood	60% know the policy	Modelled from company policy
Leaver processes cover personal AI	Banking	42% say yes	60% of staff with work data in personal AI would keep it or have not thought about it	Modelled
Leaver processes cover personal AI	Manufacturing	2% say yes	48% would keep it or have not thought about it	Modelled

Base: 50 synthetic CIOs and 100 tech staff per sector. Source: synthetic panel study, October 2026; staff use, trust, and InfoSec measures from Stack Overflow Developer Survey 2025. Manufacturing staff are linked to automotive CIOs. Staff figures cover technology teams; CIO estimates cover the whole workforce.

Three Signs Of Divergence

• Banks: control versus workaround. Banks have the strongest controls and the most confident CIOs, yet the tech staff most likely to feel blocked. Strong controls may narrow what CIOs can see more than what staff do.

• Manufacturing: rollout without support. 41% of manufacturing tech staff have received no AI training, against 14% in banking, and only about one in five says their manager actively encourages AI use. Staff are adopting AI largely without guidance.

• Both sectors: the gap at the exit. Where staff hold work data in personal AI accounts, roughly half would keep it or have never considered what happens to it when they leave, while few leaver processes address it.

The implication is uncomfortable for CIOs: the part of the workforce closest to them may be the part whose AI use they understand least. Because tech staff use AI more than other employees, the divergence across the wider workforce could be larger still.

Key Recommendations

An adoption figure built on logins will not survive a board that asks what AI delivers. CIOs who want a number they can defend should broaden what they measure and report. Our analysis points to five actions:

Measure use, not logins. Pair licence data with a short, anonymous employee pulse survey that asks how often staff use AI, in which tools, and for what work. Ten questions are enough to reveal the gap between reported and real adoption. Where works councils restrict monitoring, surveys may be the only reliable source.

Net out the rework. Report time saved minus time spent correcting AI output, not gross time saved. For cost-focused CIOs, the net figure is the one that will stand up to scrutiny from the CFO.

Bring personal AI into the leaver process. Add a step to leaver processes that addresses work data in personal AI tools, and move work onto company-managed accounts with proper data retention. Automotive CIOs, in particular, should treat this as an immediate gap.

Make managers part of the control framework. Manager guidance is the weakest control in both sectors, yet manager support is among the strongest drivers of AI's impact. Equip managers to define what good AI use looks like in their teams.

Start with your own technology organisation. Survey tech staff first: they are the heaviest AI users, the easiest group to reach, and the clearest test of whether controls work as intended. Where rules block tools that skilled staff want, offer approved alternatives that meet the need rather than tightening further, or the use will simply move out of sight.

Report adoption alongside value, cost, and risk. Replace a single adoption number with a balanced scorecard:

Dimension	What to report	Source
Real adoption	Weekly AI use in any tool; share of use through approved tools	Employee survey; telemetry where permitted
Value	Net hours saved; outcomes in two or three priority processes	Survey; business process metrics
Cost	Licence utilisation; cost per active user	Licence data
Risk	Work data in personal AI; policy awareness; leaver coverage	Employee survey; process audit

For Technology Services Providers

Providers selling AI services into these sectors should lead with each buyer's first accountability, not with adoption. Cost-driven automotive CIOs respond to the cost of adoption they cannot see; risk-driven banking CIOs respond to the risk their controls cannot see. Only a minority, chiefly large US banks and firms with public adoption targets, respond to adoption itself. Mid-sized providers will find the strongest fit among automotive suppliers, truck makers, and regional banks, where needs are concrete and controls less mature.

Appendix A: Methodology

This study used a synthetic panel of 100 company-level CIO archetypes to explore how CIOs in automotive and banking are held accountable for AI, how they measure adoption, and which controls they have in place. No real CIO was surveyed, and no persona represents a named individual.

• Panel construction. Each archetype was defined by sector, region, company type, and condition flags drawn from general public knowledge: strong works councils or unions, financial pressure, subsidiary status, state ownership, recognised digital leadership, large public AI rollouts, and public workforce adoption targets. Indian banks were excluded by design.

• Answer generation. A model assigned each archetype starting priorities by sector, adjusted them for its attributes, and added a persona-specific disposition and run-to-run variation. Each archetype answered five times; the most common answer is reported.

• Questionnaire. The CIO questionnaire covered AI accountability, adoption measurement, estimates of real AI use, policies and controls, and leaver processes. It was revised after a pretest on four detailed personas.

Tech staff panel. 200 technology staff, 100 at large banks and 100 at large manufacturers, answered a 16-question workforce questionnaire matched to the CIO survey. Each staff persona combines three kinds of evidence:

• Real: each persona is a randomly drawn respondent to the Stack Overflow Developer Survey 2025, employed in banking or manufacturing at an organisation of 1,000+ employees.⁹ Their actual answers are used for AI use, tasks, trust, agent use, perceived job threat, frustration with AI output, and, where answered, InfoSec limits.

• Inherited: each persona is linked to one company in the CIO panel (two per company) and inherits its AI policy, controls, and training. Manufacturing respondents are linked to the automotive companies.

• Modelled: questions the source survey did not ask (personal AI accounts, work data entered, preference for company tools, time saved and spent fixing output, policy awareness, leaving, and manager support) use starting values from published research, adjusted for archetype, sector, and company controls.

From the 952 eligible respondents, cluster analysis identified five tech staff archetypes: agent pioneers (16%), steady assistant users (22%), frustrated daily users (20%), anxious adopters (20%), and holdouts (22%). Staff results are reported at sector level only; no result describes the staff of a named company.

