A large language model is a machine that has read most of the public internet and learned, from all that reading, what word tends to come next. Give it the start of a sentence and it continues the sentence. Give it a question and it continues the conversation the way a good answer would continue it. That is the whole machine.
I know how that sounds. You have watched it summarize a contract in eight seconds, and the summary was right, and "it guesses the next word" reads like a description written by someone who never tried the thing. Keep it anyway. Every decision you will make about this tool gets easier once you hold onto how small the mechanism is, even while the performance is trying to talk you out of it.
Start with what the mechanism buys you. A surprising share of a workday in your accounting group is language with a known shape: the collections email, the vendor thread that ran to forty messages while somebody was out. The model has seen each shape thousands of times and produces another one, in seconds, at a quality that startles people who came in expecting a chatbot. And there is a measured version of this. Brynjolfsson, Li, and Raymond studied 5,179 customer support agents at a Fortune 500 software company and found an AI assistant lifted issues resolved per hour by 14%. The lift landed almost entirely on the newest agents, who improved 34%. The veterans barely moved. The tool had learned what the best agents already knew, and it handed that knowledge to everyone else.
Read that result again, because it is the honest ceiling. The tool spreads competence that already exists somewhere. Where the right answer is known and only needs to reach more hands, it works, measurably. Your two newest AP clerks will feel it most. Your controller may feel nothing. Budget accordingly.
Now the failure, which deserves more of your attention than the success.
The model has no separate place where facts live. Producing the sentence is the entire act; there is no inner ledger it checks first. So when the true answer is missing from everything it has read, it produces the most plausible sentence anyway, fluent and confident, because fluent and confident is what its training rewarded. Ask what your own freight contract says and it will tell you, having never seen your contract, because it has seen ten thousand freight contracts and knows what that clause usually says. A person who is guessing slows down and hedges. This machine has no such tell. The right answer and the invented one arrive in identical packaging.
Now the harder part. This equipment can fail while your best people report that it is working.
In July 2025 a research group called METR ran a controlled trial with 16 experienced open-source developers on 246 real tasks in codebases they knew well, each task randomly assigned to AI-allowed or AI-forbidden. With the AI, the developers took 19% longer. Afterward, the same developers estimated the tool had made them 20% faster. Skilled people, working in their own domain, misjudged the tool's effect on their own output by nearly 40 percentage points, in the direction that felt good. The drafts looked polished, reviewing them felt like progress, and the clock disagreed.
That perception gap is the fact I want you to carry out of this piece. Your team's enthusiasm after the pilot will be sincere, and it will be weak evidence. So will their skepticism. The one trustworthy report on this tool is a measurement: pick the tasks, count the errors, time the work, compare.
The demo that dazzles the conference room tells you even less. A demo is the model doing the one thing it was built to do, producing plausible language on a stage where nobody checks a single fact against the world, and every vendor knows it.
One more limit, briefly. The model learned from text collected up to a certain date, and it has read none of your systems unless someone deliberately connects them. Most business versions can now search the web and read files you hand them, which narrows the gap. Ask it for your current inventory position anyway and it will still answer, because a plausible answer exists and producing plausible answers is its job. When the model is working from nothing, the behavior above governs.
Where does this leave your accounting team? Give the model the work where a competent reviewer can judge the output at a glance: the drafted email, the first pass at a reconciliation memo. A wrong draft you catch costs a minute. A wrong number that reaches a customer costs the relationship. Keep a person on every judgment, weight the rollout toward your juniors because that is where the evidence says the gain lands, and settle the question with a measurement.
Verified August 2026