GLT Research

Research 02 ·

AI fluency is the next enterprise requirement

Models produce more. People direct, judge, and take responsibility.

Models can do more of the work. People still decide what to delegate, explain the task, assess the result, and take responsibility. Enterprises need a shared standard for these skills across hiring, daily work, and performance reviews.

Presentation deck

Slide 01 of 10·Title

01 Title
02 Thesis
03 AI code share
04 4D fluency
05 Coinbase loop
06 Hiring spectrum
07 Shopify baseline
08 Multi-mode
09 Practice
10 Conclusion

Open a slide in full screen to present. Use the arrow keys to move between slides.

Presentation deck

Research 02 · Executive briefing · 2026-08-13

AI fluency is the next enterprise requirement

Gary@gaplo.tech

Published on · ai.gaplo.tech

Models now generate code, tests, and implementation options as part of everyday production. Engineers define the task, evaluate the output, and take responsibility for what reaches users. As capable models become widely available, the quality of that collaboration becomes a lasting advantage.

AI fluency is the ability to choose when to use AI, explain the task clearly, recognize weak or unsafe output, and remain accountable for the result. Labs are giving these behaviors a common vocabulary, and employers are incorporating them into hiring and management. Interview methods vary, and published evidence has yet to establish how well the scores predict performance after hiring.

Models expand what can be produced. Human direction, judgment, and accountability determine what becomes valuable.

1 · AI-mediated production changes the human role

AI fluency becomes a daily requirement as model-generated work enters production. Google’s first-party figures are useful because they count AI-generated code that engineers reviewed and accepted.

Today, 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall.

Sundar Pichai, CEO@GoogleCloud Next, Apr 2026[2]
Figure 1 · Engineer-approved AI code · 2024–2026
Google’s engineer-approved AI code share rose from over 25% in Q3 2024 to about 50% in Fall 2025 and 75% on 22 April 2026. The dashed extension projects the last observed increase. OpenAI’s blue point uses a different measure, with Codex producing 64% of combined Codex and ChatGPT output tokens in June 2026.
Mediation riseGoogle engineer-accepted AI code share from Q3 2024 through April 2026, plus OpenAI Codex token share in June 2026. A human acceptance gate remains throughout.Observed production mediationAI output rises. The human gate remains.Accepted AI-generatedcode ratio0%25%50%75%100%Projection20252026>25%Q3 2024[1]~50%Fall 2025[2]75%Apr 2026[2]64%Jun 2026[3]Google · approved new codeOpenAI · Codex tokens✓Human review and acceptance gateEvery reported code milestone remains subject to engineer approval.

According to Figure 1, Google’s reported share rises from more than 25% in Q3 2024 to about 50% in Fall 2025 and 75% on 22 April 2026.[1][2] The counted code has passed engineering review. OpenAI’s 64% point measures Codex output tokens and uses a separate color to distinguish that unit.[3] Other Alphabet statements citing 75% concern Cloud product adoption or Ads support queries and belong to separate measures.[4][5]

As accepted AI-generated code becomes the majority, directing and reviewing model output becomes ordinary engineering work. AI fluency gives that human work a name.

2 · Fluency makes the human work observable

Anthropic’s 4D framework names four parts of that work. Delegation decides whether and how AI should participate. Description turns intent into context and instructions. Discernment tests the result. Diligence keeps responsibility with the person.[6][7] These behaviors travel across tools and vendors.

The most common expression of AI fluency is augmentative—treating AI as a thought partner, rather than delegating work entirely.

