--- title: "Ally Built Six AI Customers Before Shipping" slug: "ally-bank-ai-agent-personas-customer-research" author: "Mikel Amigot" date: "2026-08-31 18:00:00" category: "Premium" topics: "AI agent personas, synthetic customers, financial services AI, customer research, product testing, banking AI, self-hosted AI" summary: "Ally's Personas project built six AI agent personas modeled on its 11M+ customers, so teams can gather user feedback instantly instead of waiting on a research cycle. The interesting part is the inversion: most enterprises deploy AI to serve customers, not to understand them first." banner: "" thumbnail: "" linkedin: | Most enterprises ask: how do we deploy AI to serve customers? Ally asked a different question first: how do we use AI to understand customers before we build anything? Their Personas project launched in Q4 2025 after 16 months of development by Ally Tech Labs, working with LangChain. Six AI agent personas — Alex, Charlie, Jessie, Jordan, Logan and Sam — each modeled with its own financial background, goals, values, digital habits and trusted brands, drawn from the characteristics of more than 11 million customers. Teams use them as "the voice of the customer" to gather feedback early, often, and instantly, rather than waiting weeks for a research panel. The inversion is the lesson. Customer research is usually the bottleneck that makes teams skip it: a real study takes weeks, so it happens once, late, after the concept is already expensive to change. A synthetic panel makes the cheap version of that loop available on day one. What it is not: a replacement for talking to actual people. A persona derived from your existing customers cannot tell you what the customers you failed to acquire wanted, and it will reproduce whatever bias is in the source data. Treat it as a fast first filter, not as evidence. The part worth thinking about structurally: a synthetic panel is built from your customer data. That model IS your customer base, encoded. It belongs inside your perimeter, not on someone else's platform. On ibl.ai you own all the code and the data — self-hosted in your own environment, model-agnostic across any LLM, with no per-seat pricing. #iblai #FinancialServices #BankingAI #CustomerResearch #AgenticAI #EnterpriseAI --- ## The Short Answer **Ally's Personas project uses six AI agent personas — modeled on the characteristics of its 11M+ customers — to gather product feedback instantly instead of waiting on a research cycle. The inversion is the lesson: AI applied to understanding customers before serving them. Because such a panel is built from customer data, it belongs inside your perimeter; on ibl.ai you own all the code and the data.** The project is worth studying less for the technology than for the sequencing decision behind it. ## What did Ally actually build? Six named AI agent personas: **Alex, Charlie, Jessie, Jordan, Logan and Sam**. Each carries its own financial background, goals, values, digital habits and trusted brands, representing characteristics drawn from Ally's more than **11 million** customers and prospective customers. The project launched in the **fourth quarter of 2025**, following **16 months** of development and testing by Ally Tech Labs in partnership with LangChain, [as reported by American Banker](https://www.americanbanker.com/news/allys-personas-is-on-of-the-innovation-of-the-year-honorees). Internally the personas serve as "the voice of the customer," letting employees gather user feedback **early, often, and instantly** rather than commissioning a study. Sixteen months is the detail most summaries drop. This was not a weekend prototype. ## Why is testing against synthetic customers useful at all? Because the bottleneck in customer research is almost never the analysis. It is the latency. A real research cycle — recruit, screen, schedule, run, synthesize — takes weeks. That cost means it happens once, late, after a concept is expensive to change. Teams skip it not because they doubt its value but because the calendar does not allow it. A synthetic panel changes what is available on day one. A product manager can put three variants of a savings-account flow in front of six personas in an afternoon and find the obvious failures before anyone builds anything. The value is not that the synthetic answer is as good as a real one. It is that the cheap, fast, imperfect answer arrives while the design is still cheap to change. ## What can a synthetic panel not tell you? This is where the discipline matters, and it is worth being blunt about the limits. **It cannot tell you about people who are not your customers.** A persona derived from your existing base encodes who you already acquired. It is structurally silent about everyone you failed to reach — which is usually the more valuable question. **It reproduces the bias in its source data.** If a demographic is underrepresented in your customer base, it is underrepresented in the model of that base, and the panel will confidently give you a majority view. **It cannot be surprised.** Real users do things nobody modeled — misread a label, use a feature for the wrong purpose, abandon a flow for a reason that never occurred to the designer. That is often the finding worth the entire study. **It is not evidence for a regulator.** A synthetic panel's opinion about whether a disclosure is clear does not establish that consumers found it clear. The right framing is a fast first filter that raises the quality of what reaches real users — not a substitute for reaching them. ## Where should a synthetic customer model actually run? This is the structural question underneath the technique, and it follows from what the artifact is. A persona panel built from customer data is not a generic tool. It is **your customer base, encoded** — behavioral patterns, financial characteristics, segment structure. Building it requires exposing that data to whatever system does the modeling.
Consideration Managed platform Platform you own
Customer data used to build personas Leaves your environment Never leaves it
The persona model itself Vendor's system An asset you keep
Regulatory posture Depends on the DPA Same perimeter as your core systems
Cost as teams adopt it Per-seat, multiplied by headcount Usage-based against a cap you set
For a bank, the middle rows are the ones examiners ask about. A model of your customers carries much of the sensitivity of the underlying data. ## Is this a broader pattern in banking? Yes, and Ally is early rather than alone. Reporting on synthetic customer testing notes JPMorgan Chase generating synthetic financial data to simulate market behavior for risk management and product design, and NatWest, Monzo and Santander building synthetic data ecosystems to train models. The common thread is not "chatbot." It is using generative models against internal data to shorten a research or risk loop — a category of use that never faces a customer and therefore never appears in adoption statistics about customer-facing AI. ## How does ibl.ai support this pattern? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** For internal research agents specifically: the customer data used to build the personas stays inside your environment, the resulting agents are yours, and every interaction logs to your own systems. Because pricing is usage-based rather than per-seat, a tool meant to be used casually by many product managers does not get rationed by license count — which is exactly how a research tool dies. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ## The sequencing is the insight The technique is replicable. The decision that produced it is the part worth copying: Ally spent 16 months applying AI to understanding its customers before applying it to serving them. Most enterprises run that order backwards, ship an assistant, and then try to find out whether anyone wanted it. *Related: [AI Agent Companies Landscape 2026](/blog/ai-agent-companies-landscape-2026) · [Legal Grew 108x. Governance Didn't Move.](/blog/legal-codex-adoption-108x-governance-gap)* ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.