AI projects not paying off? Keep at it, says Uni Productivity Institute

A University of Manchester report commissioned by Bytes Technology Group says AI adoption has not yet reliably translated into organization-wide productivity gains, as upfront retraining and workflow costs precede payback and task-level efficiency may not increase overall output. Examples include 26% more tasks completed with GitHub Copilot and 40% less time spent on routine writing and analysis with ChatGPT, but AI coding activity rose about 240% while software releases increased only about 30%. The report estimates high-exposure economies such as the UK could gain 0.4–1.3 percentage points in annual labor productivity, while emphasizing management, worker skills, governance and workflow redesign as prerequisites.

Analysis

The key market implication is a potential mismatch between who monetizes AI now and who captures durable economic value. Vendors can earn from seats or usage while customers absorb workflow redesign, training, review and governance costs. If faster task completion creates more work to check rather than more shipped output, customer ROI—and ultimately renewal or expansion budgets—can lag vendor revenue. That makes AI usage growth a weak proxy for productivity or pricing power.

The constraint may shift, not disappear: more generated code and content can raise demand for testing, security, data quality and human review. Those areas are plausible relative beneficiaries, but the article does not establish that their spending will exceed the costs they help control. For labor-saving software, the relevant proof is business throughput or lower cost per completed outcome, not activity measures alone.

Near term, this is a reason to scrutinize earnings claims, not a standalone catalyst to short AI. Over the next 1–3 months, watch for customer ROI disclosures, software release/output metrics and any slowdown in AI-related budget expansion. Over 6–18 months, successful workflow redesign could support broad productivity gains; failure to convert task efficiencies would expose adoption spending to cuts. The UK’s high exposure to complementary cognitive work is a possible long-run upside, but not a near-term equity signal.

Contrarian point: delayed measured gains do not prove the technology has failed—implementation costs arrive before redesigned processes can scale. Conversely, the report’s promise of eventual payback is not independently verified. The Gartner and Apollo views strengthen the case for demanding evidence, but do not establish a company-specific earnings impact for Gartner (IT) or Apollo (APO).

Key Decisions for Investors

• Avoid a broad AI short based on this report alone. In upcoming software earnings, favor companies that disclose workflow-level outcomes—cost per transaction, cycle time, or shipped output—over those citing adoption, prompts, seats or task-volume growth without a financial bridge.
• Use a conditional relative-value watchlist: long providers exposed to testing, cybersecurity, data quality or governance against labor-replacement software only if customer evidence shows review bottlenecks rising while realized savings remain absent. Do not put on the pair without confirming valuation, customer retention and revenue exposure.
• For private-equity underwriting, treat AI labor savings as unproven until portfolio-company data show sustained output per employee or lower operating cost after implementation. The article offers no basis to infer a direct earnings effect for APO.
• Track falsifiers over the next 1–3 months: credible company disclosures of lower unit costs or higher completed output would weaken the delayed-ROI concern; budget freezes, weaker renewals, or rising review and testing costs without throughput gains would strengthen it. No direct trade in IT is warranted from this evidence alone.