Why do AI adoption results show up first as job cut numbers?

Why did job cut announcements and a results survey point in different directions in the same week?

On October 24 alone, news pointing in different directions came out one after another. According to AI Times, UK bank Barclays said in its third-quarter earnings announcement that it will cut about 5,000 jobs this year, about 5% of its total workforce (AI Times, s1). The figure includes 1,000 jobs cut in the first half, with 4,000 more to be cut by year-end. CFO Anna Cross said about 60% of the cuts will be in support functions, and reportedly said that "as AI raises employee productivity, the number of developers is also falling" source.

The same day, a Gartner survey result was shared on Reddit. It said only 13% of CIOs answered that their company had gained significant or transformative value from AI (secondhand via Reddit, s2). The shared post does not give the number of respondents, the survey period or the survey method. So this figure should be read only as a directional signal.

Pandaily reported the same day that Ma Tao, chair of OpenAnolis, said at a GOSIM event that agents now write 95% of the new code from Alibaba Cloud's foundational software team (Pandaily, s3). How this 95% was counted was not disclosed source.

On one side came announcements that most of the workforce and code are shifting to AI; on the other came a survey saying only a minority of executives have felt the results. Rather than directly contradicting each other, the two signals can be seen as revealing that each side 'counts results' differently.

What are companies using to measure AI results?

Sorting this week's news by results metric shows that each uses a different yardstick. Barclays pointed to headcount source, and the Alibaba Cloud side pointed to the share of new code written by agents source. Both are metrics showing how much of people's work AI is taking over.

Vodafone took a different path. According to AI Times Korea (인공지능신문), Vodafone unveiled an autonomous network strategy in which its network forecasts traffic demand and detects failures in advance. It said this has cut the daily power consumption of its 5G radio network by a double-digit percentage and shortened upgrade work on thousands of routers by about 70% (AI Times Korea, s4). It is also pursuing a 50% reduction in outages source. This is a case of showing results through operational outcomes rather than headcount.

The AT&T legal team case presented by OpenAI deals with changes in how work is done rather than numbers. It covered how the AT&T LegalEdge team redesigned its legal request intake process and applied legal judgment more broadly (OpenAI, s5).

Laid out this way, four yardsticks are being used under the single word 'results': headcount, share of work, operational metrics and work redesign. The 13% survey result source could also be read differently depending on which yardstick the question used.

Why do cost savings show up in numbers first?

Headcount reductions are numbers that show up directly in the accounts. The number of jobs cut and its percentage can be announced right away in an earnings release source. Productivity is different. You first have to decide what to measure it against. Even when a share figure such as 95% comes out, it is hard to compare with other organizations if the measurement method is not disclosed alongside it source.

GitHub's addition of an 'Industry comparison' section to its Copilot impact dashboard also relates to this gap (GitHub, s6). According to GitHub, organizations can see the share of agent-first users and users of multi-agent approaches. They can also compare the difference in pull request merge rates between agent users and non-users against a peer group of similar size and industry source. Organizations choose a company size such as 1–999, 1,000–4,999 or 5,000 or more employees and one of 10 industries, and scores are compared with the peer group median over a 28-day reporting period source. Participation is opt-in and off by default source.

GitHub explained that until now leaders could not tell whether their organization's adoption was ahead of similar companies, and that industry comparison provides the context needed to set realistic goals source. This piece reads job cuts as showing up first not because their effect is larger, but because they are the easiest number to count. Tools for measuring productivity and outcomes are only now emerging.

What does this gap mean for practitioners?

Changes in workforce structure are a sensitive issue for the people affected. This week's news alone is not enough to judge AI's overall impact on jobs. Barclays' job cuts were also reported as a decision that combined AI adoption with organizational restructuring source.

The more direct implication for practitioners is how results are measured. If you report AI results to executives with only workforce metrics, you are likely to be judged only on 'how much you cut'. Presenting operational metrics such as power, task time and outages, as Vodafone did source, or outcome metrics such as the merge rate in GitHub's dashboard source, alongside them can arguably show the value AI creates more broadly.

When citing figures, also check how they were measured. Even eye-catching numbers such as 95% source or 13% source are hard to use as a basis for comparison with your own organization if the calculation method or sample has not been disclosed. The starting point is to write down the definitions and time periods for your own organization's numbers.