Evidence & perspective / Reviewed October 2026What the evidence supports.
The evidence behind three workforce questions: what capability to own, how much work to deliver, and how quickly to deliver it.
These studies inform our approach. They are external findings, not Deep48 client results or forecasts of your return. The recommendation for a particular workflow depends on its baseline, pilot performance, and operating economics.
01 / What could you own?Internal development is entering the purchasing decision.
McKinsey's 2026 survey reports that 32% of respondents said their organizations avoided at least one software product or feature purchase because they could build the functionality internally using agentic coding tools.
This is self-reported purchasing behavior. It does not establish successful replacement of existing platforms, lower total ownership cost, or the return for an individual client.
What this means for your business: A focused internal tool may avoid a purchase, but compare the full cost before replacing anything. Count savings only when expenditure actually falls.
More output is possible in specific workflows.
In Generative AI at Work, Brynjolfsson, Li, and Raymond found an average 15% increase in issues resolved per hour with AI assistance in customer support, with substantial differences across workers.
The study concerns an assistant supporting human customer-service workers. It does not establish autonomous agent performance, recruiting gains, or unlimited scaling.
What this means for your business: Test whether your team can deliver more usable work per hour, including time spent checking and correcting output.
The process matters alongside the technology.
McKinsey's March 2025 survey found that workflow redesign was the attribute most associated with reported EBIT impact among the 25 organizational attributes it examined.
A survey association does not prove causation. It supports examining the operating process rather than assuming that adding a tool will produce a financial result.
What this means for your business: Look beyond a faster individual task. Delays between teams and repeated review may be better targets for improving delivery.
Task-level speed can improve. Quality still defines the boundary.
In the Harvard Business School and BCG experiment Navigating the Jagged Technological Frontier, participants using GPT-4 completed tasks within the studied capability frontier 25.1% more quickly on average and completed 12.2% more tasks.
The experiment studied selected knowledge-work tasks using a 2023 version of GPT-4. Performance could deteriorate outside the model's capability frontier. The findings do not establish a comparable reduction in end-to-end business delivery time or the performance of current agents.
What this means for your business: Choose tasks the tool can perform reliably, then measure time to a usable result. Faster drafts are valuable only if correction and review do not erase the gain.
Productivity must be measured, not assumed.
METR's randomized study of experienced open-source developers found that early-2025 AI tools increased task completion time by 19% in that setting, despite developers believing the tools made them faster.
In February 2026, METR reported that its follow-up experiment could not reliably estimate current gains because of selection and measurement issues. The earlier result should not be treated as a verdict on today's tools or all software development.
What this means for your business: Compare completion time and quality before investing further. A tool that feels faster may still cost your team more time.
Financial evaluation includes implementation, migration, operating costs, maintenance, and the timing of any spend reduction. Freed staff time is capacity; cash savings require an actual reduction in expenditure.
Deep48 perspective / John W. Syed
Who Recruits
the Machines?
Our perspective asks how Talent Acquisition could participate in sourcing human and agentic capability for outcomes. Outcome Autonomy, Human Leverage Ratio, and Experience Custody are proposed research concepts, rather than established industry standards.
Read the full essay →