Kristen Swanson, Author@AnthropicAnthropic Education Report, Feb 2026[8]
Figure 2 · Fluency as a closed collaboration loop
Anthropic’s four dimensions describe repeated human decisions in AI collaboration. Its AI Fluency Index examines behavior in 9,830 multi-turn Claude.ai conversations from January 2026; 85.7% showed iteration or refinement.
4D fluency loopFour human collaboration behaviors form a closed loop.Index · Jan 2026 · 9,830 chats · 85.7% iterateHuman judgmentdrives every turn01DelegationChoose the work and when AIshould help02DescriptionSet context, criteria, andconstraints03DiscernmentTest the output againstreality04DiligenceOwn the risks, impact, andresultOwnership closes the loopObserve the person, not polished output[6][7][8]

According to Figure 2, the four dimensions form a loop because each model response calls for another human decision. Anthropic’s Index measures those behaviors. In 9,830 multi-turn Claude.ai conversations, 85.7% showed iteration or refinement. The framework covers 24 human behaviors, of which 11 can be observed in chat.[8]

OpenAI’s literacy blueprint also emphasizes critical interpretation, effective deployment, and risk mitigation.[9] A May 2026 Codex study shows the scale of work these decisions can affect. Among sampled users, 80.6% made at least one request estimated to require more than 30 minutes of human effort, 70.2% more than one hour, and 25.6% more than eight hours.[10] Fluency belongs to the person, even as models change. Hiring needs realistic tasks that make this capability visible.

3 · Coinbase turns fluency into an interview system

Coinbase offers a detailed published example. Its H2 2025 frontend pilot revealed tasks that AI could solve too easily. In January 2026, it switched to backend tasks built around a repository. By March, the company had introduced fluency scoring across its hiring process.[11]

We can’t hire engineers to work alongside AI if we’re still selecting for the ability to work without it.

Coinbase Engineering@CoinbaseCoinbase Engineering Blog, Jul 2026[11]
Figure 3 · Coinbase scored fluency rubric
Coinbase progressed from a frontend pilot to repository-based backend interviews and company-wide scoring. Junior and senior candidates share one rubric. Checks at 45 and 90 days assess performance after hiring; completed results remain unpublished. Source: Coinbase Engineering, 13 July 2026.
Coinbase scored fluency rubricOne fluency rubric across levels, followed by post-hire checks with results pending.01Phase 1H2 2025 frontend pilot · AIcould solve the questions02Phase 2Jan 2026 backend ·repository debugging andreview03Phase 3Mar 2026 · fluency scored atevery stageSame fluency bar · junior and senior01UsageScored in loopSelect and use the right tools02ApplicationScored in loopKnow when AI fits and design forimpact03Understanding LimitsScored in loopRecognize model limits and applyjudgmentPost-hire validation methodGoal · 100% maintain or improve45-day pulse90-day pulseResults unpublishedMethod published · results pendingScore fluency in the interview. Validate it after hire.[11]

According to Figure 3, Coinbase changes the work candidates perform and applies the same rubric to junior and senior candidates. Usage assesses tool operation, Application assesses when AI is appropriate, and Understanding Limits assesses judgment about risks such as privacy and security.[11]

Coinbase also describes checks at 45 and 90 days after hiring, with a goal for 100% of new hires to maintain or improve their rating. Completed results remain unpublished. The rubric provides a consistent assessment method whose ability to predict job performance still needs validation.

We rolled out AI signals at every stage of the engineering interview loop, not as an overlay on the existing process, but how we evaluate every candidate.

Coinbase Engineering@CoinbaseCoinbase Engineering Blog, Jul 2026[11]
Figure 4 · Coinbase interview stages before and after
Coinbase’s Phase 3 design integrates AI-assisted work and fluency ratings across five stages, from pre-screening to offer approval. Source: Coinbase Engineering, 13 July 2026.
Coinbase interview stagesFive Coinbase interview stages before and after fluency was embedded.Interview StageBeforeAfter01Pre-screen(codingassessment)Standard codingassessmentAI-Assisted Coding Assessment.Candidates use AI toolsthroughout.02Live coding /Tech ExecutionNo AI tools permittedAI Support and Referenceenabled. Candidates usedocumentation and AIassistance as they would atwork.03Domain / ReverseSystem Design(senior roles)Standard systemarchitecturediscussionNow includes questionssurfacing how candidates haveused AI in real systems theyhave built.04Foundational(Hiring Managerinterview)Standard leadershipquestionsAdded AI Competency questionsand rubric. Every hiringmanager assesses AI fluencydirectly.05Round-up & offerapprovalNo AI signal inapproval workflowEvery candidate receives an AIFluency rating (Above/At/BelowThreshold), visible inexecutive offer approvals.Every stage contributes evidence. Offer approval sees the final rating.

