How to Use AI in Business: The Implementation Sequence
How to use AI in business is a sequencing problem, not a tool problem. Pick one narrow process. Measure what it costs you today. Build the smallest version that touches real work. Decide who approves the output before it leaves the building. Then compare against your baseline and either scale it or kill it. Everything else is a pilot that never ends.
That sequence is unglamorous, and it is the part most companies skip. They buy seats, run a company-wide enablement session, and six months later have a lot of people using AI to write emails slightly faster and nothing on the P&L. The fix is not more ambition. It is narrower scope and an earlier baseline.
TL;DR
- The US Census Bureau's Business Trends and Outlook Survey put overall AI use between 17% and 20% of US businesses from December 2025 through May 2026, rising to 37% among firms with 250 or more employees. Broad adoption is still ahead of you, not behind you.
- Census researchers found 57% of AI-using firms run it in three or fewer business functions and 65% limit it to three or fewer tasks. Concentration is the normal pattern, so treat breadth as a warning sign.
- The leading functions are sales and marketing at 52%, strategy and business development at 45%, and IT at 41%.
- Token cost is a rounding error. Coding 600 supplier invoices a month on Claude Sonnet 5 runs about $6. The build and the supervision are the real line items.
- Stanford's 2026 AI Index puts agent deployment in the single digits across nearly every business function. Your first build should be a workflow with a human gate, not an autonomous agent.
Step 1: pick one process, and make it narrower than feels right
The instinct is to pick a department. Resist it. Pick a task with a verb and a count: "code supplier invoices to the right GL account," "route inbound support tickets to the right queue," "draft the first version of a quote from an RFQ email." If you cannot say how many times it happened last month, you have not picked a process yet.
Two filters decide the first one. Count: at least a few hundred instances a month, so a small per-item saving adds up to something you can see. Blast radius: a wrong answer is caught cheaply and reversed cheaply. High count and low blast radius is where you start. Low count and high blast radius is where AI projects go to die.
This narrowness is not a beginner's compromise. It is what adopting firms actually do. In an April 2026 Census Bureau working paper, The Microstructure of AI Diffusion, Bonney, Breaux, Dinlersoz, Foster, Haltiwanger and Pande found that 57% of AI-using firms integrate it into three or fewer business functions, and 65% limit use to three or fewer tasks. The same paper puts the leading functions at sales and marketing (52%), strategy and business development (45%), and IT (41%), and finds that 66% of users apply AI to augment existing work rather than replace it, with employment reductions reported at only 2% of firms. If you want the map of which functions move which P&L lines, that is the subject of AI in business. This post is about what you do on Monday once you have picked one.
Step 2: instrument the process before you automate it
You cannot prove a saving against a number you never wrote down. Before anything AI-shaped touches the work, log three things for two weeks:
- Volume. How many items, by week.
- Time per item. Not a guess. Have the person doing it start a timer on ten real items.
- Error rate and cost of error. How many needed rework last quarter, and what each one cost in hours or credits.
Two weeks of a spreadsheet is enough. This is the cheapest step and the one that gets skipped, which is why so many AI programs end with a vague claim about productivity and no budget renewal. If you want the full measurement rig, including the haircuts that separate a demo number from a P&L number, I worked that through in how to measure AI ROI. For step 2 you only need the baseline.
Step 3: build the smallest version that touches real work
Smallest version means: one prompt, twenty real examples pulled from last month, and output that a human reads. Not a platform. Not an agent. A workflow with fixed steps, because that is what almost everyone is actually running successfully right now.
Stanford HAI's 2026 AI Index reports that 88% of surveyed organizations use AI in at least one capacity and 70% deploy generative AI in at least one business function, while agent deployment remains in the single digits across nearly all business functions. That gap is the whole lesson. The stuff that works at scale today is bounded: read this, classify it, draft that, hand it to a person. If you are unsure which shape your task needs, AI workflow vs AI agent draws the line. When you do move from one-off prompting to something repeatable, the pattern to steal is in Write Loops Not Prompts.
Run the smallest version in parallel with the manual process for two weeks. Same inputs, both outputs, a human comparing. You are not looking for perfection. You are looking for a defensible percentage: how often the AI output was accepted with no edit, with a small edit, or thrown away.
Step 4: set the review gate before the first output leaves the building
Every output belongs in one of three lanes, sorted by whether a mistake can be undone and what it costs if nobody catches it. Decide the lane before you write the prompt, because the lane determines how much accuracy you actually need.
The lane also sets your quality bar honestly. A routing decision that is right 92% of the time is fine if the 8% lands in a queue someone clears daily. A GL code that is right 92% of the time is fine only if a controller reviews every exception. The same accuracy is excellent in one lane and unacceptable in another. The mechanics of building that checkpoint into the system, rather than into someone's good intentions, are in human in the loop AI agents.
Then name a person. Not a team, a person. Lanes without an owner degrade into unreviewed output within about a month.
Step 5: measure against the baseline, then scale or kill
At the end of the parallel run, you have four numbers: acceptance rate, time per item now, volume, and exceptions. Put them next to your two-week baseline and make one of three calls.
Scale if the time saved clears the build cost inside a quarter and the exception queue is stable. Re-scope if the acceptance rate is strong on a subset. Most first builds fail on 20% of cases that share a trait, so cut those out and ship the other 80% rather than trying to fix everything. Kill if the accepted output still needs as much review time as doing it manually. That is a real answer and it cost you three weeks, which is a good trade.
