If your picture of business AI is still a chat window where you paste in text and get a draft back, you are about a year behind. That picture was accurate in 2024. Since then, three changes have reshaped what the major AI tools can do for a small or mid-sized business: connectors that plug AI into the systems you already use, skills that package up how your team does a task, and agents that carry out multi-step work instead of just answering.
None of these is a gimmick, and none of them removes the need for a person to own the work. This post explains each shift in plain English, where it came from, which tools support it as of September 2026, and what it means for how you should set up AI in your business.
Key Takeaway
The short version: prompts became procedures. In 2024 the skill that mattered was writing a good prompt. In 2026 it is writing down how your business actually does something, clearly enough that an AI tool can follow it and a person can check the result.
Shift 1: Connectors, so AI can see your real work
The early limitation of chat tools was that they could only work with whatever you pasted in. Connectors fix that. They let an AI assistant read from, and sometimes write to, your email, calendar, file storage, CRM, or project tools, within permissions you set.
The plumbing that made this practical is the Model Context Protocol, or MCP, which Anthropic released as an open standard in November 2024. MCP gives AI tools a common way to talk to outside systems, so a connector built once can work across several AI products. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a new body under the Linux Foundation co-founded with Block and OpenAI and supported by Google, Microsoft, and AWS among others. At that point there were already more than 10,000 public MCP servers.
Source: Anthropic, Introducing the Model Context Protocol (November 2024)
Source: Anthropic, Donating MCP and establishing the Agentic AI Foundation (December 2025)
For a business, the practical meaning is simple: an assistant that can see your shared drive and your inbox can answer "what did we quote this client last spring" without anyone digging. The practical risk is equally simple: it can see your shared drive and your inbox. Connecting tools is a permissions decision first and a productivity decision second.
Shift 2: Skills, so AI does things your way
The second shift is the one most business owners have not heard of yet, and it is arguably the most useful. In October 2025 Anthropic introduced Agent Skills: folders of instructions, examples, and scripts that an AI tool loads only when a task calls for them. A skill might hold your proposal format and pricing rules, your brand voice guide, your month-end checklist, or the steps for turning a client intake form into a welcome package.
Think of a skill as a standard operating procedure written for an assistant. Instead of re-explaining your process in every chat, you write it once, and the AI pulls it in when it recognizes the task. Because only a short description is loaded until the skill is needed, a team can keep dozens of them without cluttering every conversation.
Source: Anthropic, Introducing Agent Skills (October 2025)
In December 2025 the format was published as an open standard. The project site lists support across a growing set of tools, including ChatGPT and OpenAI Codex, GitHub Copilot, VS Code, Cursor, and Google's Gemini CLI. That portability matters for small businesses: a procedure you write as a skill today is not locked into one vendor.
Source: Agent Skills open standard, agentskills.io
Source: GitHub Docs, About agent skills
Shift 3: Agents, so AI can finish the task
The third shift is agents: AI that plans and carries out a multi-step task, using tools along the way, rather than handing you a draft. OpenAI added an agent mode to ChatGPT in July 2025 that could browse websites and compile research reports on its own. Google launched Gemini Enterprise in October 2025 as a workbench for building and running agents across a company's data. In September 2026 Meta launched Muse, a personal agent that runs in its own secure virtual machine and asks a separate oversight agent to approve actions before it takes them online, which we cover in our guide to Meta's Muse.
Source: TechCrunch, OpenAI launches a general-purpose agent in ChatGPT (July 2025)
Source: Google Cloud, Introducing Gemini Enterprise (October 2025)
Agents are genuinely capable now, and they are also where most of the new risk sits. An agent that can send email, update records, or make purchases can make a mistake at scale. The sensible pattern for a business is the same one we use in every automation build: let the agent do the gathering and drafting, and keep a person at the point where money moves, a client hears from you, or a record changes permanently.
Business AI in 2024 A chat box you paste text into A new prompt every time One vendor, one product Output is a draft you copy somewhere
Business AI in late 2026 Connectors to email, files, and your CRM Skills that encode your procedures once An open standard that travels across tools Agents that complete steps, with a person approving
What happened to custom GPTs
If your business built custom GPTs in 2024 or 2025, this shift affects you directly. OpenAI is retiring custom GPTs in favour of plugins that bundle skills and connected apps. According to OpenAI's migration guidance, in affected Enterprise workspaces GPTs stop running on December 11, 2026, and new GPT creation there ended on September 25; on personal plans, new GPTs can no longer be created, though existing ones remain available for now. OpenAI's migration tool turns a GPT's instructions into a skill, but custom actions have to be rebuilt as connectors.
Source: OpenAI Help Center, Custom GPT retirement and migration FAQ
The good news is that the work you put into a custom GPT, the instructions and reference files that describe how your business does something, is exactly what a skill needs. It is a migration, not a rebuild from scratch, as long as someone on your team owns it before the deadline.
The model lineup, as of late September 2026
Models still matter, but they now change monthly, which is another reason to invest in skills and connectors that outlast any one model. As of September 26, 2026:
- Anthropic: Claude Opus 5.5, released September 22, which Anthropic says performs at the level of Claude Fable 5.1 on most work at 40 percent lower cost than Opus 5. It is available on the Pro, Max, Team, and Enterprise plans.
- OpenAI: the GPT-6 family, with GPT-6 Sol and GPT-6 Luna released September 22 and GPT-6 Astra earlier in the month. Luna is available to free users.
- Google: Gemini 3.1 Pro remains Google's top Pro model, with Gemini 3.8 Flash released September 2.
- Meta: Muse Spark 1.3, released September 2.
Source: Anthropic, Claude Opus 5.5 (September 2026)
Source: TechCrunch, OpenAI launches GPT-6 Sol and Luna (September 2026)
Source: Google AI for Developers, Gemini API changelog
For a side-by-side of the practical differences between the main assistants, see our ChatGPT, Claude, and Gemini comparison.
What this means for your business
Four practical steps follow from all of this, and none of them requires picking a winning vendor:
- Write down your procedures. The businesses getting the most from AI in 2026 are the ones whose processes exist on paper. A clear checklist for quoting, onboarding, or month-end is the raw material for a skill, and it helps new staff too.
- Treat connectors as access decisions. Before connecting an assistant to email or files, decide what it may read, what it may change, and who reviews that list. Put it in your written AI guidelines.
- Keep a person at the judgment point. Let agents gather, draft, and prepare. Keep approval with a person wherever money, clients, or permanent records are involved.
- Stay portable. Keep your skills, examples, and reference documents in your own files. With an open standard for skills and a shared protocol for connectors, switching tools is far easier than it was, as long as your know-how is not trapped inside one product.
If you want help deciding where to start, an AI readiness assessment maps your workflows and identifies which ones are worth turning into skills or agents first, and a hands-on workshop gets your team writing and using them safely.
Not sure whether your team needs connectors, skills, or agents first? Book a free 30-minute call and we will map one workflow with you and show you what each would look like.
Get Started