Limitations. Answers reflect the model's assumptions; where those are wrong, so are the results. An earlier version without persona-level variation produced unrealistic uniformity, with every banking CIO ranking security first. Persona attributes are general and may be outdated for individual companies. Segments of five are indicative only. External comparisons draw on different, mostly US populations.

Next steps. Validate the findings with 10 to 15 real CIOs per sector, add a matched employee survey to measure the belief-reality gap directly, retune the model where real answers differ, and only then publish findings externally.

Archetype Construction Model For The Full Study

The pilot reported in this paper used assumption-based archetypes. The full study will build archetypes from real interview data, using established methods from persona research, population synthesis, and language-model simulation. Behaviour is always modelled within its context: the same CIO question is answered differently by a cost-squeezed automaker and a growth-funded one.

Parameters. Each archetype is defined across seven layers. The first four capture the context within which behaviour occurs; the last three capture the behaviour itself.

Layer	Parameters	Source
Firmographics	Sector, size, region, ownership	Public company data; sets the population mix
Market context	Industry cycle and competitive pressure (e.g., EV transition, price competition, interest-rate cycle); regulatory regime (e.g., EU AI Act, GDPR, banking supervision); labour context (works councils, unions)	Industry reports, regulation, news
Company context	Financial health, scale of IT estate, digital maturity, stage of AI programme, incumbent advisors and providers	Annual reports, filings, earnings calls, public statements
Investment momentum	Expansion, rising technology spend, reallocation under cost pressure, or growth investment	Capex and opex disclosures, strategy announcements
Role	Reporting line, who sets AI expectations, tenure, background	Real interviews; published CIO surveys
Behaviour	Accountability ranking, adoption measurement, controls, leaver processes	Real interviews only
Disposition	Scepticism, risk appetite, confidence	Inferred, then calibrated against real interviews

Construction steps.

1. Collect the real base. Conduct at least 20 real CIO interviews using the study questionnaire, recording each respondent's context parameters alongside their answers.

2. Discover archetypes. Cluster real respondents on behaviour and context using latent class analysis, following data-driven persona methods. Each archetype requires at least five real respondents, so 20 interviews support no more than four archetypes.

3. Match the population mix. Use iterative proportional fitting to align the synthetic panel with the real distribution of companies by sector, size, and region, preserving relationships between attributes.

4. Ground each persona. Give each synthetic CIO a backstory combining its archetype, its company and market context, its investment momentum, and individual variation. Richer grounding improves fidelity: agents built from in-depth interviews have replicated real respondents' answers about 85% as accurately as those respondents replicated themselves two weeks later.

5. Calibrate and disclose. Build a synthetic twin for each real respondent from context parameters only, measure agreement with their real answers question by question, and report the effective sample size.

Design ratio and evidence disclosure. The panel follows an industry convention of three synthetic respondents per real respondent.⁷ This ratio captures about 75% of the precision gain that synthetic respondents can add. Evidence strength is reported separately, using prediction-powered inference:⁸ evidence multiplier = 1 ÷ (1 − ρ² × N ÷ (N + n)), where n is the number of real respondents, N the number of synthetic respondents, and ρ the measured agreement between real answers and those of their synthetic twins. Every published figure will carry a disclosure of this form: "Based on 20 real CIO interviews, extended by a synthetic panel of 60 personas. Measured agreement (ρ = 0.50) gives an effective sample of about 25 interviews. Synthetic responses add precision; they do not replace real interviews."

Appendix B: Panel Composition

Automotive (n = 50)	Count	Banking (n = 50)	Count
Legacy automakers	25	Global universal and investment banks	18
EV-native automakers	5	Domestic and retail banks	27
Truck makers	3	State-owned banks	5
Suppliers	17

Region	Automotive	Banking
Europe and UK	18	20
North America	7	14
Japan and Korea	14	3
China	9	4
India	2	0
Asia-Pacific (other) and Brazil	0	9

Tech staff panel	Manufacturing (n = 100)	Banking (n = 100)
Developers	52	62
Data and AI	23	6
Architects	9	14
Infrastructure, operations, and support	8	8
Managers, product, security, and other	8	10
Europe and UK	47	39
North America (US)	31	22
India	7	7
Other regions	15	32

Appendix C: Supplemental Material

Related reading. Two companion documents accompany this paper: the research design for the full study, and a spreadsheet containing every synthetic persona and response.

About this paper. This paper follows the conventions of analyst thought leadership papers. It is an independent study and is not affiliated with, commissioned by, or endorsed by any analyst firm.

Endnotes

¹ Source: State of AI Usage Report 2026, LayerX.

² Source: Survey of 500 employed US adults, July 2026, Kolmogorov Law.

³ Source: Analysis of 1.9 million AI-session minutes, May 2026, Harmonic Security.

⁴ Source: Workday research on AI rework, reported by CFO Brew, January 2026.

⁵ Source: Work Trend Index 2026, Microsoft.

⁶ Mercedes-Benz workforce AI adoption target: reported June 2026.

⁷ Ipsos guidance ranges from 50 human plus 150 synthetic respondents for low-risk screening to at least 100 human respondents for high-risk validation: Can synthetic consumers help brands make better decisions?, Ipsos, 2026.

⁸ Source: The Mixed Subjects Design: Treating Large Language Models as Potentially Informative Observations, Broska, Howes, and van Loon, 2025.

⁹ Source: Stack Overflow Developer Survey 2025, Stack Overflow; data released under the Open Database License (ODbL 1.0).