"What's wrong with my portfolio?" is a question UK investors started asking ChatGPT in 2023 and have been getting unsatisfying answers to ever since. The answers were generic because the question was generic — the model could see the question but not the portfolio. That gap between "AI that knows my situation" and "AI that knows about investing in general" is the gap most retail AI tools have not yet closed.
In 2026 the landscape has changed. The Model Context Protocol (MCP), introduced by Anthropic and now widely supported, lets AI assistants read data from your accounts when you explicitly authorise it. That changes what is possible. It does not — yet — change what is regulated, what is safe, or what AI is actually good at. This article walks through what works, what does not, and where AI portfolio analysis usefully sits in a UK investor's workflow.
This is not financial advice. Past performance does not guarantee future returns. Consider speaking to an FCA-authorised financial adviser for personalised guidance.
Why ChatGPT alone gives generic answers about your portfolio
A language model generates text based on patterns in its training data and the prompt you give it. If you paste in "Should I rebalance my SIPP?", the model has no idea what is in your SIPP, when you bought it, what it cost, or what your other accounts hold. The answer it gives — a list of generic considerations about diversification, fees, time horizon — is competent and applies to most portfolios precisely because it cannot apply to yours specifically.
You can paste a list of holdings into the prompt. People do. The problems are practical:
Stale data. The list you paste reflects a moment in time. By the time you have asked three follow-up questions, the prices have moved and the model is reasoning about a snapshot.
Limited context. A typical UK investor's holdings list, including units, cost bases, dividends, and corporate actions, is enough text to push the model into using context budget on data parsing rather than analysis. Even with long-context models, you spend tokens describing the portfolio rather than thinking about it.
No look-through. ChatGPT cannot reach into a fund and pull its holdings. If you tell it "I hold VWRL", it knows VWRL is the FTSE All-World UCITS ETF in the abstract but it does not know your specific exposure to Apple, Nvidia or Taiwan Semiconductor — let alone how that exposure aggregates across your VUSA and VEVE holdings as well.
No memory. Unless you are using a memory-enabled mode, the next session starts from scratch. You re-paste the portfolio. You answer the same questions about wrapper structure. The conversation is genuinely useful only for one-off questions, not ongoing engagement.
The result is that ChatGPT is excellent for general investing education, average for question-answering about specific named instruments, and weak for "what should I do about my portfolio" because it cannot see your portfolio in any structured way.
MCP — Anthropic's protocol for connecting AI to real data
The Model Context Protocol, released by Anthropic in late 2024 and adopted across the AI assistant landscape through 2025–26, is a standard for letting AI assistants connect to external data sources and tools. It is roughly the equivalent of "API standards" for the AI era — a common shape for how a model asks for data and how a server provides it.
A typical MCP setup has three pieces:
- The AI assistant — Claude, ChatGPT (which now supports MCP), or any compliant client.
- The MCP server — software you connect, which exposes a set of tools and data the AI can call. The server lives on infrastructure run by the service provider, not on the AI side.
- You — the user, who explicitly authorises the connection and decides which tools the AI can use.
When you ask Claude a question that needs data, Claude can call a tool exposed by the MCP server, get a structured response, and incorporate it into the answer. This is fundamentally different from "I pasted my holdings into the prompt": the data is fetched in structured form on demand, refreshed automatically, and stays out of the conversation log unless explicitly surfaced.
Invormed exposes seven read-only tools through its MCP server: portfolio summary, holdings list, allocation by wrapper, allocation by sector, look-through holdings, performance history, and dividend income. Each tool returns structured data that Claude (or any MCP-compatible client) can reason about. The portfolio data is served from your authenticated Invormed account — you authorise the connection once, the AI can read your portfolio for the lifetime of the consent, and you can revoke access at any time.
The point of MCP is not that "AI can do more things". The point is that AI can answer specific questions with current, structured data, without you needing to manually copy-paste your portfolio into every conversation.
Read-only versus read-write — and why we do not trade for you
MCP tools can in principle be read-only (the AI fetches data) or read-write (the AI takes actions). For investments, this distinction is not a detail.
A read-only MCP server lets the AI see your portfolio. It cannot place trades, move cash, modify holdings, or change account settings. The worst it can do is show you a summary you do not like.
A read-write MCP server in an investment context would let the AI execute trades on your behalf. This is technically possible. It is not what Invormed does, and there is a deliberate reason.
UK investment advice and execution are FCA-regulated activities. A platform that takes trading instructions from an AI agent acting on a customer's portfolio is in execution-only territory at best, and arguably in advice territory once the AI is making the trade selection. The current FCA position on AI-driven execution is unsettled — the regulator has consulted on it and supervisory expectations are still forming. Operating in that grey zone with retail customer money is not where a consumer product should be in 2026.
The other reason is more pragmatic. Language models hallucinate. They confidently produce answers that are wrong. The cost of a hallucinated paragraph in a chat is that you read it, notice it is wrong, and discard it. The cost of a hallucinated trade is real money. Until the models are demonstrably stable enough that the second category disappears — and they are not, in 2026 — the right design is to keep the AI on the analysis side and the human on the execution side.
