The most common version of this question is "how do I train ChatGPT on my business data". It is worth correcting the premise first, because it sends people down an expensive road for no benefit.
You don't train it — you give it context
Training changes a model's weights. It is slow, costly, requires substantial data, and produces something that is out of date the moment your prices change. Almost nobody asking this question wants that.
What they want is for the assistant to have their information available while answering. That is a context problem, and it is solved by supplying material at the time of the question rather than by baking it into a model. Everything below is a way of supplying it.
The seven levels, at a glance
| Level | Mechanism | Best for | Freshness | Maintenance | Can it act? | Portable across assistants? |
|---|---|---|---|---|---|---|
| 1 | Paste into the chat | One-off questions | Correct at the moment you paste | Re-paste every time | No | No |
| 2 | Custom instructions / projects | Stable rules and identity | Until you edit it | Occasional review | No | No |
| 3 | Uploaded files | Written, bounded knowledge | As of the file's last edit | Manual re-upload | No | No |
| 4 | Knowledge base with retrieval | Large document sets | As fresh as the index | Ongoing curation | No | Only if you build it that way |
| 5 | Connected live systems | Numbers that change | As current as the last sync | Mostly automatic | Depends on the connection | No, per integration |
| 6 | Governed tool access (MCP) | Multiple assistants, controlled actions | Set by the server | Server-side | Yes, within what is exposed | Yes |
| 7 | Persistent context layer | Context spanning systems and sessions | Maintained continuously | Automatic, with human curation | Yes, typically gated by approval | Yes |
Escalating is a cost, not an achievement. The right level is the lowest one that answers the questions you actually ask.
Level 1 — Paste it in
What it is. You put the relevant facts directly into the conversation.
Best for. One-off analysis. A specific report. Anything you will not ask again.
Limitations. It dies with the conversation, it is a snapshot from the moment you pasted, and there is a practical ceiling on how much you can supply before the answer quality drops. More is not free: as the window fills, accuracy and recall degrade — an effect Anthropic calls context rot.
Verdict. Underrated for genuine one-offs. A trap as a habit, because the effort recurs forever and nothing accumulates.
Level 2 — Custom instructions and projects
What it is. A standing block of text applied to every conversation, or to every conversation inside a project or workspace: who you are, what you sell, who your customers are, how you want answers formatted.
Best for. Identity, positioning, standing constraints, tone. The context that is true every day and rarely changes.
Limitations. Small. Static until you edit it. Confined to one product — Claude projects do nothing for ChatGPT, and vice versa. And because it is always present, it consumes context on every question whether relevant or not, so it should be short.
Verdict. The highest-return thirty minutes most businesses can spend on AI. Write it once, review it quarterly. Do this before anything more elaborate. Packaging levels 2 and 3 together as a shareable assistant is what a Custom GPT does — worth reading before you build anything larger.
Level 3 — Uploaded files
What it is. Documents attached to a conversation, project, or assistant: price lists, brand guidelines, past reports, policies.
Best for. Knowledge that is genuinely written down, bounded in size, and changes on a human timescale.
Limitations. A file is a photograph. It is accurate as of its last edit and gives no signal when it drifts, which is the most common way AI answers go quietly wrong. Uploads also stay inside the product you uploaded them to.
Verdict. Good for stable documents. Dangerous for anything with numbers in it, unless you commit to a re-upload rhythm you will actually keep.
Level 4 — A knowledge base with retrieval
What it is. Documents indexed so the system can search them per question and pull in only the relevant passages. This is retrieval-augmented generation: search the corpus at question time, place the passages in context, then answer.
Best for. Large bodies of written material where any given question touches a small part — support content, documentation, policy libraries, past research.
Limitations. Retrieval quality is now a thing you own; a relevant passage that ranks poorly is invisible to the answer. It still only knows what somebody wrote down, and it still ages at the speed of the underlying documents.
Verdict. The right answer for document-heavy businesses. Not an answer for operational questions, because operational facts are not documents. Where retrieval runs out goes through this in detail.
