Perspective

AI is not one thing.

If a vendor tells you AI means agents, they are describing their product line, not your options. Here is the full board.

01

Every seller describes AI as the thing they sell.

Agent platforms say AI means agents. Automation shops say it means workflows. Your software vendors say it is already in the product, just turn it on. None of them are wrong about their square. They are selling you the square.

This matters because AI advice is hard to judge before you buy it, and often after. The simplest protection is a map of the whole board, so you can notice when someone presents one region as the entire territory. We hold ourselves to the same test.

02

The map.

Two questions place any AI initiative. How do you get it: already in your stack, configured, or built? And who is in control: does a person drive every step, supervise the work, or does it run on its own?

WHO'S IN CONTROLRuns itselfYou superviseYou drive itAssistantsEveryday knowledge workCoworkersDelegate, then reviewEmbedded AIInside software you ownAutomationsAI steps in workflowsSpecialised agentsScoped to one functionSystems and productsWhere advantage livesAlready in your stackYou configure itYou build itHOW YOU GET IT

Specialised agents are one region of six. A vendor who lives there will describe the whole map as that region.

03

The six forms.

  1. 01

    Assistants

    A person thinking, drafting and analysing with AI in the loop on every step. Bought, not built. The foundation of workforce fluency and usually the fastest payback in the business.

  2. 02

    Coworkers

    Agentic tools that take a whole task, work on it for minutes or hours, and hand back a result for review. Coding agents and computer-use agents sit here. Still bought, lightly configured.

  3. 03

    Embedded AI

    The AI already inside software you pay for: your CRM, your accounting package, your office suite. Nobody builds anything. The decision is whether to turn it on and how far to trust it. Usually the cheapest win in the stack, and the most overlooked.

  4. 04

    Automations

    Lightweight workflows with AI doing judgment at specific steps: classify, extract, summarise, route. Deterministic pipes, AI at the joints. They run until an exception needs a person.

  5. 05

    Specialised agents

    Software agents scoped to a business function: inbox triage, invoice matching, content production. Configured on a platform or bought as a vertical product. Useful, and one square of six.

  6. 06

    Systems and products

    AI at the core of something custom: an internal platform, a reimagined process, a client-facing product. The most expensive form, the slowest to ship, and the one that builds durable advantage. Also where governance matters most.

04

What this means for strategy.

An AI strategy is not a decision about which agent to buy. It is a deliberate mix across all six forms, sequenced. Most mid-market businesses should have assistants, coworkers and embedded AI working well before specialised agents or custom builds make sense. Autonomy is earned over time, not declared at setup.

The edges of these categories blur. A coworker given a schedule becomes an automation. An automation with enough scope becomes an agent. That is fine. The map is not a classification exam. It is a tool for noticing when someone sells you one square as the whole board.

The right question is never "should we get an agent". It is "which form fits this problem, and are we ready for it". Our 90-day AI Strategy Jump Start answers that question system by system: integrate, extend, replace, or leave alone.

Start with a map of your own stack.

The Jump Start puts every meaningful system on this board in 90 days.

Learn about the Jump Start

By Paul at delvr.ai. Published July 2026.