The shape of useful AI in investing today is: AI can read everything, summarise everything, surface things you missed; you make the decision and place the trade.
What questions you can usefully ask
Once an AI assistant can see your real holdings through MCP, a different class of question becomes useful. Worked examples of what people actually ask:
"What is my real exposure to Apple across all my accounts?"
The AI calls the look-through tool. It returns: 4.7% of your total portfolio, made up of direct exposure inside VWRL (in your ISA), VUSA (in your ISA), and a small direct holding in your GIA. The AI summarises: "Apple is your largest single-name exposure. About four-fifths of it comes from ETFs rather than direct holding. If you wanted to reduce concentration, the lever is fund choice rather than the direct position."
"Which of my holdings have not paid a dividend in twelve months?"
The AI calls the dividend income tool, filters by the absence of payments in the last twelve months, returns the list. Useful for an income investor reviewing whether a yield assumption still holds.
"What is my GBP performance this year, including FX?"
The AI calls the performance tool, returns the time-weighted return in GBP terms with FX effects included separately. It can break it down by wrapper, by asset class, or by individual holding.
"Show me where my SIPP and ISA hold the same fund."
The AI cross-references holdings across wrappers. Useful when your SIPP and ISA have drifted into the same allocation by accident — a common state for investors who set both up at different times with similar logic.
"If I sold my GIA holding in VUSA today, what would the CGT implication be?"
The AI calls a tool that calculates realised gain against your annual exempt amount. It cannot tell you whether to sell — that depends on your wider plan — but it can do the maths instantly so the decision is informed.
The pattern is that AI is most useful when it is doing the data-handling work that you would otherwise do manually, and when the question has a structured answer that benefits from being phrased in plain English on both ends.
Pure AI tools versus robo-advisors versus portfolio intelligence
These are three different categories often discussed together. They overlap but solve different problems.
Pure AI tools — ChatGPT, Claude (without MCP), Gemini. General-purpose assistants. Strong at education, weak at portfolio-specific advice without data access. Free or near-free. Not regulated as financial services because they offer no regulated activity — they are language models, not advisors.
Robo-advisors — Wealthfront and Betterment in the US, Nutmeg, Moneyfarm and the now-discontinued Vanguard UK Personal Financial Planning service in the UK. They take your money, build you a portfolio, and rebalance it according to a model. They are FCA-regulated as discretionary investment managers. They are not really AI in the modern sense — they are rules-based portfolio construction that gets called "AI" in marketing copy. Useful for hands-off investors who want a managed solution. Not useful if you want to keep self-directing and just want better analysis on the portfolio you already have.
Portfolio intelligence — Sharesight, Snowball Analytics, Invormed. They do not manage your money or place trades. They aggregate your holdings, analyse them, and surface insights. The newer entrants (Invormed) layer AI on top of that aggregation through MCP. Not regulated as advice because they do not give advice — they show you data about your own portfolio and answer factual questions about it. You make decisions; the platform helps you see the data clearly.
| Category | Example | Manages money | Gives advice | Reads your portfolio | UK regulation |
|---|---|---|---|---|---|
| Pure AI assistant | ChatGPT, Claude (without MCP) | No | No (general info only) | No | None (not a regulated activity) |
| Robo-advisor | Nutmeg, Moneyfarm | Yes | Yes (discretionary) | Yes (their own platform) | FCA discretionary investment manager |
| Portfolio intelligence | Sharesight, Invormed | No | No | Yes (with your consent) | Generally not regulated; AISP if using Open Banking |
| AI portfolio intelligence | Invormed via Claude (MCP) | No | No (information service) | Yes (read-only via MCP) | Same as portfolio intelligence |
The choice depends on what problem you actually have. If you want to delegate the whole investing job, a robo-advisor or human IFA is the right shape. If you want to keep self-directing but make better-informed decisions, AI-augmented portfolio intelligence is the right shape. If you just want general investing education, a pure AI assistant is fine.
Privacy and data flow — your holdings should not train a model
A reasonable concern: if you connect your portfolio to an AI assistant, where does the data go?
The honest answer depends on the provider, and the question is worth asking explicitly.
The MCP transport. The AI client sends a tool request to the MCP server. The server returns structured data. The data flows back to the AI client to inform its answer. This conversation is, by default, in memory for the duration of the session.
Training. Reputable AI providers — Anthropic's API tier, OpenAI's API tier — do not train on customer data by default. Consumer tiers sometimes do. If you are connecting your real financial data, use the API tier or a product built on top of it, not the consumer-facing chat where training opt-out is sometimes only a setting buried in preferences.
Logging. AI requests are typically logged for safety, abuse-prevention, and product improvement. The logs do not normally include the data that came back from MCP tool calls, but they do include the prompts. "What is my Apple exposure?" is in the log; "your Apple exposure is 4.7%" usually is not. Check the provider's data retention policy.