Level 5 — Connected live systems
What it is. A direct connection to the systems that own your changing numbers: commerce, analytics, search, advertising, payments, email.
Best for. Anything where the correct answer changes: revenue, spend, stock, rankings, open orders, campaign status.
Limitations. Each connection is its own setup, its own permissions, and its own maintenance. Connecting a system to one assistant does nothing for the next one. Freshness now depends on sync behaviour, which you need to understand rather than assume — "connected" does not automatically mean "live to the second".
Verdict. The step that changes what AI is useful for. Also the step where people start rebuilding the same integration per tool, which is what level 6 exists to stop.
Level 6 — Governed tool access (MCP)
What it is. Exposing systems through a standard interface an AI assistant can discover and call. The Model Context Protocol is an open standard for exactly this: a server advertises named tools with defined inputs, and an authorized client calls them.
Best for. More than one assistant, more than one person, or any situation where an AI should be able to do something and you need control over what.
Limitations. Setup is real. The server's design decides everything that matters — what proportion of tools can change things, whether a human approves, how narrowly access is scoped. A badly designed server is a liability regardless of the protocol.
Verdict. The point at which context stops being per-tool plumbing and becomes infrastructure. MCP for business covers how to evaluate a server before connecting it; Claude and ChatGPT each have their own connection path.
Level 7 — A persistent context layer
What it is. Business context maintained outside every assistant: assembled from connected systems, structured into entities and relationships, refreshed on a schedule, scoped by permission, and exposed to whichever AI client is authorized.
Best for. Context that spans several systems, has to stay current without anyone maintaining it by hand, and needs to serve more than one assistant and more than one person.
Limitations. It is architecture. It only pays off when the questions genuinely span systems and recur. A business whose context really is three documents will get no return on it, and should stop at level 3.
Verdict. Worth it when levels 2 to 5 have each stopped scaling — usually visible as the same numbers being re-pasted weekly by several people into several tools.
How to choose, in four questions
1. Does the answer depend on a number that changes? No → levels 2 and 3 will take you a long way. Yes → you need a connection. Nothing below level 5 solves this.
2. Is the written material bigger than you can supply at once? Yes → level 4. No → level 3 is simpler and more predictable.
3. Does more than one assistant, or more than one person, need it? Yes → level 6 or 7, because everything below is trapped in one product.
4. Should AI be able to do things, or only tell you things? Only tell → stop at 5. Do → level 6 at minimum, and evaluate the approval model harder than the feature list.
A sensible starting sequence
For a business that has done none of this:
- Week one. Write the level 2 block. Three or four paragraphs: what the business is, who it sells to, what matters, how you want answers. Put it in whichever assistant you use most.
- Week two. Upload the three documents you would hand a new employee. Not thirty. Three.
- Then stop and observe. For a fortnight, note every question the assistant got wrong. Sort the failures into "did not know a written fact" and "did not know a current number".
- Escalate where the failures are. Written-fact failures point to level 3 or 4. Current-number failures point to level 5 and above. Most businesses discover their failures are lopsided, which makes the next investment obvious.
That observation step is what stops people building level 7 architecture for a level 3 problem.
Where Zylx fits
Zylx implements levels 5 through 7 as one thing. A Business Brain is persistent, structured business context — assembled from the systems a company already runs and kept current — that authorized AI assistants can query through a governed interface. Supported sources include Shopify, Stripe, Google Search Console, Google Analytics 4, Google Ads, Microsoft Clarity, Klaviyo, Ahrefs, Semrush, GitHub, and a crawl of the company's own website. Assistants reach it over MCP with reads by default; changes become proposals a human approves.
It is not the right tool for a business whose context is a handful of documents. Levels 2 and 3 are free, and if they answer your questions, the honest recommendation is to stay there.
Sources
- Anthropic: Effective context engineering for AI agents — what makes up an agent's context.
- Anthropic: Context windows — window behaviour and context rot.
- Anthropic: third-party connectors with remote MCP — how Claude connects to outside systems.
- OpenAI: building MCP servers — transport and authentication expectations for ChatGPT connectors.
- Model Context Protocol documentation.