AI Is a Tool, Not a Strategy:
Where It Fits and Where It Doesn't

There's a meeting happening right now in a mid-market company near you where someone is presenting an "AI strategy" deck. The slides have words like augmentation and enablement and transformation. There are arrows. There are quadrants. Somewhere in the deck there's a bullet that says "leverage generative AI to unlock value." Nobody in the room is sure what to do on Monday morning.

I want to say something unpopular about AI in business. It's a tool, not a strategy.

You don't have a Microsoft Excel strategy. You have a finance team that uses Excel for the things Excel is good for, and uses other tools for the things Excel isn't good for. The same posture is the right one for AI. It's a category of tool. The interesting question isn't "what's our AI strategy?" The interesting question is "where in our actual work does this category of tool make us better, and where doesn't it?"

That second question is the one this post is about.

Where AI fits well

A few honest fits, based on what I've actually used AI for in real businesses.

• Drafting and summarizing. AI is genuinely good at the first version of an email, a memo, a job posting, a vendor outreach, a meeting summary. It also handles longer-form summarization well: condensing a 40-page contract into the parts that need attention, turning a year of customer feedback into the recurring themes, distilling a meeting transcript into action items. The time savings compound across a team that does a lot of correspondence and reading.
• Repetitive document work. Pulling structured data out of PDFs. Reformatting a stack of vendor invoices into a consistent shape. Categorizing receipts. Extracting key dates out of contracts. Anything that used to involve a person reading the same kind of document over and over and typing the same kind of field into the same kind of form.
• Processing large piles of data. Drop in a spreadsheet, a CSV export from your accounting system, or an ODBC pull from your ERP, and AI can answer the kind of "I just want to know X about this data" questions that used to need either someone who knew the right query tool or hours of manual sorting. Finding duplicates. Spotting outliers. Grouping records into themes. Pulling out the rows that match a pattern you can describe in plain English. This is genuinely useful work that didn't have a good answer before.
• Search and Q&A across your own documents. This is the capability people are usually referring to when they say "RAG." Connect AI to your own policies, procedures, contracts, manuals, and email archives, and it can answer questions from inside your business instead of from generic web knowledge. "What did we decide about that vendor last quarter?" "Where in the manual does it talk about reimbursable mileage?" "What's our standard COI requirement for subs?" AI plus your own documents beats Ctrl-F for the kinds of questions an employee would otherwise ask a coworker.
• Multi-step workflows that used to require a person. A newer capability, sometimes called "agents." Instead of asking AI to do one thing, you set it up to do a sequence: read an incoming vendor email, look up the vendor in your master file, draft a reply, queue it for your approval, file the email when you confirm. For routine work where the right answer is usually clear, this is a different category of capability than "help me write this one email." The catch is the same as elsewhere: there needs to be a human checkpoint somewhere, and the chain needs to be observable when it goes wrong.
• Classification and routing at scale. Sorting customer messages by topic. Routing AP invoices to the right cost code. Flagging contract clauses that don't match your standard terms. Categorizing expenses against the chart of accounts. AI is good at applying consistent judgment to a high volume of small decisions, where the criteria are stable but the volume makes it expensive to do by hand.
• A second set of eyes on the work. Reviewing a contract for unusual terms before you sign. Reading a financial statement and asking "what assumptions does this imply?" Walking through a forecast and flagging the numbers that depend on something you haven't actually verified. AI isn't a substitute for the expert, but it's a useful first reader for documents an expert would otherwise read cold.
• Software, where you have someone who can verify it. AI can produce real working software at a speed traditional development can't match. The catch is the word "verify." If you don't have someone on your side who can read what came out and tell good from bad, the speed is a liability, not an asset. With verification in the loop, the cost curve is different from traditional development.
• Translation between human and software languages. Turning a written description of a workflow into a working procedure. Turning a spreadsheet of inconsistent data into a clean import. Turning a vague request from a stakeholder into a clear specification.

Where AI doesn't fit

The same honest accounting on the other side.

• Anything where a small error embarrasses you. A misstated number in a board deck. A wrong figure in a customer-facing report. A confident hallucination in a contract clause. AI will produce text that looks completely correct and is wrong in a way you'll only catch on careful reading. If the cost of a small mistake is high, the right place for AI is upstream of a human review, not as the final output.
• Anything that requires institutional context to verify. AI doesn't know that this customer always gets net-45 terms, that this vendor's invoices are habitually short by the freight line, or that your CFO hates being CC'd on anything routine. Those facts live in your team's heads. Without them, AI fills in the gaps with plausible-sounding wrong answers.
• Anything where the consequence of an error compounds silently. A miskeyed AP entry catches itself at the bank reconciliation. A miskeyed AI-driven AP entry catches itself maybe never, because the next AI step downstream accepts it as input and processes it further. Putting AI in the middle of a chain without checkpoints turns a small error into a long tail of derivative wrong outputs.
• Anything where the value was the human conversation. When a sales rep talks to a customer, the value is the relationship being built, not the transcript. When a manager has a hard conversation with a direct report, the value is the eye contact, not the words. AI is bad at the things that are actually the point of those moments.
• Anything you should be doing yourself because it's how you learn the work. Some tasks are productive precisely because the doing of them teaches you something about your business. Outsourcing those tasks to AI saves time and costs you the learning. For a junior employee, that trade is almost always bad. For an experienced operator, it might be fine. Be honest with yourself about which side of that line a given task is on.

How to think about it without overthinking it

If your team isn't using AI at all, you're probably leaving real time on the table on the drafting, summarizing, and document-shuffling kind of work. Start there. Low risk, fast feedback, easy to measure.

If your team is using AI everywhere without checkpoints, you're probably accumulating quiet errors. Step back. Find the places where output goes straight to a customer, a regulator, or a downstream system without a human read, and put a checkpoint there.

If someone is selling you an AI strategy, ask them what specific task in your business it makes better, and how they'll measure the improvement. If the answer is more deck slides, you're being sold the wrong thing.

The right posture for AI in a small or mid-market business is the same as the right posture for any other tool. Use it where it fits. Don't use it where it doesn't. Be willing to change your mind as the tool gets better at things it isn't good at yet. That's it.

← Back to all posts