Microsoft’s Copilot may be the most visible face of generative AI inside business apps, but the reality on the ground is stark: widespread experimentation has not yet translated into widespread benefit for organisations or users. The latest Computing research of UK IT leaders shows heavy adoption of branded copilots and LLMs — Copilot, ChatGPT and Gemini lead the pack — yet respondents repeatedly report that most deployments are narrow, tactical and, crucially, under‑measured for business impact. rview
Generative AI’s rapid rise over the past three years has produced a familiar pattern: early consumer buzz, followed by hurried enterprise trialing and product teams racing to fold LLM capabilities into existing software suites. Surveys from independent consultancies confirm the scale of interest: large studies from Deloitte show millions of UK workers have used GenAI and report perceived productivity gains, while PwC’s CEO survey finds nearly all UK chief executives claim some level of GenAI adoption in their organisations. These broader signals match the Computing findings — many organisations are using GenAI tools, yet few say they are realising durable, measurable benefits.
The Computing sample is a targeted snapshot of 138 UK IT leaders. It documents which models and platforms these professionals are using, what problems they are trying to solve, and — just as importantly — the dissonance between vendor promises and day‑to‑day value. The survey captures both tool preferences (Copilot 79%, ChatGPT 70%, Gemini 45%, Claude 24% and muor specialist platforms) and candid practitioner commentary about real outcomes.
For CIOs and IT leaders the imperative is clear: stop chasing shiny demos and start running disciplined experiments thatscaled. Where vendors promise one‑click transformation, ask for baselines, sample scripts, and real evidence of improved KPIs. In the meantime, expect GenAI to deliver real but incremental gains in development and knowledge work, while remaining wary and defensive in high‑risk, regulated workflows. The technology is not a panacea — but with the right engineering and governance it is already a pragmatic accelerator for teams who focus on outcomes rather than hype.
Source: Computing UK Everyone’s using GenAI, few are truly benefiting
Generative AI’s rapid rise over the past three years has produced a familiar pattern: early consumer buzz, followed by hurried enterprise trialing and product teams racing to fold LLM capabilities into existing software suites. Surveys from independent consultancies confirm the scale of interest: large studies from Deloitte show millions of UK workers have used GenAI and report perceived productivity gains, while PwC’s CEO survey finds nearly all UK chief executives claim some level of GenAI adoption in their organisations. These broader signals match the Computing findings — many organisations are using GenAI tools, yet few say they are realising durable, measurable benefits.
The Computing sample is a targeted snapshot of 138 UK IT leaders. It documents which models and platforms these professionals are using, what problems they are trying to solve, and — just as importantly — the dissonance between vendor promises and day‑to‑day value. The survey captures both tool preferences (Copilot 79%, ChatGPT 70%, Gemini 45%, Claude 24% and muor specialist platforms) and candid practitioner commentary about real outcomes.
What the survey actually shows
Who organisations are using (brand share and platform penetration)
- Microsoft Copilot is the most commonly reported assistant among the surveyed IT leaders (79% of 138 respondents). This mirrors the commercial strategy Microsoft has used since Azure and Office 365: bundle AI into existing product lines and win by convenieather than by being technically superior in every scenario.
- OpenAI / ChatGPT remains highly prevalent (70%), benefitting from early‑mover recognition and broad consumer familiarity. Many professionals stifor quick drafting and ideation tasks.
- Google Gemini (45%) occupies a credible third position, especially with organisations embedded in Google Workspace or Google Cloud. Several respondents explicitly mentioned trusting Gemini or preferring it to Cn some workflows.
- Anthropic Claude (24%) shows traction among developer and enterprise B2B users who value its enterprise positioning. Other platform mentions (Llama 6%, Bedrock, Cursor, Replit, AlphaCode) were small by comparison. Image generators and specialist search‑centric models (Perplexity, Midjourney) appear much less commonly used in pro Beyond discrete assistants, hyperscaler cloud AI platforms are widely present: the Computing survey reports cloud AI usage in roughly 80% of organisations polled, including offerings from Microsoft, Google, Amazon, IBM, Oracle and others. This highlights that many organisations are experimenting with or buying managed AI services rather than only running third‑party chatbots.
actually use GenAI for (use‑cases on the ground)
Respondents named an unexpectedly ordinary set of use cases where generative AI is being trialled or deployed:- Software development acceleration (code scaffolding, conversion, documentation and basic “vibe” coding for prototypes).
- Documentation and knowledge work (transcription, summarisation, bid drafting, tender discovery, and training content generation).
- People‑facing automation (customer support agent assists, HR admin automation, sales enablement).
- Security and SOC workflows where generative techniques support scenario generation for red/blue team exercises or easier explanations for non‑technical stakeholders.
