New architectures for capital markets
We started Telescope because we recognized that large language models would enable entirely new architectures in capital markets. Not incremental improvements to existing workflows, but fundamentally different customer experiences that couldn’t exist before this technology shift.
The retail investment landscape has transformed over the past decade. Fractional shares removed minimum capital barriers. ETFs democratized diversification. Commission-free trading eliminated friction. Retail inflow has never been stronger. There’s no longer any meaningful excuse for not having a portfolio of some kind.
But access to markets has created new problems that the industry hasn’t solved.
The paradox of choice
When you remove barriers to entry, you expose the underlying complexity. Investors now face thousands of ETFs, tens of thousands of listed equities, multiple asset classes across global exchanges. The infrastructure works brilliantly. The decision-making process doesn’t.
This manifests in two ways:
Investor paralysis. People freeze when confronted with overwhelming choice. They know they should invest, but they don’t know what to buy. So they don’t buy anything. Cash sits idle while they wait for clarity that never arrives.
Irrational deployment. Others make impulsive decisions driven by headlines, social sentiment, or fear of missing out. They chase momentum without strategy, concentrate risk without realizing it, and exit positions emotionally when volatility hits.
Even among those who take the responsible route—building simple ETF portfolios—you find concentration risk and suboptimal allocation strategies. A 60/40 stock-bond split sounds sensible, but which stocks? Which bonds? What about sector exposure, geographic weighting, currency risk? The details matter, and most retail investors don’t have the tools or knowledge to work through them systematically.
Access alone doesn’t solve the decision-making problem.
What foundation models enable
Large language models represent a shift in what’s computationally possible. Not because they’re smart in some general sense, but because they can coordinate multi-step reasoning processes across unstructured information at scale.
This matters for finance because portfolio construction is exactly that: a multi-step reasoning process that requires synthesizing client circumstances, market conditions, risk frameworks, allocation constraints, and instrument-level details into a coherent recommendation. It’s not a lookup problem. It’s not a rules engine. It’s structured thinking applied to context-specific inputs.
Before foundation models, automating this process meant either:
- Reducing it to rigid rules (model portfolios, risk tolerance questionnaires)
- Relying entirely on human advisors (expensive, doesn’t scale)
Now there’s a third option: agentic workflows that reason through the same stages a human advisor would, but do so in minutes instead of hours, at marginal cost instead of hourly rates, while maintaining transparency about every decision made along the way.
This is what we’ve been building.
What we’ve learned
Over the past few years, Telescope has shipped multiple products that leverage agentic architectures in production finance environments. While our platform spans investment discovery, research automation, news curation, and compliance orchestration, two products in particular demonstrate the core principles we’ve developed.
Ripple turns prompts or news events into diversified instrument baskets. It interprets intent, identifies relevant instruments, applies diversification logic, and runs compliance checks before returning results. Ripple has generated over 500,000 baskets and executed more than 2 million compliance validations across our partner platforms.
Insights runs focused agents that monitor individual assets over time. For an equity like Tesla, it tracks earnings reports, quarterly filings, product launches, and news events, synthesizing signal from noise. For an ETF like EEM, it monitors macro trade flows, emerging market dynamics, and geopolitical developments. For a commodity like gold, it tracks supply fundamentals, central bank activity, and inflation expectations. We’re currently rolling Insights out to large listed broker platforms globally.
These products taught us how to build reliable agentic systems in regulated environments: how to orchestrate multi-step reasoning, validate outputs against compliance frameworks, expose decision traces for auditability, and manage behavioral consistency across model updates.
Atlas: agentic advice infrastructure
The same architecture that powers our existing products can be applied to portfolio construction and financial advice.
Atlas is an AI-powered advice engine that generates tailored portfolio recommendations through structured reasoning workflows. It doesn’t retrieve pre-built model portfolios or apply static rules. It reasons through the client’s circumstances—risk tolerance, investment goals, existing holdings, constraints—and constructs allocations designed specifically for them.
The process involves over 600 individual reasoning steps. Atlas works through the same stages a human advisor would: assess the client, evaluate their situation, formulate a strategy, construct a portfolio, allocate across assets, document the rationale. This takes 3 to 10 minutes depending on complexity. The output is a draft Statement of Advice that can be reviewed, validated, and presented to clients.
Every decision is logged. Which assets were considered. Why certain options were selected. How competing constraints were balanced. If the system can’t explain why it made a recommendation, the recommendation shouldn’t be trusted.
