How we read the signal

Analysis frame

Evidence level

Primary-source evidence

Analytical lens

Track how the financial, data, and reputational costs of informal AI adoption are distributed between workers and employers.

Affected groups
  • Workers paying for AI tools or using them without approval
  • Employers receiving productivity gains and carrying data risk
  • Customers whose information may enter external systems
  • IT, security, legal, and labor teams responsible for policy
What remains unknown
  • Whether self-reported spending matches audited subscription payments
  • How much reported time saving becomes measured productivity or longer workloads
  • Which kinds of sensitive data enter unapproved tools
  • How usage and employer support vary by income, role, and industry
Second-order effects to watch
  • AI access may widen workplace inequality between people who can and cannot pay
  • Prohibition without alternatives may push adoption deeper into the shadows
  • Employers may capture more output without sharing savings or reducing workload
  • Clear reimbursement and training could convert hidden risk into governed adoption

The workforce is financing adoption

Deloitte estimates £958 million in annual personal spending on work-related generative AI. Seventeen percent of users paid for at least one tool themselves, while 46 percent used free tools and 34 percent used externally supplied tools funded by employers.

The figures come from a weighted self-report survey, so they estimate behavior rather than audit it. The scale nevertheless challenges the idea that enterprise adoption begins with an executive procurement decision.

Time savings do not settle who benefits

Respondents reported saving an average of 70 minutes each week, and Deloitte says most used that time to do more work for the same employer. Sixty-four percent of weekly users feared managers might conclude that AI could perform their jobs.

Productivity can rise while worker bargaining power falls. Measuring output without measuring workload, compensation, and job design hides the distributional effect.

Govern the tool instead of the secret

Employers should publish approved services, prohibited data, verification duties, reimbursement rules, and an incident channel that does not punish honest reporting.

Blanket bans can preserve appearances while moving use onto personal accounts. A governed alternative makes both the benefit and the responsibility visible.

Primary trail

Go to the source

Read the evidence behind this analysis. External links open in a new tab.

Deloitte — British workers spend their own money on generative AI for work