If your business doesn't ship software, the phrase "AI transformation" is being weaponised against you. Most of what's sold under that label is theatre, a deck, a license, a workshop, and an invoice. The deployments that pay back look nothing like the deployments the vendors are selling.
This is the pillar piece. The rest of the AI category links back here.
The mental model
A useful AI deployment in a non-software business sits in one of four places:
- Inside an existing workflow: between two human steps that should never have needed a human in between.
- Across a data silo: connecting what marketing knows to what operations knows, or what sales saw to what finance recorded.
- In the customer interface: the place a customer asks the question they currently ask a human.
- In the operator's review loop: Friday afternoon, the weekly metrics, the anomalies.
If a proposed deployment doesn't sit in one of those four, it's probably a science fair project, not a transformation.
What actually pays back
In the field, year-one ROI clusters around five patterns:
- Demand forecasting that beats the spreadsheet. Retail, hospitality, logistics. Not because the model is magic, because the spreadsheet was using last year's number plus 8%.
- Field-ops triage. Field-service businesses where the first hour of a job is dispatch + diagnosis. AI tightens both.
- Customer-service deflection that doesn't piss off the customer. Narrow, scripted, well-instrumented. Not a chatbot pretending to be a person.
- Document workflows. Insurance, legal, finance, ops. Anywhere a human is reading PDFs all day.
- Anomaly detection in the metrics that matter. Inventory variance, payment fraud, transit times. Pattern: replace the dashboard that nobody reads with the alert that ships when the pattern breaks.
What doesn't pay back
Same field, same year:
- "AI strategy" engagements with no production system at the end.
- Custom LLM training when fine-tuning a strong base model would have done the job in a tenth of the time.
- Internal "AI assistants" rolled out to everyone without a single owned workflow. Adoption hits 4% and stays there.
- Chatbots on the marketing site that didn't need to exist.
- "AI strategy" with no production system
- Custom LLM training
- Assistants for everyone, no owned workflow
- Chatbots the marketing site didn't need
- Demand forecasting that beats the spreadsheet
- Field-ops triage
- Narrow, scripted service deflection
- Document workflows
- Anomaly alerts on the metrics that matter
The diagnostic to run before you spend a pound
Three questions, asked of every proposed AI project:
- What is the human action this replaces or augments? If you can't say in one sentence, walk away.
- Where does the data live, and who currently looks at it? If neither answer is clean, fix that before the AI project.
- How will we measure pay-back in 90 days? If the answer involves "we'll see," it isn't a project, it's a hobby.
Get those three answers and you've already filtered 80% of the noise the vendors will throw at you.
What an AI-native partner looks like
Three tells:
- They name the workflow before they name the model. The first slide is your process, not their tech stack.
- They quote a 90-day pay-back, not a 36-month roadmap. The roadmap can exist behind the 90-day win.
- They tell you which deployment to not do. If everything on your wishlist gets a thumbs-up, you're talking to a vendor, not a partner.
This is the work I do at Rope, and it's the lens I write from on this site. If your business runs on physical things, vehicles, inventory, locations, people, and you've been told "add AI" without a specific workflow attached, the consulting page lists the five ways to start a conversation.