AI vs automation: build with AI, run without it
Use AI once to write the rules, scripts and dashboards, run them for almost nothing, and pay per case only where the work needs it.
Many companies now pay for AI every time a task runs, even when a plain automation would do. Each email answered, report drafted or ticket tagged is another model call on the bill. The bill grows with volume.
The cheaper way is to use AI once, to write the thing that does the work, and then let that thing run without it.
Use AI to build
AI is very good at writing the parts of a process. Give it a clear description and a few examples, and it will draft:
- the formula that reconciles two reports
- the rules that route each request to the right team
- the script or workflow that moves orders between systems
- the query behind a monthly dashboard
A person who knows the process tests each piece on real cases before it goes live.
Then run without it
Once built, a rule or a script works on structured data such as order numbers, dates, amounts and status codes. It needs no model for that. Each run costs close to nothing, whether it runs 100 times a month or 100,000. It gives the same answer to the same input every time, and an auditor can read why.
The run cost is hosting, a workflow tool or a spreadsheet you already pay for. We state that monthly number before the build starts.
Returns point the same way. Fortune’s coverage of the MIT NANDA report notes that more than half of generative AI budgets go to sales and marketing tools, while MIT found the biggest ROI in back-office automation. The back office is where most repeatable, rule-based work lives.
Three examples
All three are hypothetical. Your numbers will differ.
Order-status emails. Say a distributor gets 2,000 emails a month asking where an order is. Most include an order or invoice number. A script finds the number, looks up the status in the ERP and replies with a template giving the delivery date and tracking link. Emails with no number, or with a complaint, go to a person. AI wrote the script and the templates. Each answer costs close to nothing.
The monthly sales report. Say a finance analyst spends two days a month downloading exports, pasting them into a spreadsheet and rebuilding the same ten charts. A script pulls the same exports and fills the same template in minutes. AI helped write the formulas and the script. At month end nothing calls a model. The analyst spends those two days explaining what changed.
Ticket tagging at high volume. Say a support team gets 30,000 tickets a month in free text, with typos and screenshots, in Spanish and English. Keyword rules tag some and miss many, because reading them takes judgment. Here a model call on every ticket can pay for itself. Say it costs $0.40 MXN per ticket, review included. That is $12,000 MXN a month. Tagging by hand at one minute per ticket would take 500 hours a month, so the model costs $24 MXN for each hour of work it replaces. A wrong tag costs little to fix. AI stays in the loop here.
When AI should stay in the loop
Pay for a model on every case when three things are true:
- The input is messy. Free text, scanned documents, images or voice that no form captures.
- The call needs judgment. Cases differ, and nobody can write the rule down.
- The volume is there. Enough cases for the savings to cover the build, the cost per case and the checking.
Then check what a mistake costs. If errors are expensive or hard to catch, a person reviews the output, and that time belongs in the cost per case.
Our default order starts with a rule or a spreadsheet, then a script or workflow, then an off-the-shelf AI tool. Custom AI comes last.
Price the case before you switch it on
A cost per case looks small in a pilot. At full volume it becomes a monthly line that grows with the business, and once a process depends on it, it is hard to switch off.
Cost per case is model fees, plus review time, plus the cost of fixing mistakes. Put it next to what each case is worth and what the simple version costs. Multiply by monthly volume to see the bill. If the simple version handles most cases, build it first and send the rest to a person or a model.
Write the number down before go-live. Check it again after the first month at real volume, and again whenever model prices change.
A short checklist
Before you pay for AI on every use, ask:
- Can a rule, a formula or a lookup handle most cases?
- Is the input structured, or could a form make it structured?
- If you need AI, can it write a script once instead of running on every case?
- Do you know the cost per case at full volume?
- Do you know what a mistake costs and who catches it?
- Is there a date to check the number again?
Score one task of your own with the free “AI or not?” tool.
To find which of your AI lines could run without a model, book a diagnostic.