The portfolio platform side. The MCP server is operated by the portfolio platform. Your holdings sit in their database under your authenticated account. The MCP layer is just a way for the AI to ask for them. Your trust model with the portfolio platform itself is the same as it would be without AI — you are deciding to let them hold your data, with whatever security and data-handling commitments they publish.
The general position: your portfolio data does not need to train a model for you to benefit from AI portfolio analysis. The architecture exists to keep them separate. Insist on that as a basic property of any AI portfolio tool you use.
Limitations of AI for investing
Worth being clear about what AI is not good at, even with MCP and full data access.
Forecasting. AI cannot tell you whether a stock will go up. Nothing can. Models that claim to predict prices are either trivially wrong or have data-snooping errors that disappear out of sample. AI is good at describing what is in your portfolio; it is not magic at telling you what to buy.
Hallucinations. Even with structured tool data, models occasionally produce confident-sounding wrong answers. The risk is reduced when the answer is grounded in real data, but it is not zero. Treat AI output as a draft to verify, not a final answer.
Regulation. AI assistants are not FCA-authorised to give personalised financial advice. They can describe your portfolio; they should not, and reputable products do not, tell you to buy or sell specific things. The legal line between "information" and "advice" is one the FCA polices and AI products operate carefully on the information side of.
Recency. Foundation models have training cutoffs. The model "knows" the world as of some date in the past. Real-time data has to come from MCP tools, not from the model's own knowledge. Ask "what is the current price of Apple?" and the model needs a tool, not its own memory.
Tax precision. UK tax rules — CGT, dividend allowance, ISA contribution rules, ERI on offshore funds, bed-and-breakfast rules — are detailed enough that AI summaries are worth verifying with a UK tax professional before acting on. AI is useful for "explain how bed-and-breakfasting works"; less useful for "tell me whether this specific disposal triggers it" without a human cross-check.
The shape of trustworthy AI portfolio analysis in 2026 is: AI handles description, summarisation, and routine data work; the human handles judgement, decisions, and the final tax-and-execution stretch.
Frequently asked questions
Does ChatGPT know my actual investment holdings?
Not by default. ChatGPT only knows what is in its training data and what you put in the prompt. It cannot reach into your broker accounts. Through MCP integrations, ChatGPT can connect to external data sources you authorise — including portfolio platforms — and read your real holdings on demand. Without that explicit connection, any portfolio question you ask is answered based on whatever you typed in.
What is MCP?
MCP is the Model Context Protocol, an open standard introduced by Anthropic in late 2024 for connecting AI assistants to external data sources and tools. It defines how a model can request data, what shape the response takes, and how user authorisation works. MCP is supported by Claude, ChatGPT and most major AI clients. In a portfolio context, it lets the AI read your holdings from a portfolio platform after you authorise the connection.
Is it safe to connect my portfolio to Claude?
Through a properly designed MCP integration, yes — provided the integration is read-only, the portfolio platform is reputable, and the AI provider does not train on your data. Read-only means the AI can see but cannot trade. Reputable platforms publish their data-handling commitments. Anthropic's API tier does not train on customer data by default. The combination is reasonable for serious financial data.
Can AI rebalance my portfolio?
It can suggest a rebalance — describing what your current allocation looks like, what a different allocation would look like, and the size of trades needed to get there. It should not place those trades on your behalf. Currently, the right architecture is AI on the analysis side, human on the execution side. The FCA's position on AI-driven execution for retail customers is unsettled, and language models still hallucinate often enough that automated trade execution is not a responsible product design in 2026.
Will AI replace my financial adviser?
No, not for most situations where you would actually use a financial adviser. AI is good at describing data, answering factual questions about your portfolio, and educating you about general investing concepts. It is not authorised to give personalised financial advice in the FCA sense, and it is not good at the parts of advice that depend on understanding your full circumstances — life goals, family situation, tax position, behavioural risk tolerance. AI is more usefully thought of as a layer that makes your interactions with your own data faster, not as a replacement for the human relationship some investors want.
What questions should I ask AI about my portfolio?
The questions where AI adds value are the ones that combine multiple data sources or require structured calculation. "What is my real exposure to Apple across all wrappers?" "Which of my funds overlap most heavily?" "How has my portfolio performed in GBP terms this year, with FX broken out?" "If I sold this position in my GIA today, what is the realised gain?" Less useful: "Should I buy this stock?" — AI cannot tell you. "What will the market do next year?" — nobody can.
Are AI portfolio insights regulated as financial advice?
Generally not, when the AI is summarising your own data and answering factual questions about it. The line the FCA polices is between "information" — describing facts about a portfolio — and "advice" — recommending a specific course of action with regard to a specific investment. Reputable AI portfolio tools stay carefully on the information side. If a product offers "personalised investment recommendations" generated by AI, check whether it is FCA-authorised; if not, treat the recommendations as information rather than advice.
Want AI that can actually see your portfolio?
Invormed exposes your holdings to Claude and other MCP-compatible AI assistants through seven read-only tools — portfolio summary, holdings, allocation, look-through, performance, dividends, and wrapper analysis. Your data does not train a model. The AI cannot trade. We are in early access — join the early-access waitlist and we will let you in as we open up.