Atlas operates within the parameters each firm defines: approved investment universe, house views, allocation constraints, custom rules. A firm focused on REITs and commodity producers can configure that as their core framework. A firm with rounding rules for large portfolios or preferences for domestic ETFs can encode those directly. Atlas recalls this environment at every stage of the reasoning process.
The advisor remains at the center. Atlas handles analytical workload—scenario modeling, diversification optimization, risk assessment—but every recommendation flows through a human who validates it, contextualizes it for the client, and makes the final call. Clients need human judgment and reassurance, especially on decisions that will shape the rest of their lives. Preserving that relationship while removing analytical bottlenecks is the correct architectural choice.
Behavioral monitoring as fiduciary responsibility
When you deploy foundation models in financial decision-making, model selection becomes a fiduciary question. Which model are you using, and why?
Most institutions default to whatever provider has the best knowledge benchmarks or the most recognizable brand. But selecting a foundation model for financial advice is more like hiring an advisor than choosing a database. You need to understand behavioral characteristics, not just performance metrics.
When you hire a human advisor, you assess their decision-making patterns. How do they respond to market stress? Are they prone to overconfidence or excessive caution? Do they show consistency across similar client scenarios? Do they exhibit biases toward certain asset classes or investment styles?
The same standards should apply to foundation models.
Regulators should be asking firms: what model did you select, and why? The answer needs to go beyond “it scored highest on MMLU” or “it’s the industry standard.” You need to demonstrate that you understand how the model behaves under different market conditions, how its recommendations compare to alternative providers, and whether it introduces systematic biases or correlation risks into client portfolios.
This is why we built BEAM (Behavioural Benchmarking), a methodology for quantifying foundation model behavior in finance-specific contexts. We test how models respond to market stress scenarios, allocation decisions under competing constraints, and portfolio rebalancing triggers. We measure reasoning diversity across providers to identify convergence risks.
One finding from this work: three out of four major foundation model providers cluster tightly when responding to identical market conditions, within 1 to 12 degrees of separation in reasoning space. When models converge this strongly, institutions using them will generate correlated portfolio shifts during exactly the moments when differentiation matters most for stability. This is a measurable exposure that belongs in risk management frameworks.
But convergence is just one example. The broader point is that behavioral monitoring needs to be systematic, ongoing, and treated with the same rigor as any other fiduciary responsibility. We pre-test models for bias, validate reasoning diversity, and actively manage which foundation model we use based on behavioral performance. If a model shows troubling conformity, systematic bias, or degraded reasoning quality, we adjust prompts, fine-tune on additional data, or switch providers entirely.
Model behavior isn’t static. Providers update their systems constantly. What worked well in June may perform differently in December. Continuous monitoring is the only defensible approach.
What’s now possible
The technology foundation has shifted. Experiences that weren’t feasible three years ago are now practical to build and deploy at scale.
You can offer personalized portfolio construction to clients with $5,000 in assets, not just $500,000. You can provide allocation recommendations tailored to specific circumstances—dependents, income streams, existing holdings—without requiring hours of advisor time per client. You can explain every decision made in the process, not just present final outputs. You can do this in any language, across any major market, for any asset class your firm supports.
Advisors can focus on the parts of the relationship that require human judgment, empathy, and trust, while systems handle the analytical heavy lifting. As these systems mature, the cost of delivering high-quality financial guidance falls. Not because we’re cutting corners, but because we’re using technology to handle tasks that don’t require human cognitive load. The outcomes improve because reasoning becomes more systematic, biases get audited and corrected, and edge cases get caught before they reach clients.
The mission
We’re building infrastructure for a new generation of investor experiences. Experiences that leverage foundation models safely within regulated environments. Experiences that preserve human oversight while eliminating bottlenecks that prevent people from getting the guidance they need.
Capital markets have done a remarkable job providing access. Now we need to help people navigate what they have access to. That requires new architectures, not just better interfaces on old systems.
Atlas is one exploration of what becomes possible. We’ll continue investing in research to ensure these systems remain safe, unbiased, and effective. We’ll continue partnering with institutions willing to pilot new approaches in real advisory environments. And we’ll continue working with regulators to align our platforms with the transparency and accountability standards this industry requires.
The technology has shifted. What we build next is up to us.
Luc Pettett
Founder & CEO, Telescope
Atlas is a research and development project. It is not a licensed financial advice product and is not available for deployment to end clients. Telescope is working with licensed financial services providers globally to navigate regulatory pathways appropriate to each jurisdiction.