According to Figure 4, fluency is assessed throughout all five interview stages. Candidates use AI during pre-screening and technical work, discuss previous AI experience in senior system design, and answer rubric-based hiring-manager questions. The resulting Above, At, or Below Threshold rating is visible during offer approval.[11] Each stage contributes evidence about the candidate’s fluency. Other employers use different methods to gather that evidence.

4 · Interview rules select for different skills

Employers publish rules ranging from permitted AI collaboration to independent live assessment. Their choices reflect the skills they want to observe, including realistic production work, fundamentals, and original reasoning.

  • Meta expects AI use inside many technical interviews through a built-in assistant, with outside tools unauthorized.[12]
  • IBM allows responsible AI in preparation when it reflects true capabilities, and forbids AI on assessments and live interviews unless the assessment states otherwise.[13]
  • Microsoft allows responsible AI in preparation and forbids outside assistance during assessments and interviews unless explicitly permitted.[14]
  • Cisco allows GenAI for prep and brainstorming and bars real-time generated coding solutions and scripted live answers.[15]
  • Adobe asks for concrete examples of AI-amplified impact and bars recording, live scribing, and conversation prompting during hiring conversations unless invited.[16]
  • GoDaddy publishes stage-specific expectations, with disclosure for applications and assessments and generally independent live phone or video rounds.[17]
  • Salesforce evaluates AI fluency as a core competency across every role, with a possible technical assessment of how candidates use AI.[18]
  • Anthropic seeks people who collaborate with Claude in production, generally forbids AI on take-homes and live interviews unless allowed, and documents AI-resistant evaluation tradeoffs when models match top human scores on PE take-homes.[19][20]
  • Google listings name AI fluency as a role expectation and ask candidates to discuss AI tool use during hiring, while official how-we-hire pages still omit whether AI may be used in coding exercises.[21][22]

It’s a core competency we evaluate across every role, just in different ways.

Dani Laven, Senior Manager Employer Brand@SalesforceSalesforce Blog, May 2026[18]
Figure 5 · What hiring selects for
Nine hiring programs are grouped by the skills they aim to observe. The range runs from independent work to authorized AI collaboration. Each card pairs a skill with its assessment method.
Hiring selection signalsNine employers clustered by the characteristic they try to observe.01Independent craftUnaided capability in live workIBMLooks forIndependent capabilityPreparation AI isallowed when it reflectsreal skill. Assessmentsand live interviewsremain independentunless permitted.MicrosoftLooks forIndependent liveperformanceResponsible preparationis allowed. Assessmentsand interviews requireexplicit permission foroutside help.CiscoLooks forOriginal live thinkingPreparation andbrainstorming areallowed. Generated livesolutions and scriptedanswers are not.GoDaddyLooks forIndependent liveconversationApplications andassessments may usedisclosed AI. Live phoneand video rounds remainindependent.02RestrictedevaluationAI at work, independentassessmentAnthropicLooks forClaude collaborationProduction rewardsClaude collaboration.Take-homes and liveinterviews generallyrestrict AI unlessallowed.03Named fluencyNamed competency and prior impactAdobeLooks forAI-amplifiedimpactCandidates showAI-amplifiedimpact. Recording,live scribing, andprompting remainoff unlessinvited.SalesforceLooks forAI fluency acrossrolesFluency is a corecompetency.Technicalcandidates may beassessed on howthey use AI.GoogleLooks forAI fluency as arole skillListings ask aboutAI use. Officialhiring pages donot say whethercoding exercisesallow it.04In-loop useAuthorized AI in theinterviewMetaLooks forIn-environment AIcollaborationMany technicalinterviews include abuilt-in assistant.Outside tools are notallowed.Each policy is an instrument for observing a different skill.

