Artificial intelligence in business: 12 practical uses, with the numbers to check

Artificial intelligence in business is most useful where its output can be checked: drafting and summarizing text, classifying documents and tickets, searching company knowledge, writing code, and turning questions into analyses that tools compute. It is weakest where it must produce exact figures unaided. Start with one process, measure time or error rate before and after, and keep a person responsible for the result.

Where artificial intelligence earns its place in a company

Every good use shares one property:

The test: can a person check the output faster than they could have produced it?

A drafted email is checked by reading it. A summary is checked against the meeting. An extracted invoice total is checked against the PDF. Where checking is cheap, the time saved is real. Where checking is as hard as doing the work, such as a figure produced by a language model with no rows behind it, the saving is an illusion and the risk is a wrong number in a board pack. Why AI gets numbers wrong explains the mechanism.

12 practical uses

Each use comes with the one measure that shows whether it worked.

1. Drafting emails, proposals and RFP answers. Measure: time to first draft, and edits needed before sending.
2. Summarizing meetings and long documents. Measure: minutes spent per summary, and how often a reader returns to the source.
3. Customer-service triage. Classifying and routing tickets by topic and urgency. Measure: time to first response and share of tickets routed correctly first time.
4. Invoice and document extraction. Reading supplier invoices, receipts and forms into fields. Measure: minutes per invoice and exception rate in accounts payable.
5. Knowledge search. Answering staff questions from policies, contracts and past proposals, with links to the passage. Measure: time to find an answer; share of answers with a valid citation.
6. Code assistance. Writing and reviewing scripts, SQL and spreadsheet formulas. Measure: cycle time per change and defects found in review.
7. Contract review, first pass. Flagging clauses that differ from the standard template for a lawyer to read. Measure: hours per contract and clauses missed in a sample check.
8. Translation. Product sheets, support replies and supplier correspondence. Measure: cost per page and corrections by a native reader.
9. Sales-call notes to the CRM. Turning a call into next steps, contacts and a logged activity. Measure: share of calls logged within a day.
10. Anomaly flags in transactions. Duplicate payments, unusual journal entries, out-of-pattern expenses. Measure: flags raised, share confirmed as real, and dollars recovered.
11. Questions of business data answered by tools. A manager types a question; the external AI model chooses the analysis, deterministic tools compute it from the rows, and the answer cites them. Measure: time from question to checked answer. AI data analysis and AI dashboards cover this use in depth.
12. Demand planning with statistical models fitted on your data. Projecting order volumes from history and seasonality. Measure: forecast error against actuals, compared with the old method. Predictive analytics sets out the approach.

Uses 1 to 9 are text work, where language models are strong. Uses 10 to 12 are number work, and they work only when the arithmetic is done by tools and statistical models, with the language model reading and explaining.

What the adoption surveys say

Adoption figures vary widely with who is counted.

The US Census Bureau's Business Trends and Outlook Survey asks a large sample of US employer businesses every two weeks whether they use AI. As of May 3, 2026, 19.8% of businesses reported current AI use, and between 20% and 23% expected to be using it within six months. Use rises with size: 37% of firms with 250 or more employees, 32% of those with 100 to 249. By sector, information led at 39.7%, and finance and insurance was at 33.9%.

The Stanford AI Index 2026 reports that organizational adoption reached 88%. The gap between 88% and 19.8% is mostly a difference in who is asked and what counts as use: a survey of organizations versus every employer business, from the largest to the smallest. For a big or medium company, the honest reading is that peers are using AI somewhere, and the question is which processes and with what result.

Worked example: a question asked of a sales ledger

Use 11 in detail. The question typed: "Which customers bought less in Q3 2026 than in Q3 2025, and by how much?" The data: the invoice ledger, rolled up to customer. A period-comparison tool computes each customer's quarter totals and the change; the external AI model only chose the tool and wrote the sentence.

Change = current-period value − prior-period value

Change % = current / prior − 1

The answer returned: three customers bought less, and together they fell $75,600: Harbor Freight Services −$40,500, Marlow Clinics −$22,500 and Bayside Hotels −$12,600. The three growers added $24,500. The answer cites the tool, the two periods and the 6 rows it read. In Excel, the same change column is =C2-B2 and the change % is =C2/B2-1.

The check

Decliners plus growers must equal the net change:

−75,600 + 24,500 = −51,100

Both quarter totals, 753,300 and 702,200, must equal the ledger's revenue for those quarters, filtered the same way. If the AI-written sentence says anything the table does not, the table wins. This is the design rule in why the arithmetic must never be left to the language model: a figure that cannot be traced to rows is not reported.

How to measure payback

Pick one process, measure it before, pilot on one team, measure it after, the same way.

Payback (months) = one-time cost / (monthly hours saved × loaded hourly cost)

If the tool has a monthly fee, subtract it from the monthly saving first. A template, filled in for invoice extraction in accounts payable:

3,000 × 9 / 60 = 450 hours; 3,000 × 4 / 60 = 200 hours; 250 × 48 = 12,000; 36,000 / 12,000 = 3.0 months. Track the exception rate beside the time: a faster process that sends more invoices to manual review has moved the work, not removed it.

Risks to manage

• Wrong figures, stated confidently. Keep arithmetic in tools and require cited rows for any number.
• Data leaving the company. Check retention, training and location terms before uploading customer or employee data; is it safe to upload customer data to an AI analytics tool lists the questions.
• Bias in automated decisions, such as credit limits or candidate screening, where a person must review the outcome.
• Over-reliance, where staff stop checking output because it is usually right.

The US checklist is the NIST AI Risk Management Framework, released January 26, 2023 and intended for voluntary use. It organizes the work into four functions, Govern, Map, Measure and Manage, and NIST added a Generative AI Profile on July 26, 2024 for the risks specific to language models.

Where it goes wrong

• A general chatbot asked to add up or rank figures, and its answer pasted into a report with no rows behind it.
• Measuring adoption instead of outcome. Logins and prompts per week say nothing about time per invoice or first-contact resolution.
• Customer or employee data uploaded to a tool before anyone read its retention and training terms.
• No baseline. A pilot on a process nobody measured before cannot show a gain.
• Automating a step that should have been removed. Faster approval of a report nobody reads is still waste.

Artificial intelligence for your own business data

Covirage is use 11. You ask a question of your own files; its tools compute the answer from your rows and cite them, and the external AI model chooses the tool and explains the result without doing the arithmetic. See AI analytics for how questions about your business files are answered by tools that compute every figure. For single uses in more depth, see AI for small business, AI for financial analysis and AI data analyst.