AI that fits the job: the right tool for the risk
Three made-up stories: a phone assistant, a paperwork helper and a privacy-first setup. Not client projects.
4 min readMade-up stories, not client work
Story 1 of 3
A friendly assistant for the phone and website
The short version: Let the assistant handle the routine calls, and hand people the conversations that matter.
A home services company misses calls after hours and wants a phone and website assistant to answer questions, schedule visits and take messages.
What's going wrong
People ask questions every which way, and phone menus drive them mad. But a wrong price would hurt trust.
What we'd do
- Pick a strong online AINatural conversation matters most here, so a leading online AI that keeps improving fits.
- Give it the factsAnswers come from the company's own prices, hours and policies; bookings go through its real calendar.
- Set firm limitsIt declines off-topic requests and never promises refunds or deals.
- Hand over to a personComplaints, emergencies and tricky questions go to staff, or become a callback with a summary.
- Keep an eye on itWeekly conversation reviews, and new AI versions tested on saved conversations before switching.
How we'd know it worked
- Anything off-topic reaches a person
- A new AI version must pass the saved tests first
- The catch
- Online AI changes over time and charges by use. Locking the version and re-testing keeps changes on your terms.
- What the business gets
- Fewer missed calls, and staff free for conversations that need a human.
Story 2 of 3
Taking the typing out of paperwork
The short version: Let AI do the typing, and let people check the exceptions.
A distributor gets purchase orders by email and PDF in every format imaginable, and retyping them causes wrong shipments.
What's going wrong
Reading messy documents is a great job for AI. But prices, quantities and account details must be exactly right.
What we'd do
- Split the job in twoAI reads each document and pulls out the details; exact rules then check them against the real catalog, prices and accounts.
- Double-check the formatEverything must be in the expected shape, and totals must add up, before an order is created.
- Know when it's sureOrders that pass every check go straight through. Anything uncertain is flagged.
- A person checks the restFlagged orders appear beside the original document, so staff confirm or fix them in seconds.
How we'd know it worked
- No order is created without passing the checks
- Clear limits on what can skip a person
- The catch
- Some manual checking stays, shrinking as the system proves itself.
- What the business gets
- Less retyping, fewer wrong shipments, staff focused on real exceptions.
Story 3 of 3
AI for sensitive, must-be-consistent work
The short version: Keep sensitive data at home, and change the AI only when you decide to.
A financial firm wants AI to summarize client files and flag missing documents. Client data can't leave its own systems, and results must stay comparable over time.
What's going wrong
A public AI service breaks the firm's privacy rules, and surprise updates would make results inconsistent.
What we'd do
- Compare the optionsA private AI in the firm's own cloud, one on its own equipment, or an online service with strict data terms, each tried on sample files.
- Freeze the versionThe chosen AI and its settings stay fixed, changing only after testing and sign-off.
- Be honest about consistencyA private or frozen AI changes less, but still won't give word-for-word identical answers. Checks and human sign-off stay.
How we'd know it worked
- Client data stays in approved systems
- Results on a fixed set of test files stay within agreed limits
- Every change has a sign-off and a way back
- The catch
- Private AI gives more control but needs upkeep, may lag the newest online abilities, and updates slowly on purpose.
- What the business gets
- Useful AI that respects the firm's privacy rules, with changes it controls.
What all 3 stories have in common
AI changes fast. Your business needs are specific.
The right answer sits between the two, and no single AI suits everyone, so we ask before we pick.
- 1
What are you trying to do?
Your goals, and how the work gets done today.
- 2
How much wobble is OK?
We walk through real situations to see how much variation the work can truly handle.
- 3
What happens if it's wrong?
An awkward reply, a wrong invoice and a privacy leak need very different safety nets.
- 4
Where can the data go?
Who may see it, and what must stay inside the business.
- 5
Is the newest model worth it?
Whether new abilities justify the risk of switching.
- 6
What will it cost to live with?
Running costs, upkeep and future flexibility.
The goal isn't the newest AI. It's AI your business can trust, afford and change on its own terms.
How we run every AI project
- Try several AI models on real tasks
- Write down the choice, its limits and what counts as good enough
- Test reliability and consistency before going live
- Watch for problems and surprises after launch
- Keep versions, a way back and human oversight
- Leave room to switch to better AI later
- Agree who is responsible and how changes are made
If you liked this, read
About these stories. These stories are made up to show how we think. They are not client projects, describe no real business, and contain no measured results.
Next: IT and security that make sense, and that you can run yourself
Three made-up stories about backups, security and getting free of IT vendors. Not client projects.
Read the next story