According to Figure 5, the spectrum moves from independent live work to authorized AI use inside the interview. IBM, Microsoft, Cisco, and GoDaddy preserve independent live rounds while allowing some preparation use.[13][14][15][17] Anthropic hires for Claude collaboration but usually restricts AI during assessment.[19][20] Adobe, Salesforce, and Google ask for prior impact or named fluency, while Meta observes collaboration through a built-in assistant.[12][16][18][21] Interview rules determine which skills an employer can observe. That choice only becomes credible when daily work rewards the same behavior.

5 · Operating practice must reinforce the hiring signal

Shopify connects hiring expectations with everyday management. Its CEO describes routine AI use as a baseline, requires headcount proposals to explain why AI cannot do the work, and includes AI use in performance and peer review.[23] The Apprentice Product Manager process also assesses tool selection and vibe coding.[24]

It doesn’t matter if you’re early in your career, mid or late. There’s no excuse to not become a builder — the AI tooling makes this so much more approachable and you won’t have an excuse for the dreaded: “Do you know how to work / build with AI?” question

Farhan Thawar, Head of Engineering@ShopifyX, Jun 2026[25]
Figure 6 · Three interview modes and the policy boundary
The three preferred interview modes restrict, allow, or require AI use. They sit below Shopify’s established operating expectations. The dashed boundary distinguishes this preference from company policy, which the official engineering guide and Careers FAQ have yet to confirm.
Multi-mode policy bandA published operating baseline, a preferred three-mode interview, and the policy boundary between them.Operating baseline[23][24]Reflexive AIA company baseline,not a side experimentHeadcountExplain why AI cannotdo the workPerformanceAI use in performanceand peer reviewAPM hiringTool fit and avibe-coding stagePreferred practice · Head of Engineering[26][27]01Not allowedFundamentalsMeasuresIndependent work withoutAI02OptionalJudgmentMeasuresWhether candidates knowwhen AI helps03MandatoryUnder pressureMeasuresDirect and judge AI underpressurePublished how-we-hire text[28][29]Engineering guide + Careers FAQ · three modes not documentedThe operating baseline is published. The three-mode interview remains preferred practice.

According to Figure 6, the preferred three-mode approach gives each interview exercise a purpose. Restricted AI use tests fundamentals, optional use tests judgment, and required use tests collaboration under pressure.[26][27] This fits Shopify’s use of AI in headcount planning, performance management, and early-career assessment. The dashed line marks an evidence limit. Shopify’s engineering guide and Careers FAQ have yet to establish these three modes as company policy.[28][29]

Salesforce shows the same direction at workforce scale. A May 2026 announcement targeted 1,000 graduates and interns and said its Emerging Talent Playbook evaluates AI fluency early.[30] Its FY26 Stakeholder Impact Report also names an AI Fluency Playbook for working alongside agents.[31] Fluency becomes durable when hiring, daily work, and performance management reinforce the same observable behaviors. A shared standard connects these decisions.

6 · Build one standard across hiring and work

Leaders can adopt a shared standard now and test its predictive value as post-hire evidence develops.

  1. Define observable dimensions. Choose portable language such as Delegation, Description, Discernment, and Diligence or Usage, Application, and Understanding Limits. Use it in role expectations, interview rubrics, and review conversations.[6][11]
  2. Publish the rule for every stage. State whether AI is expected, optional, disclosed, or restricted in applications, take-homes, assessments, and live interviews. Meta, IBM, Microsoft, and GoDaddy demonstrate different policies with explicit boundaries.[12][13][14][17]
  3. Assess collaboration rather than product familiarity. Ask candidates to frame work, direct a model, inspect its output, and respond to failure. These behaviors connect Coinbase’s rubric with Anthropic’s discernment and diligence dimensions.[8][11]
  4. Preserve the fundamentals the role requires. Use task modes deliberately so unaided reasoning, judgment about tool use, and AI-assisted execution each have a clear purpose.
  5. Measure after hiring. Compare interview ratings with evidence at 45 and 90 days before expanding score cutoffs. Coinbase has published this method, while completed predictive results remain open.[11]
  6. Reinforce the standard in management systems. Bring the same fluency language into performance, peer review, and workforce planning, following the operating direction visible at Shopify and Salesforce.[23][31]

A fluency standard becomes useful when hiring, production, and performance management ask for the same observable behaviors. Candidates gain clearer expectations, and employers gain a better basis for testing whether assessment predicts performance.