Why the gulf ed reality
There are five structural reasons the Computing research — and larger industry studies — find serious dissonance between AI marketing and real outcomes.1. Convenience beats capability
Vendors that embed LLM features into ubiquitous software win adoption through proximity. Microsoft’s Copilot appears in the apps teams use every day; that lowers friction and explains the high usage figure despite mixed satisfaction. Adoption driven by convenience does not guarantee value if the integration lacks adequate grounding in business data, workflows and verification controls.2. Data foundations are weak or fragmen are powerful as synthesizers, but they require clean, governed, queryable data and reliable retrieval layers (RAG) to produce grounded answers. The core engineering work is not model access — it is data unification, metadata cleaning, access controls and versioning. Organisations that treat LLMs as magic endpoints without investing in the data plumbing will see brittle, inconsistent results. Numerous practitioner guides and enterprise playbooks highlight the same point: the non‑glamour work of data engineering is decisive.
3. Governance and verification are nascent
Respondents repeatedly noted ad‑hoc use, contractual bans on AI for client work, and internal policies that restrict model access. When companies lack robust human‑in‑the‑loop verification, audit trails and role‑based access, the risk of inaccurate outputs or data leakage increases. This both reduces business confidence and caps the scale of deployment. These governance gaps help explain why many GenAI uses remain limited to internal admin, meeting summaries, and “first pass” drafts.4. Poor metric design and absent ROI tracking
Most pilots are not set up with realistic KPIs narrow, measurable outcomes tied to revenue, cycle time or quality, many trials remain open‑ended. The result: leaders cannot credibly claim productivity gains beyond anecdotal minutes saved. To change that calculus, teams must define baselines, control groups, and ongoing measurement — exactly the discipline missing in many early pilots. Industry playbooks increasingly recommend “outcome‑first” pilots as the path to durable value.5. The hallucination problem and domain risk
Generative models still make confident, plausible‑sounding errors. In regulated domains — legal, procurement, healthcare — a single erroneous suggestion can cause reputational, financial or compliance harm. Practitioners therefore constrain LLM usage to low‑risk tasks or force outputs through expert review, which reduces scale and speed. Several respondents in the Computing survey explicitly cautioned that legal and regulated projects either ban or strictly limit AI use.What is working — practical, measurable use cases
The gulf is real, but threads of clear, replicable value are emerging. Thlace the worker” plays; they are practical process improvements with measurable KPIs.Development acceleration
Generative tooling shortens early engineering chores: scaffolding, test generation, code translation and documentation. Teams report:- Faster prototype cycles and reduced time-to-first-draft for code components.
- Improvements in developer throughput when models are tuned to internal patterns and code style.
- Measurable reductions in routine engineering hours (vendor case studies and practitioner surveys often cite 10–30% reduction in routine tasks when models are properly integrated and instrumented).
Documentation, transcription and knowledge work
This is the most mature commercial application area today.- Meeting transcription + summarisation rede and speeds downstream work. One respondent estimated up to 20 minutes saved per customer call through auto‑transcription and summarisation.
- Tender and bid automation — discovery, draft creation and versioning — can compress timelines if legal and compliance reviews are built into the workflow.
Customer suppore ops
When used as agent‑assist rather than agent‑replace, GenAI can increase first‑contact resolution, reduce average handle time and improve consistency of replies. The key success factors are clean knowledge bases, strict grounding, and monitoring of containment rates. Independent practitioner reports show double‑digit agent productivity lifts in cases where knowledge is well organised.Security playbooks and scenario generation
Cybersecurity use cases are often more ML‑centric than generative, but GenAI helps teams write incident narratives, generate red‑team scenarios, and tndings for executives. These tasks are high value because they improve communication and speed decision cycles without exposing sensitive data to unvetted public models.Critical limitations and risks
No discussion is complete without plainly stating the risks that are already material in production deployments.- Accuracy and compliance risk: Hallucinations can produce l consequential errors. In regulated sectors the tolerance for error is near zero.
- Data leakage and IP exposure: Sending proprietary documents to third‑party models without contractual or technical safeguards risks training leakage and potential IP claims. Organisations must apply encryption, private endpoints, or oappropriate.
- Vendor lock‑in and governance capture: Heavy reliance on a single hyperscaler’s proprietary pipelines can make future migration costly and limit negotiation leverage. Several respondents flagged the risk that vendor‑embedded features (Copilot in Office) create a ‘soft’ monopoly through convenience.
- Skill gaps and distributional inequality: External research finds that while many employees use GenAI, employers often do not formally support or pay for tools; Deloitte reports that a nontrivial share of UK employees pay out of pocket for GenAI tools and that employers dcourage usage. This creates fragile, unsanctioned pockets of adoption that are hard for IT and governance teams to manage.
