MCC AI Weekly · Thursday, 23 July 2026

The productivity gap nobody is measuring

Abstract illustration of a widening gap between two data curves

Three quarters of enterprises now report AI in production. Almost none report a measurable change in output per employee. This week we look at why that gap persists — and at the two organisations that closed it.

Insights

Only 6 % of enterprise AI pilots reach measurable P&L impact

A study of 1,100 large enterprises finds that while 78 % have deployed generative AI somewhere in the organisation, just 6 % can attribute a documented earnings effect to it. The differentiator is not model choice or budget: it is whether the deploying team owned the underlying process before automating it. Organisations that redesigned the workflow first saw a 3.4× higher rate of measurable impact than those that layered AI onto the existing process.

So what

Your AI budget is not the constraint — your process ownership is. Before approving the next pilot, ask who owns the end-to-end process and whether they have authority to redesign it. If the answer is unclear, the pilot will produce a demo, not a number.

MIT Sloan Management Review →

Siemens cut RFQ turnaround from 11 days to 32 hours

Siemens Digital Industries deployed an agentic workflow across its request-for-quote process, where engineers previously spent days reconciling specifications against a 400,000-part catalogue. The agents draft the technical response; engineers review and sign off. Notably, headcount stayed flat — the team took on 2.6× the RFQ volume instead, moving into segments it had previously declined to bid on.

So what

The value did not come from cost reduction but from capacity that unlocked new revenue. When you model the business case for an agentic deployment, a pure cost-savings frame will systematically undervalue it — and will lose to the CFO's discount rate.

Siemens / Handelsblatt →

Weekly Short News

EU Commission opens consultation on compute sovereignty

The Commission published a consultation paper acknowledging that 92 % of frontier-model training compute used by European firms sits on US or Chinese infrastructure, and floating procurement preferences for EU-hosted capacity.

So what: If procurement preferences land, vendor selection becomes a regulatory question, not just a technical one. Worth asking your CIO now where your inference actually runs.

European Commission →

Anthropic and OpenAI both raise enterprise seat pricing

Two of the three major providers raised list pricing for enterprise tiers within the same fortnight, in both cases bundling agentic capabilities that were previously priced separately.

So what: The era of deflationary AI pricing is pausing at the application layer even as token costs fall. Budget for flat-to-rising per-seat cost in 2027 planning.

The Information →

German works councils win co-determination ruling on AI monitoring

The Federal Labour Court ruled that AI systems capable of inferring individual performance require works council approval even when performance monitoring is not their stated purpose.

So what: In Germany, 'we're not using it for monitoring' is no longer a sufficient defence. Loop in your works council at design time — retrofitting consent after rollout is where these projects die.

Bundesarbeitsgericht →

AI Player of the Week

HeyGen

The video-avatar company crossed $100M ARR with roughly 200 employees, driven almost entirely by enterprise localisation: training and internal-comms content produced once and rendered into 40+ languages with a synthetic presenter. Its fastest-growing segment is not marketing but HR and compliance training.

So what: The interesting signal is where the demand came from — not the creative department, but the functions with a translation backlog. Look for the same pattern internally: your highest-value AI use case is probably sitting in a queue nobody calls innovation.

HeyGen →

Tip of the Week

Give your AI assistant a written scope, not just a prompt

Most executives use AI assistants conversationally and re-explain their context every session. Instead, write one page: your role, your recurring decisions, your standing constraints, and how you want pushback delivered. Paste it at the start of every working session, or store it as persistent instructions if your tool supports it. Users who do this report the largest single jump in output quality — larger than any model upgrade.

Found this useful? Forward it to a colleague — they can subscribe here.

You are receiving this email because you confirmed your subscription to MCC AI Weekly. Questions? Just reply to this email — it reaches us directly.

Imprint · Privacy Policy · Unsubscribe