Conclusion

AI fluency brings direction, judgment, and accountability to work produced with models. Employers can apply one standard across hiring, daily work, and performance reviews. Each interview stage needs a clear AI rule and a way to observe collaboration. Interview scores should remain provisional until evidence shows how they relate to performance on the job.

The published evidence supports this direction, with limits. Interview designs vary, detailed expectations by job level remain limited, and completed predictive results are still unpublished. Measure outcomes within your organization and assess the same behaviors during hiring and on the job.[8][18][32]

The enterprise advantage will come from people who direct AI clearly, judge it rigorously, and remain accountable for the result.

References

  1. Sundar Pichai, CEO@Alphabet / Google (). “A letter from our CEO and Alphabet Q3 2024 earnings”. Google Blog. Public report that more than 25 percent of new code at Google was generated by AI and then reviewed and accepted by engineers (Q3 2024).
  2. Sundar Pichai, CEO@Alphabet / Google (). “Cloud Next 2026 keynote remarks on AI-generated code”. Google Blog. Public Cloud Next snapshot that 75 percent of all new Google code is AI-generated and engineer-approved, after a Fall 2025 public figure near half of new code.
  3. OpenAI (). “How enterprises put AI to work”. OpenAI. Assigns a June 2026 enterprise figure that Codex is 64 percent of combined Codex and ChatGPT output tokens.
  4. Alphabet (). “Alphabet investor presentation, June 2026”. Alphabet / Google Blog. A later official 75 percent line measures the share of Cloud customers using Google AI products, not engineer-accepted new-code share.
  5. Sundar Pichai, CEO@Alphabet / Google (). “Alphabet Q2 2026 earnings call”. Alphabet. A later official 75 percent line measures Ads Customer Support queries handled by Gemini-powered agentic solutions, not engineer-accepted new-code share.
  6. Anthropic (). “AI Fluency Framework and Foundations”. Anthropic. Defines AI Fluency as portable human collaboration skills across four dimensions: Delegation, Description, Discernment, and Diligence.
  7. Anthropic (). “Reflect with Claude and the 4D fluency framing”. Anthropic. Product and education framing that operationalizes 4D fluency as how people decide what to hand off, how they specify work, how they judge outputs, and how they remain accountable.
  8. Anthropic (). “Anthropic Education Report: The AI Fluency Index”. Anthropic. January 2026 sample of 9,830 multi-turn Claude.ai conversations; 85.7 percent show iteration or refinement. Treats fluency as 24 observable human behaviors, including 11 chat-observable signals.
  9. OpenAI (). “Teen AI Literacy Blueprint”. OpenAI. Centers critical interpretation of model outputs, effective deployment of AI tools, and risk mitigation, including work that should not be automated.
  10. OpenAI (). “How agents are transforming work”. OpenAI. Measures May 2026 individual Codex use: 80.6 percent of sampled users made at least one request estimated at more than 30 minutes of human work, 70.2 percent more than one hour, and 25.6 percent more than eight hours.
  11. Coinbase Engineering (). “Interviewing engineers in the AI era — lessons from a year of rebuilding”. Coinbase. Describes a three-phase rebuild of the engineering interview loop around AI Fluency, with Phase 2 in January 2026 and Phase 3 in March 2026, dimensions Usage, Application, and Understanding Limits, and a 45-/90-day pulse method (results not yet published).
  12. Meta Careers (). “Meta hiring process”. Meta. States that many technical interviews include a built-in AI assistant candidates are expected to use, with outside AI tools unauthorized beyond the interview environment.
  13. IBM Careers (). “Application steps and FAQs — acceptable use of AI by candidates”. IBM. Encourages AI for prep, grammar, and mock interviews with disclosure of significant assistance, and prohibits AI on assessments, coding challenges, and live interviews unless the assessment states otherwise.
  14. Microsoft Careers (). “How we hire — candidate code of conduct and AI guidance”. Microsoft. Allows responsible AI use in preparation when it reflects true capabilities, and forbids outside assistance during assessments and interviews unless explicitly permitted.