The re-scope arrow in the first figure matters more than it looks. The common mistake is to respond to a mediocre result by running a second pilot on a different department. Run it again on a narrower slice of the same process instead. You already have the baseline, the examples, and the reviewer.
What the smallest version actually costs
Take supplier invoice coding at 600 invoices a month. Each one is roughly 3,000 input tokens (extracted text, your chart of accounts, a few worked examples) and 400 output tokens.
- Input: 600 × 3,000 = 1.8M tokens. On Anthropic's published pricing, Claude Sonnet 5 is $2 per million input tokens, so $3.60.
- Output: 600 × 400 = 0.24M tokens at $10 per million, so $2.40.
- Total: about $6.00 a month, or $3.00 if you run it through batch processing at the stated 50% discount.
Now the baseline. Four minutes per invoice across 600 invoices is 40 hours a month. At a loaded $28 an hour, that is $1,120. If the model handles 70% cleanly and a human reviews the 180 exceptions plus spot-checks 10% of the automated lane, you land near 13 hours, call it $364. The saving is roughly $756 a month against $6 of usage.
Which means the token bill is not the decision. The decision is the build, which is two or three days of a competent operator's time, plus the recurring slice of a controller's week spent on the exception queue. Price those two lines and ignore the API bill until you are running millions of items. If finance is the function you are starting with, the agent patterns in Claude Finance Agents are a faster starting point than a blank prompt.
Three ways business AI projects die
Buying seats instead of picking a process. A company-wide rollout produces a usage chart and no baseline. Usage is not a result. If nobody can name the process, nobody can measure the change.
Automating the irreversible lane first. It is tempting because that is where the money is. It is also where a single bad output costs more than the whole year's saving, and where one incident ends the program permanently.
Widening before proving. The Census data says most adopting firms run AI in three or fewer functions. Breadth comes after a working narrow build, not instead of one. The same US Census Bureau BTOS release from May 2026 put overall AI use between 17% and 20% of businesses, and 37% among firms with 250 or more employees, with the Information sector at 39.7% and finance and insurance at 33.9%. Most of the market is still early. There is no penalty for being deliberate, and a real one for being everywhere at once.
The bottom line
How to use AI in business reduces to five decisions made in order: which process, what does it cost today, what is the smallest thing that touches real work, who approves the output, and did it beat the baseline. Everything expensive happens when those get done out of order or not at all. The tooling is the easy part in 2026. The sequencing is the job.
Pick your process this week. Time ten items. That is step one and step two, and it costs you an afternoon.
Want the next one in your inbox? Join the newsletter for one practical AI build a week. And if you want the workflow pattern behind step three, start with Write Loops Not Prompts, the loop I use to turn a one-off prompt into something that runs every day.
How do I start using AI in my business?
Start with one process, not one tool. Write down a task your team already does dozens of times a month, where the input is text or documents and the output is a draft, a classification, or a routing decision. Time it for two weeks so you have a real baseline. Then build the smallest version that touches actual work: one prompt, twenty real examples, one person reviewing every output. Run it in parallel with the manual process for two weeks and compare. If it does not beat the baseline on time or accuracy, you have learned that cheaply. If it does, you now have a number you can defend when you ask for budget to widen it. The sequence matters more than the model you pick.
Which business process should I automate with AI first?
Pick the one with the highest count and the lowest blast radius. High count means the task happens at least a few hundred times a month, so a small per-item saving compounds into something visible. Low blast radius means a wrong answer is caught cheaply and reversed cheaply: an internal draft, a routing decision, a first-pass classification. Census Bureau researchers found the leading AI functions are sales and marketing at 52%, strategy and business development at 45%, and IT at 41%, which mostly tracks where repetitive text work lives. Avoid anything that sends money, signs a contract, or emails a customer unsupervised on day one. Those become the second build, once you have a review gate that already works.
How much does it cost to use AI in a business?
The API bill is almost never the constraint. Claude Sonnet 5 is listed at $2 per million input tokens and $10 per million output tokens on Anthropic's pricing page, with batch processing at half that. A job that reads 600 supplier invoices a month at roughly 3,000 input and 400 output tokens each comes to about $6 of usage. The real cost is the build and the supervision: two or three days of someone's time to write the prompt, assemble examples, and wire up the review queue, then a recurring slice of a human's week to check output. Price the project on those two lines. If the token cost is what makes your business case work, the case is too thin.
Do I need to hire an AI engineer to use AI in my business?
Not for the first build, and hiring one before you have a defined process is the most common way to waste six figures. The first version of most business AI work is a prompt, a set of real examples, a spreadsheet or queue for review, and a person who owns the output. An operator who knows the process well can build that in a few days with a general model and no custom code. You need engineering when the thing has to run unattended against live systems: pulling from your database, writing back to your ERP, handling retries and rate limits, and surviving a model change. That is a real project, and it is much cheaper to scope once a working manual version has proved the value.
How do I stop AI from making expensive mistakes in my business?
Sort every output by two questions before you write a line of prompt: can a mistake be undone, and what does it cost if it is not caught. Reversible and cheap output can go straight through with sampling, where you check a fixed percentage each week. Reversible but costly output gets drafted by AI and sent by a human. Irreversible output, meaning money moving, contracts, public statements, or anything touching a regulator, stays human-executed with AI only preparing the material. Then name one person who owns each lane. The failure mode is not that AI makes mistakes, because humans do too. It is that nobody was assigned to catch them, so nobody did.
OpusJake is Jake Schincariol's operating system for building with AI: agents, workflows, prompts, and the free resources behind them. Get the next move every week.