What AI Is Actually Good At: Four Business Tasks You Can Automate Today
Modern AI is astonishingly good at four things — summarizing, categorizing, ranking, and calculating. Here's how each becomes business automation.
Strip away the hype and the headlines, and modern AI is astonishingly good at exactly four things: summarizing, categorizing, ranking, and calculating. That's it. Those four jobs are the foundation underneath nearly every practical business use of AI I build, and once you see them clearly, the question stops being "what can AI do?" and becomes "where in my business is a human doing one of these four things by hand?"
Here's the useful reframe: forget the word "AI" for a minute. Each of these four tasks used to require a person to sit down, read something, think about it, and produce a judgment. That's slow, it's expensive, and it doesn't happen at 2 a.m. or across ten thousand records. What's genuinely new is that these four judgments can now happen automatically, at any scale — and the output is structured data your other systems can act on.
Forget the word "AI" for a minute. The real question is: where in your business is a human doing one of these four things by hand?
Eric Lovgren, lovgren.ai
Summarize
a record → the gist
Categorize
an item → a bucket
Rank
a list → a score
Calculate
a sentence → a number
1. Summarizing
Give the model a record — a long email thread, a support case, a contract, a meeting transcript, a customer's entire history — and it returns the gist in whatever length and shape you ask for. Three bullet points. A one-line status. A neutral paragraph a manager can skim.
This is the capability people underestimate, because summarizing sounds trivial until you count how often your team does it. A rep opening an account reads back through months of notes before a call. A support agent inheriting a ticket scrolls through forty messages to catch up. A manager assembles a weekly status by reading everyone's updates. Every one of those is a human compressing information by hand.
Now picture it happening on its own: the moment a support case is closed, a two-sentence resolution summary writes itself into a field. Before any call, the rep sees an auto-generated account brief. A long inbound email gets distilled to "what they're actually asking for" the second it lands. The reading still happens — a machine just does it, every time, in the same format.
2. Categorizing
Hand the model an item and a set of options, and it picks the right one. That covers assigning a category from a list (which department should this email go to?), and it covers the simplest, most powerful version of all: a yes/no, true/false, or yes/no/maybe verdict.
Classification is the workhorse. Inbound emails routed to the right team. Support cases tagged by issue type. Receipts sorted into expense categories. Customer messages scored as positive, negative, or neutral. A new lead's company description read and judged: does this business actually fit who we serve — yes or no?
A new lead comes in with nothing but a company name and a website. Read the site, and answer one question: "Is this an organization that needs the thing we sell — yes or no?" That single yes/no, written to a field the instant the lead is created, can decide whether it routes to a salesperson or quietly drops to a nurture list. No one had to read the website. No one had to decide.
3. Ranking
Give the model a set of criteria and a scale — say 1 to 10 — and it scores each item against the rubric. This is categorizing's more nuanced cousin: instead of a bucket, you get a graded judgment you can sort and threshold on.
Lead scoring is the obvious one — rate how well each lead fits your ideal customer on a 1-to-10 scale, then let anything above a 7 jump the queue. But the pattern is everywhere once you look. Triage a support backlog by urgency. Score open deals by health so the shaky ones surface before they slip. Rank applicants against a hiring rubric. Flag the invoices most likely to go unpaid. Anywhere a person is eyeballing a list and deciding "this one first," there's a ranking task waiting to be automated.
4. Calculating
This is the one that surprises people, because it isn't arithmetic in the spreadsheet sense — it's solving the "story problem." The hard part of a word problem was never the math; it was reading the messy situation and figuring out which math to do. That translation step — from a sentence a human wrote to the calculation it implies — is exactly what the model is good at.
Think about a quote request that arrives as a paragraph: "We've got about a dozen people who'd need access for the spring season, plus two managers." A person reads that, infers headcount, infers duration, applies your pricing rules, and produces a number. The model can do the reading-and-setting-up part instantly — extracting "12 users, 3 months, 2 admin seats" and applying the logic.
For exact, money-on-the-line arithmetic, the right pattern is to let AI interpret the situation and set up the numbers, then hand the actual computation to a calculator or a formula. The model's strength is turning a messy sentence into structured inputs — not being a calculator itself. Build it that way and you get the best of both: human-like reading, machine-precise math.
The pattern underneath all four
Notice what these have in common. Each one used to require a human to intervene — to read, judge, and type the answer in. Now each can run on its own, produce a piece of data, and that data can trigger the next step automatically. Unstructured in, structured out, automation fires.
| By hand (before) | Automated (now) | |
|---|---|---|
| Summarizing | Read it all yourself | Gist written to a field |
| Categorizing | Decide the bucket by hand | Verdict assigned instantly |
| Ranking | Eyeball the list | Scored 1–10 automatically |
| Calculating | Work the numbers out | Inputs extracted, math applied |
- Generate the judgment. AI reads the record and produces a summary, a category, a score, or a number.
- Store it as data. The answer lands in a field — a verdict, a rating, a dollar figure — right where your systems expect it.
- Trigger the automation. That field drives what happens next: routing, assignment, an alert, an email, the next stage.
For anyone running a business on Salesforce or a similar system, that third step is the real prize. The judgments AI produces aren't just nice to read — they're values in fields, and fields drive everything. What if every new lead were automatically assigned an Industry the moment it arrived, instead of sitting blank until someone got to it? What if it were dropped into the right Territory based on a quick read of the company? What if a company description were analyzed to stamp a simple yes/no on whether it meets your criteria — and the "yes" leads routed straight to a rep while the rest waited?
None of those require a single human to read a single record. The data generates itself, and the automations you already know how to build take it from there.
The exercise worth doing
The specific answers depend entirely on your business — and I'll share concrete use-cases in the Insights and Solutions sections of this site. But the thinking is the same everywhere, and it comes down to two questions:
What are you doing manually today that could instead happen automatically? And what aren't you doing at all today — because it would simply take too much time — that suddenly becomes possible?
That second question is where most of the value hides. There's a whole category of work that's obviously valuable and obviously never gets done, because no one has the hours: reading every inbound lead's website, summarizing every closed case, scoring every deal in the pipeline every week. When the cost of a judgment drops to near zero, the work that was never worth a human's time becomes worth doing on everything.
If you can frame what you want in terms of summarizing, categorizing, ranking, or calculating, it's very likely something I can automate. That's usually the most useful first conversation to have.