Retavon Retavon

Blog

Designed for AI, not AI included

People expect their assistant to know them. That context lives with the assistant. Build software their assistant can work.

Michael Cooper,

For a few years the way to put AI in business software was to put an assistant in it: a chat panel in the corner, trained on the product's help pages, able to draft an email or summarize a screen. "AI included" became a line on the pricing page.

That made sense while people had no assistant of their own. Most people on a dealer's team now use ChatGPT or Claude, and the panel in the corner knows nothing about them.

Most people at work use one assistant

LayerX's State of AI Usage Report 2026, on how people at work use AI tools, found that 70.44% of users rely on a single AI assistant. 21.16% use two, and about 8% use three or more.1

In Claude, a person adds a business application under Customize, then Connectors, by choosing Add custom connector, pasting the address of the application's MCP server and signing in as themselves.2 ChatGPT connects a workspace to outside applications through MCP servers in the same way,3 and OpenAI is retiring custom GPTs on December 11, 2026, with custom actions not transferring automatically.4

Context lives with the assistant

People expect their assistant to know them. It can hold their past conversations, their email and documents, and how they like things done. Asked to follow up with a customer, it can know that the customer wrote last Tuesday and that this salesperson writes short emails.

An assistant bundled into business software holds the product's records and nothing about the person typing: not the email thread, not the spreadsheet the owner keeps on the side, not the way this salesperson lays out a quote.

The obvious fix is for the software to collect that context: read the person's mail, keep its own memory of them, learn their habits. We think business software cannot do that well and should not try. That context belongs to the person. It spans every application they use, not just ours, and gathering it into a point of sale would mean holding far more about each employee than a store's software has any business holding.

The person's assistant brings the context about the person. The application brings the context about the business: the live records, what this person is allowed to do, and what can happen next to each record.

A bundled assistant usually runs under one account for the whole business, so the audit trail shows one name whoever asked. And it is another bill, for people who already pay for ChatGPT or Claude.

What "designed for AI" means in practice

In Retavon, business software for specialty retailers, designing for AI comes down to five things.

  1. The assistant acts as its person. Each person issues a key for their own assistant. It carries exactly that person's roles, at the stores they work, and can be made read-only or set to expire. A salesperson's Claude can write a quote and cannot give a refund the salesperson could not.
  2. Every record says what can happen next. A record with a status carries a next list: the operations its state allows, each with the reason. The assistant does not have to infer the workflow from the documentation. The buttons on the screen come from the same list.
  3. Refusals say what to do. When the system says no, it says why, with the numbers, and what would work instead: "Elena Vasquez has $120.00 of account credit: $200.00 cannot be paid with it. Pay at most $120.00 on account and the rest another way." The response gives the assistant the amount and an alternative, so its next call can succeed. The MCP specification asks for the same: errors with "actionable feedback that language models can use to self-correct and retry".5
  4. Money moves once. Every write that takes money or moves stock takes an idempotency key, so an assistant that retries after a dropped connection does not take a deposit twice, the way Stripe's API handles retries.6
  5. It is all on the record. The audit trail marks an assistant's work as made by its key, beside the person's own, with what each record was before.

Where we learned this

Most of this came from building AGLedger, change control for AI agents. Working with agents that do real work inside systems of record taught us to start with identity, permission, what the agent may do next, and an account of what it did. Retavon starts from those.

Sources

  1. LayerX, State of AI Usage Report 2026: layerxsecurity.com
  2. Claude, custom connectors using remote MCP: claude.com/docs/connectors/custom/remote-mcp
  3. OpenAI, MCP servers in ChatGPT: developers.openai.com/api/docs/mcp
  4. OpenAI, custom GPT retirement and migration FAQ: help.openai.com/en/articles/20001519
  5. Model Context Protocol specification, tools and error handling: modelcontextprotocol.io
  6. Stripe, idempotent requests: docs.stripe.com/api/idempotent_requests