Avrom Gilbert

The through-line

Gilbert's argument returns again and again to one gap: large language models know the world but do not know your business. "Genai has an amazing understanding of the whole world and very strong reasoning skills but it doesn't actually know very much about the details of what's driving my business because it deals with language and has read the whole internet but hasn't really looked into my data" [1]. Everything else he says follows from that diagnosis. The secrets, in his telling, sit in structured operational data, "my CRM database my ERP database my customer transactions my IoT," and the work worth doing is educating an LLM on those patterns until it becomes an expert on a specific company [1][2]. He frames the payoff in deliberately popular terms: the Jarvis of the Iron Man films, an AI "which deeply understands everything including his business and operations," able to answer questions like how to reduce costs or how to reach the customers most at risk of churning [1].

He splits the history into two eras and places himself in both. In the pre-generative period, when he was at Similar Web and Seeking Alpha, "we didn't really have great AI for decision-m we didn't really know of any good tools," even though SparkBeyond was already selling AI for decision-making to hundreds of large enterprises, including Santander, Equinor and HDFC, and to partners such as McKinsey, by applying AI to structured data to find what was dragging down key KPIs [1]. The post-ChatGPT period of the last two to three years added something different in kind, and the direction he now argues for is the merger of the two: world knowledge and reasoning from the LLM, operational specificity from the data [1][3].

On what generative AI actually changed

The first and most widely felt shift he calls rapid decision support, the AI everyone already knows: ask questions about the world and get answers in seconds rather than weeks. He speaks from experience here, recalling "many of the times I spent days or my team spent even weeks doing research at similar web seeking alpha and it can be done in minutes now" [1]. That is genuinely valuable, but he is careful about its limits: it gets you "facts about the world at scale," which is not the same as facts about your company [1]. Asked today why costs are rising on a specific product, ChatGPT or Perplexity would return "generic advice from the internet," and his test for that is blunt: "I certainly wouldn't hire an employee to give me generic generic advice and I say I don't really want AI to do that either" [1]. Moving businesses past generic AI-generated insight, by grounding models in CRM, ERP and IoT data, is the shift he argues for [2][3].

On agents and the app store analogy

The second evolution, already underway, is people trusting agents to make decisions rather than making them themselves, in personal life and at work alike, from asking an agent to decide dinner and order the ingredients through to handling customer service problems, with products such as Agentforce from Salesforce as the visible edge of it [1]. He explains the momentum behind agentic AI not as hype but as a collapse in the cost of building: LLM quality is "advancing at a shocking speed," models now have "incredible knowledge, intuition and reasoning," and their ubiquitous availability means "there's a very low bar for creating agencies" [1]. Soon, he expects, a simple natural language instruction will be enough for almost anyone to create a usable agent [1].

His analogy is the smartphone app ecosystem. He remembers trying to convince Nokia or telecoms carriers to put a startup's game on a feature phone and finding it almost impossible; the App Store launched with around 500 apps, passed 80,000 within a year and a million after five [1]. Building a real app still took investment, and he recalls the effort of building both iPhone and Google apps at Seeking Alpha in the early days [1]. Agents will follow the same curve with an even lower barrier: "if I want to book 30 there'll be an agent for that I want to pay my bills do my taxes answer questions for my customers," until using agents is as second nature as using phone apps [1].

On the "always optimized" loop

He defines the capability plainly as using AI constantly in the background to analyze data, work out what is driving problems in the operating environment, and propose actions that agents can then execute where they exist [1]. He argues for it by first describing the alternative. A telco or bank sees churn rising on a dashboard; when the data science or analyst team has time, they hunt for insights; those go to marketing, which reviews them, proposes actions, and eventually sends emails or promotions. The process is "lengthy and infrequent and expensive," and he reaches for Churchill on democracy to describe it, "the worst for all government except for all other types of governments which have been tried," concluding that "this is a terrible way to optimize your business. But without AI, it's actually the best and most practical way to do that" [1].

The always optimized version does the same work systematically and continuously. The AI divides the customer base into micro segments by location, age and behavior, much as an analyst would, derives insights on which groups are leaving and why, then hands those insights to an LLM that "very rapidly then becomes an expert on what's driving customers to leave the business" and drafts custom emails per customer [1]. The cycle runs as often as the data changes, with no waiting for people to free up [1]. He generalizes the pattern to fraud detection on bank transactions, cost analysis as costs rise, and energy optimization [1]. The core claim under all of it is a method claim: "the key to all of this is that we figured out the way to educate an LLM how to become an expert in the relevant parts of your business" [1].

On where autonomy stops, for now

Gilbert is consistent that trust is the gating factor rather than capability. Agents making decisions instead of people "is just starting," and in the churn example he draws the line explicitly: emails "would generally need to be approved by a marketing team," with direct sending by the AI arriving "in the future with agents once they could be trusted" [1]. He expects the boundary to move, describing always optimized as something that will feel "perfectly natural for organizations in the coming years as we grow to trust LLMs and generative AI much more" [1]. He also acknowledges the source of the earlier skepticism, that the first versions of generative AI "made a lot of mistakes and had hallucinations" [1].

On operational and industrial use cases

Beyond marketing and churn, he keeps returning to physical operations as the clearest illustration of business-specific knowledge. A factory operator wants an AI that understands the details of its machines well enough to recognize when temperature or pressure crossing a threshold has historically predicted failure, alert an engineer, and send that engineer the maintenance procedure, so "that machine never fails" [1]. The revenue equivalent is granularity at a scale no team can staff: knowing that "middle-aged men over the age of 30 tend to buy organic goods on a Thursday and it's a good time to give promotions" is useful on its own, but "if you multiply that by a thousand or 10,000 different micro segments," each generating its own automatic messages and promotions, "that's a gamecher" [1].

Takeaways

• The bottleneck for enterprise AI is not reasoning but context: LLMs have read the internet and still know nothing about a specific company's CRM, ERP, transaction and IoT data [1][2].
• Generic answers fail the hiring test. Gilbert would not hire an employee who gave generic advice from the internet, and holds AI to the same standard [1].
• Agentic AI is gaining momentum because the cost of building agents has collapsed, repeating the App Store curve from roughly 500 apps at launch to over a million in five years [1].
• The manual analytics-to-marketing chain is "lengthy and infrequent and expensive," which is why continuous background analysis and action beats periodic dashboard-driven investigation [1].
• Autonomy will arrive gradually: today AI proposes actions and humans approve them, with direct execution waiting on trust rather than on capability [1].
• The same always optimized loop is meant to generalize across churn reduction, fraud detection, cost analysis, energy optimization and predictive machine maintenance [1].
• Value scales through micro segmentation, applying thousands of small, specific behavioral patterns automatically rather than a handful of broad ones manually [1].