- Superficial training and token credentials: Public programmes and short courses raise awareness but do not automatically create the operator skills required to verify outputs or run safe deployments. National training ambitions therefore need to be paired with deeper, role‑specific learning pathways.
How successful organisations are approaching GenAI (what winners do)
The research and practitioner guides converge on a compact set of success factors for moving from novelty to durable value:- Outcome‑first pilots — Start with narrowly scoped projects tied to clear KPI improvements (time saved, error reduction, conversion uplift). Design measurement into the pilot from day one.
- Invest in data foundations — Centralise, govern and index enterprise data for reliable retrieval; invest in RAG layers and versioned knowledge bases. Without this, models will remain brierational discipline (ModelOps / AgentOps) — Continuous monitoring, rollback processes, version control of prompts and models, and audit trails are essential when outputs can affect customers or comverification and role design — Embed humans into review loops and design responsibilities for acceptance testing, verification and exception handling. Avoid blanket automation where trust is low.
- **Hybrid build/buy strpilots for rapid seat‑level value but build custom layers (RAG, policy, workflow orchestration) where differentiation, compliance or defensibility matters. Enterprise deployments that combine both paths tend to
Practical checklist for IT leaders (an operational playbook)
If you are an IT or business leader tasked with turning GenAI experiments into value, this is a succinct operational checklist to convert hype into repeatable outcomes.- Define the hypothesis and metric b, reduce average handle time by X% or cut first‑draft write time by Y minutes).
- Choose a narrow domain and curate a single source of truth dataset to ground the model.
- Use retrieval‑augmented generation (RAG) with strict source attribution rather than blind LLM outputs.
- Run A/B tests and include control groups to measure real impact.
- Put fail‑safe governance in place: mandatory human sign‑off for high‑risk outputs, logging and rollback.
- Protect IP and customer data using private endpoints, contract indemnities, or on‑premise models where required.
- Audit for bias and set a cadence for periodic model re‑evaluation and prompt tuning.
- Budget beyond model access: allocate 50–70% of project effort to data, integration and monitoring work rather than raw model costs.
What the numbers from broader studies tell us (context and calibration)
The Computing survey’s sample of 138 UK IT leaders provides a focused industry snapshot, but larger independent studies offer context about scale and sentiment:- Deloitte’s national surveys estimate tens of millions of UK residents have used GenAI and report substantial self‑reported productivity boosts, even while noting large gaps in employer endorsement and understanding of risk. This aligns with the practitioners’ claim that usage is widespread but unevenly supported.
- PwC’s CEO surveys show near universal adoption claims at the executive level, yet leaders are less confident that GenAI will quickly translate into profits. That gap — between adoption and clear financial impact — echoes the Computing study’s central finding.
A critical reading of the Coitations and caveats)
The Computing research is honest and valuable, but readers should interpret the results with these caveats in mind:- Sample size and representativeness: The survey interrogates 138 IT leaders — a useful but small sample. Their views are insightful for organisational practice, but not a statistically representative national barometer. Extrapolating absolute market shares from this cohort risks overstating prevalence.
- Self‑selection and early adopter bias: Respondents ting are likely to be more technologically curious and potentially more inclined to trial vendor products. That can skew the observed brand mix toward the tools these professionals can more easily access or justify buying.
- Snapshot timing and product churn: The GenAI product landscape moves qui pricing changes and new vendor features appear frequently. Any survey is a snapshot in time; leaders must combine such data with ongoing vendor evaluation and proof‑of‑concepts in their specific environment.
Conclusion — realistic expectations, disciplined execution
The Computing research documentractical truth: everyone is trying generative AI, but few organisations yet reap sustained, measurable value. That reality does not mean GenAI is a mirage — it means the industry is in the transition from tool novelty to disciplined engineering. The winners will be organisations that treat GenAI as an integrated, governed productivity layer: they will invest in data plumbing, define outcome‑focused pilots, bake in operational controls, and measure impact with the same rigor they apply to any critical enterprise system.For CIOs and IT leaders the imperative is clear: stop chasing shiny demos and start running disciplined experiments thatscaled. Where vendors promise one‑click transformation, ask for baselines, sample scripts, and real evidence of improved KPIs. In the meantime, expect GenAI to deliver real but incremental gains in development and knowledge work, while remaining wary and defensive in high‑risk, regulated workflows. The technology is not a panacea — but with the right engineering and governance it is already a pragmatic accelerator for teams who focus on outcomes rather than hype.
Source: Computing UK Everyone’s using GenAI, few are truly benefiting
