AI tool or AI workflow: the difference that matters
When two founders say they 'use AI in the business', one may be pasting into ChatGPT and the other may have it wired into their pipeline. Same technology, very different dependency.
AI tool or AI workflow: the difference that matters
Two founders. Same industry. Both say they "use AI in the business."
The first has ChatGPT open on a second screen. Customer sends a tricky email; she pastes it in, gets a draft back, copies it, tweaks it, sends. Speeds up her day. Nothing else changes.
The second has AI embedded in the pipeline. Customer sends an email; it gets classified, routed to the right person, drafted with the relevant context pulled from the CRM, and queued for one-click approval. All before anyone reads it.
Same technology. Two very different businesses.
Tool users have a helper. Workflow users have a dependency.
If the AI vendor doubles the price tomorrow, the tool user notices, weighs it up, might switch to Claude. The workflow user has a broken system. The pipeline stops. Emails pile up. The team waits: they didn't write the pipeline and can't fix it.
That's the trade-off nobody talks about. Embedded AI delivers real leverage. It also creates real dependency. You cannot have one without the other.
Which do you have? Look for these signals.
Signs of AI as a tool. Your team opens a chat window, prompts, copies output somewhere else. Removing the AI slows them down but nothing breaks. There's no automation calling the model on your behalf. Switching to a different provider is a Chrome tab change.
Signs of AI as a workflow. Something in your business calls the API without a human in the loop. Data flows into it and structured output flows out. The output is used downstream — populating a record, triggering a task, sending a message. Switching providers means rewriting or re-testing a system.
Most businesses have both. That's fine. What matters is knowing which is which.
Why this matters now
When we ask founders "how are you using AI?", they list tools. ChatGPT. Copilot. Perplexity. Fine, those are tools. Then somewhere in the same estate there's an automation nobody wrote a runbook for, quietly moving data through OpenAI's API on a schedule. That's the load-bearing workflow. That's the one you need to reason about.
The interesting risks live in the second column. Which vendor. Which model. What happens if the API returns garbage. What happens if it stops returning anything. Who owns the fix.
None of these questions come up when the AI is a helper. All of them come up when it's a workflow.
The action
Draw two columns. Left: tools your team uses on demand. Right: workflows that call AI without a person watching. Write down what's in each.
If the right column is empty, you're not getting the leverage yet, but you have no dependency either. If the right column has anything in it, you have work to do: mapping the vendor risk, the fallback, the human failsafe.
Both columns are legitimate. Confusion between them is not.
Want to know which column your AI use falls into? Book a conversation and we'll walk through your stack.
Robin Carswell
More on
The AI features you didn't sign up for
Every SaaS tool you pay for is quietly shipping AI features by default. You didn't procure them, you can't cleanly turn them off, and your team is now building workflows around them.
'AI-first' is a paint job, not an architecture
Most tools claiming to be AI-native haven't changed their architecture at all. Here's the test that separates genuine AI-native systems from legacy software with a chatbot bolted on.
The three workflows every business should automate first
Most automation projects fail because the starting point is wrong. Here are the three workflows that almost always justify the effort — and how to pick which one to tackle first.