Image by Google Nano Banana
I have been fascinated for a while about distribution and connectivity in software, as well as a big proponent of data driven metrics in operating efficiencies. Calculating the opportunities here has long been an area of interest. With the huge focus on whether AI was having a material impact on the ecosystem I started following reports and research on returns here.
The lack of ROI in AI has fuelled a lot of debate, certainly the big picture is not showing material impacts yet. But its also clear there are many benefits and it is early in the adoption curve, main street is always late to the table in adoption. But my world is software businesses specifically AI and SaaS in B2B. Tech often disrupts its own industry first.
To start to address the gaps here I started looking for indications, market reports and research that demonstrate business value from AI and associated infrastructure demands.
(For the bigger picture large scale corporate enterprise context on AI ROI follow Amin Mrini The AI Value Gap here on Substack)
Each month I run a number of automated jobs to gather insights on data which indicates AI enabled sales growth, engineering productivity and compute consumption, and each week I update and add new parameters and sources to the list.
The jobs I run verifies every figure at its source. So investors, operators and leadership teams can make decisions on numbers that gives them the insight they need. However..
All data is from freely available sources, I do not have routine access to paid enterprise software data analytics. This data is aggregated and verified by Claude and only a subset is human reviewed, so please be aware it may contain errors.
ROI on AI - November 2026
For two years the AI productivity debate has run on vendor claims and survey sentiment. This month, something changed: the controlled evidence started to land. A randomised trial on revenue. A 718-firm difference-in-differences study on engineering output. Production telemetry from Microsoft on how agentic workloads actually consume compute. Read together, they say the same thing. AI gains are real, and they are measured in tens of per cent, not multiples. And in every case, the gain only reaches the P&L when the system around the tool changes: the sales motion, the review process, the way inference is routed. For investors, that is a more useful finding than any hype curve. It tells you where value creation will be decided: not in the procurement of AI, but in operating model redesign.
AI’s commercial impact: where we can find a lift
The cleanest revenue evidence I’ve seen this year is a randomised field trial posted to arXiv on 18 October. Across 14.8 million visitors on 100 live websites in 14 industries, an AI sales agent raised revenue per visitor by 10.0% against a no-agent control (95% CI +5.1% to +14.9%). That echoes earlier field experiments at a cross-border online retailer (Fang et al., revised June 2026), where generative AI workflows produced sales effects from zero to 16.3%, mostly through higher conversion. Both sets of evidence are largely B2C, and the October paper is a single-author preprint, not peer-reviewed, whose author is affiliated with the vendor of the agent tested.
The counterpoint is in the same paper. On 20 sites with staffed live chat, the agent converted engaged visitors at 11.26% against 11.20% for human associates: non-inferior, not better. On B2B, Octane11’s telemetry across 400 million sessions shows AI search touching around 9% of closed-won deals while supplying under 2% of search referrals, a sign that AI influence is concentrated in high-intent buyers. Yet Salesloft’s survey of 500 US revenue leaders found only 20.6% describe their AI strategy as production-ready with measurable outcomes. Company disclosures tell a similar story: The AI Value Gap’s census of 535 S&P 500 and FTSE 100 companies found only 31% can show a realised, quantified AI result in production (August 2026).
In due diligence: can the team show AI-influenced pipeline and conversion, against a control or a baseline?
Engineering operating efficiency: more output, same finished work
One of the most inciteful engineering papers of the year, is Fiona Chen and James Stratton’s Harvard study of 718 firms and nearly 726,000 workers using Jellyfish data from 2021 to March 2026. Adopting AI coding agents raised pull requests per worker-month by 23%, commits by 20% and lines of code by 30%. Microsoft’s own study of its early-2026 rollout of CLI coding agents found adopters merged roughly 24% more PRs.
However in the Harvard data, Jira issues resolved per worker showed no significant change, and the authors can rule out gains larger than 12%. PR open-to-merge time rose 49%, from a baseline of about seven days, and review comments per PR rose 35%. Faros AI’s telemetry from 22,000 developers points the same way: monthly incidents up 125.4% and time in QA up 300.6% among high-adoption teams. That’s vendor data, but it is consistent with the academic evidence.
For a PE- or VC-backed business, this is the gap between an activity metric and a value metric. A board pack showing PR volume up a quarter is not a productivity story if cycle time and incidents are also rising. What I’d want to see: issues or features delivered per engineer, PR cycle time, change failure rate, and evidence that review capacity and test automation were funded alongside the licences. The constraint has moved from writing code to verifying it.
