# Making the Energy Decision Stack interesting to Demis Hassabis ## Core reframe This should not read like an "AI in energy" market memo. It should read like **a real-world benchmark for frontier reasoning systems in a root-node industry**. The strongest Demis-facing angle is: > This analysis maps where frontier models can become persistent infrastructure in energy — not by replacing the final decision-maker, but by compressing the evidence-prep stack around capital, regulation, permitting, planning, title, and project execution. That matters twice: first because these workflows are excellent testbeds for agentic reasoning with tools, memory, and citations; second because the same decision loops constrain the build-out of energy and compute that advanced AI itself depends on. --- ## Why this is interesting to Demis specifically ### 1. It connects to the "root-node problems" frame Demis is much more likely to care if this is framed as part of the machinery that governs **how fast energy capacity can actually get built**, not as a sector digitization story. Your current analysis already points toward power, permitting, siting, interconnection, project finance, and utility regulation as bottlenecks. That is the right bridge. The pitch is not: - "Energy companies can save analyst hours" The pitch is: - "The hidden bottlenecks behind energy expansion are evidence-heavy decision loops. Frontier models can compress those loops." That makes the work adjacent to AI infrastructure, scientific progress, and industrial throughput. ### 2. It is an eval environment for useful agency Your package is strongest when it looks like a benchmark suite for real work: - long-context reading across PDFs, spreadsheets, and messy packets - cross-document consistency checking - citation / provenance - exception handling - workflow memory across cycles - bounded human signoff in high-consequence settings That is much closer to what a frontier lab cares about than generic productivity language. ### 3. It supports the "AI as instrument" story Demis often gravitates toward AI as a scientific or cognitive instrument, not just automation software. This analysis becomes more interesting when you stress that it helps organizations: - ask more questions - run more scenarios - widen the search space - improve decision quality under constraint That is stronger than a headcount-arbitrage pitch. ### 4. It is human-in-the-loop by design One of your best insights is: **the seat survives, the prep stack compresses**. That is exactly the sort of story a safety-conscious frontier lab should like: - the lender still signs - the regulator still rules - the board still decides - the engineer still attests - the model handles the evidence assembly, first-pass synthesis, conflict detection, and drafting beneath the decision node That is a more serious and credible framing than "AI replaces analysts." --- ## What to change in the current package ### Change the headline Current flavor: - ~90% of AI-exposed wage dollars sit above field operations Better Demis-facing headline: - **Energy is a root-node benchmark for agentic reasoning, not just another enterprise vertical.** The 90% result is still useful, but it should become supporting evidence rather than the hook. ### Promote the frontier-lab material to the front Right now, the package already contains the right raw material: - the frontier-lab brief - the partnership map for frontier labs - the Jevons / more-questions-than-answers insight - the rubric for inter-rater scoring - the proof trace for treasury / lender readiness For Demis, that material should move from supporting appendices to the front door. ### Lead with the proof trace, not the thesis The most Demis-relevant artifact in the package is the treasury proof trace. Why: - it shows tool-using reasoning on messy source materials - it includes error detection - it shows where the model succeeded - it shows where a human had to intervene - it implies an eval framework, not just a narrative That is much more compelling to a frontier lab than a polished static deck. ### Reorder the beachheads For a DeepMind audience, the most compelling wedges are not the most operationally mundane ones. Lead with: 1. utility planning / rate recovery 2. interconnection / permitting / regulatory writing 3. treasury / lender readiness / project finance 4. EPC / project controls 5. title / non-op / minerals That order emphasizes energy-system throughput and high-consequence agency. ### Rename "sensitivity analysis" in the pitch layer Keep the analysis itself as-is, but when pitching upward, translate it into: - robustness of the directional claim - confidence bounds on the benchmark - signal stability under scoring perturbation That sounds closer to an eval discipline and less like a consulting appendix. --- ## Suggested opening paragraph Demis, this is not a sector deck about who gets automated first. It is a map of where frontier reasoning systems can become persistent infrastructure in a root-node industry. The key finding is that the near-term surface is not field robotics but the evidence-heavy decision loops around capital, regulation, planning, permitting, title, and project execution. Those loops are exactly the kind of work that tests long-context reasoning, tool use across PDFs and spreadsheets, workflow memory, citation, exception handling, and calibrated human handoff. That makes energy interesting not only as a commercial market for reasoning, but as a practical benchmark for useful agency in a sector that also constrains the physical expansion of AI itself. --- ## Suggested titles - **Energy as a Root-Node Benchmark for Agentic Reasoning** - **Where Frontier Models Become Persistent Infrastructure in Energy** - **Not Energy AI — Useful Agency in Capital, Regulation, and Operations** - **The Evidence Stack Behind Energy Is an AGI Problem in Disguise** - **A Real-World Eval Environment for Frontier Models: Energy Decision Loops** --- ## Three claims most likely to resonate ### Claim 1 This is not mainly a robotics story. The early surface is document-heavy, evidence-heavy, and review-heavy. That makes it an immediate testbed for model reasoning rather than a hardware waiting game. ### Claim 2 The value is not only efficiency. It is expanding the number of scenarios that can be explored. That maps to AI as an instrument for search, hypothesis generation, and decision expansion. ### Claim 3 Energy is special because the customer and the bottleneck are entangled. The same workflows that are likely to buy reasoning also influence how fast power, transmission, permitting, and capital formation move. --- ## What to de-emphasize Avoid making this sound like: - a labor-displacement deck - a generic enterprise TAM memo - a static consulting taxonomy - a score-obsessed ranking exercise Demis is unlikely to care about "which clerical role automates first" unless it is attached to a deeper systems insight. Also de-emphasize low-value back-office automation examples in the opening sequence. Keep them in the benchmark, but do not lead with them. --- ## The best one-line ask > Treat this as a candidate benchmark suite for Gemini-class agents in high-consequence, evidence-rich energy workflows — and test where the model still breaks. That is a much stronger ask than "this is a good market." --- ## If you want a 30-second spoken version "I wouldn't pitch this to Demis as an energy market map. I'd pitch it as a benchmark for useful agency in a root-node industry. The interesting thing is not who gets automated; it's that the critical bottlenecks in energy are messy, evidence-heavy decision loops — rate cases, interconnection, lender packs, permitting, project controls — that demand long-context reasoning, tool use, citations, and human handoff. Your package already has the right ingredients: a robustness layer, an eval rubric, and a proof trace. The move is to foreground those and make the whole thing feel like a frontier-model testbed, not a vertical SaaS thesis." --- ## Concrete edit list 1. Replace the current top-line hook with the root-node / useful-agency framing. 2. Put the treasury proof trace above most of the narrative exposition. 3. Add a section called **Why this is a benchmark, not just a vertical**. 4. Move utility / interconnection / regulatory workflows ahead of generic back-office workflows. 5. Turn the sensitivity appendix into a short confidence panel in the executive layer. 6. Add one explicit slide or section connecting energy workflow compression to AI infrastructure build-out. 7. Use language like **agentic reasoning**, **evidence operating system**, **bounded autonomy**, **workflow memory**, and **human-governed deployment**. --- ## The one sentence to keep repeating **The point is not that energy uses AI. The point is that energy contains some of the clearest real-world tests of whether frontier models can become reliable, useful agents in high-consequence systems.**