  15. Cisco Careers (). “GenAI Best Practices and Permitted Uses”. Cisco. Allows GenAI for interview prep and brainstorming while forbidding real-time generated coding solutions and scripted live-interview answers.
  16. Adobe Careers (). “AI and your hiring experience”. Adobe. Asks for concrete examples of AI-amplified impact and bars AI recording, live scribing, and conversation prompting during hiring conversations unless invited.
  17. GoDaddy Careers (). “AI Expectations for candidates”. GoDaddy. Publishes stage-specific candidate AI rules: AI may support applications and assessments with disclosure; live phone and video interviews are generally independent; transcription AI is restricted except accommodations.
  18. Dani Laven, Senior Manager Employer Brand@Salesforce (). “Ask a Recruiter: How Do I Get AI-Ready Without a Technical Background?”. Salesforce. Treats AI fluency as a core competency evaluated across every role, with a possible technical assessment of how candidates use AI, and frames Agentblazer Champion, Innovator, and Legend as a product learning ladder rather than a junior-versus-senior job rubric.
  19. Anthropic (). “Guidance on candidates' AI usage”. Anthropic. Seeks people who collaborate effectively with Claude in production while generally forbidding AI on take-homes and live interviews unless the stage explicitly allows it.
  20. Anthropic Engineering (). “Designing AI-resistant technical evaluations”. Anthropic. Documents PE take-home redesign tradeoffs when models match top human scores, including why bans alone are misaligned and how AI-resistant evaluation design is pursued.
  21. Google Careers (). “Staff Software Engineer, Android Hotword — job listing”. Google. States that demonstrating AI fluency is an expectation in the role and that candidates should be prepared to discuss their knowledge and use of AI tools during the hiring process.
  22. Google Careers (). “Our hiring process”. Google. Official how-we-hire page; does not publish whether candidates may use AI during coding interviews, assessments, or leveling.
  23. Tobi Lütke, CEO@Shopify (). “Reflexive AI usage as company baseline”. X. Sets reflexive AI use as a company baseline, ties headcount requests to why AI cannot do the work, and adds AI usage questions to performance and peer review.
  24. Shopify (). “Apprentice Product Manager Program”. Shopify. APM hiring design includes AI tool fit and vibe-coding assessment as program stages.
  25. Farhan Thawar, Head of Engineering@Shopify (). “Do you know how to work / build with AI?”. X. States that early, mid, and late-career builders now face the hiring question of whether they know how to work and build with AI.
  26. Farhan Thawar, Head of Engineering@Shopify (). “Preferred multi-mode technical interview format”. X. Describes a preferred three-problem technical format with AI not allowed, optional, and mandatory modes to probe fundamentals, judgment, and pressure use of AI.
  27. Farhan Thawar, Head of Engineering@Shopify (). “Waterloo-style AI modes applied to interviewing”. X. Links multi-mode AI evaluation (not allowed / optional / mandatory) to interviewing practice and candidate assessment design.
  28. Shopify Engineering (). “Nail your technical Shopify interview”. Shopify Engineering. Official engineering interview process guide; does not publish multi-mode AI not-allowed / optional / mandatory bands as formal how-we-hire policy text.
  29. Shopify Careers (). “Careers FAQ — how we hire”. Shopify. Careers FAQ on hiring process; does not document multi-mode AI interview bands as published policy.
  30. Salesforce (). “Salesforce Commits to Hiring 1,000 AI Native Grads”. Salesforce News. States that Salesforce is recruiting 1,000 graduates and interns through the Builder / Futureforce program, and that the Emerging Talent Playbook Assess step evaluates AI fluency early. The 1,000 figure is a recruiting target.
  31. Salesforce (). “FY26 Stakeholder Impact Report”. Salesforce. Names an AI Fluency Playbook for working alongside AI agents.
  32. Saffron Huang and colleagues@Anthropic (). “How AI is transforming work at Anthropic”. Anthropic. December 2025 workplace study of Claude use among Anthropic engineers and researchers. Not a selection-validation study.