Compute efficiency: cheaper token consumption, dearer capacity
Efficiency innovation is compounding at every layer. At the model layer, Anthropic launched Claude Haiku 5.5 on 7 October at $0.10 per million input tokens and $0.50 output, 90% below Haiku 4.5 for requests up to 100,000 tokens. At the training layer, Microsoft AI reports its MAI-Base-2 recipe was 4.2–12.3x more FLOPs-efficient than DeepSeek-V4-Pro-Base, on a 38-day run on 16,000 GB200s. At the serving layer, a study of more than 121 million Microsoft 365 model calls found about 80% of agentic context tokens are served from the prefix cache, and that cache reuse falls from 89% to 29% when consecutive calls land on different deployments.
That last point is the one I’d underline. For agentic products, cost per task is now determined as much by routing and caching architecture as by the list price of the model.
The counter-evidence is in hardware. The Ornn Compute Price Index, built from volume-weighted transacted rental prices, had H100 SXM at $3.17 per GPU-hour on 26 October, up 24.8% in a month, and B200 at $9.76, up 21.4%. Tokens are getting cheaper while the capacity underneath them gets more expensive. In diligence on any AI-heavy business I’d ask for cost per unit of customer outcome, cache hit rates and GPU utilisation, and how much capacity is contracted versus exposed to spot pricing.
What to do with this
One action per theme:
Commercial: require a controlled baseline for every AI revenue claim. AI should add value, but it can also indicate changing buyer behaviour.
Engineering: report delivered work (issues, features, cycle time, change failure rate) alongside AI adoption, and fund review and test capacity before more generation.
Compute: make utilisation and efficiency a separate leadership cost metrics for AI-heavy products, and review GPU contract exposure now that rental prices are rising.
If you want the numbers without the commentary, the monthly digest is on the Ingeniva Consulting LinkedIn page.
Gemma Whitehouse is a fractional and interim CTO/CPO working with PE and VC-backed B2B software and AI-native businesses. She publishes monthly benchmark digests and original analysis at ingeniva.ai. Follow Ingeniva Consulting on LinkedIn for the monthly benchmark summary. This summary is aggregated and commentary drafted by Claude. It may contain errors.
Sources
“Selling in the Wild: Benchmark and Randomized Field Evidence for a Sales Agent Trained on Commercial Conversations”, B. Tannenbaum, arXiv:2610.25461, 18 Oct 2026 — https://arxiv.org/abs/2610.25461
“Generative AI and Sales Productivity: Field Experiments in Online Retail”, Fang, Yuan, Zhang, Donati & Sarvary, arXiv:2510.12049, v6 29 Jun 2026 — https://arxiv.org/abs/2510.12049
State of B2B AI Search Q2’26, Octane11, 29 Jul 2026 — https://www.octane11.com/insights/insights-state-of-b2b-ai-search-q2-26/
2026 US Revenue Benchmark Report, Salesloft, 2 Sep 2026 — https://www.salesloft.com/company/newsroom/salesloft-releases-2026-us-revenue-benchmark-report
“No.40: The Enterprise AI Value Map”, Amin Mrini, The AI Value Gap (Substack), 31 Aug 2026 —
“Artificial Intelligence in the Firm: Bottlenecks in Software Production”, F. Chen & J. Stratton, Harvard University, 4 Aug 2026 — https://fion.ac/jellyfish.pdf
“Adoption and Impact of Command-Line AI Coding Agents”, Murphy-Hill, Butler & Savelieva, arXiv:2607.01418, 1 Jul 2026 — https://arxiv.org/abs/2607.01418
AI speed trap takeaways (blog), Faros AI, 18 Sep 2026 — https://www.faros.ai/blog/ai-speed-trap-takeaways
Introducing Claude Haiku 5.5, Anthropic, 7 Oct 2026 — https://www.anthropic.com/claude-haiku-5-5
Frontier pre-training at Microsoft AI, Microsoft AI, 15 Oct 2026 — https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/
“Understanding KV Residency in Production-Scale Agentic LLM Workloads”, Arya et al., arXiv:2610.22863, 16 Oct 2026 — https://arxiv.org/abs/2610.22863
Ornn Compute Price Index, Ornn, settled 26 Oct 2026 — https://data.ornn.com/