How is AI being used in operations work?
Operations staff use AI mostly for the written and analytical work that keeps processes running, not for running the process itself. The value comes from tasks that are frequent, tedious and easy to check against a source.
The list below is what a general assistant such as ChatGPT, Claude, Gemini or Copilot can do with material you are allowed to share:
Draft or update an SOP
Turn rough notes or a talk-through of a task into a numbered procedure, then have the process owner correct it.
Compare supplier quotes
Lay several quotes side by side on the same basis: unit price, surcharges, minimums and terms. Recalculate the totals yourself.
Write checklists
Turn a procedure into a checklist for a new starter, a month-end close or an office opening.
Clean up lists
Standardise inconsistent supplier names, addresses or product codes, or suggest a spreadsheet formula to do it.
Summarise incidents
Turn an email chain about a failed delivery or system outage into a timeline, cause as stated, and open actions.
Find gaps in a handoff
Ask what information the next team needs that a handoff form does not ask for.
Answer policy questions
Paste an internal policy and ask where it covers a situation, then read that section yourself.
Should you use an AI assistant or automate the process?
Start with an assistant, where a person checks each output, and consider automation only for steps that are stable, well documented and cheap to get wrong. Many searches for AI in operations are really about automation and AI agents, which are a different kind of project with a different risk profile.
The comparison below is a planning aid, not a technical assessment of any product.
| Approach | Who checks the output | Typical failure | Good first use |
|---|---|---|---|
| Assistant used by a person | The person, every time | An invented step or a wrong figure that the person misses | Drafting an SOP or comparing quotes |
| Rule-based automation | Whoever designed the rules, once | A case the rules did not anticipate | Moving data between two systems in a fixed format |
| AI agent acting on its own | Often nobody, at the moment it acts | A confident wrong action repeated at scale | Only after the process is mapped, owned and tested |
What does an AI-drafted SOP get wrong?
It fills gaps with steps that sound standard but are not how your organisation works. That is more dangerous in an SOP than in an email, because people follow an SOP without questioning it.
A fictional example. Ashford Print Co is a small print shop. The operations lead's notes on handling a rush order: 'Check artwork is print-ready. Confirm paper in stock. Rush fee applies under 48 hours. Production manager must OK anything over 500 units. Customer gets proof by email.'
A loose prompt, 'Write an SOP for rush orders', produced a tidy ten-step procedure that included 'Notify the customer by SMS when the order ships', 'Orders over 250 units require director approval' and 'Offer a 10% discount if the deadline cannot be met.' None of those is in the notes. The approval threshold is wrong and the person is wrong, and the discount is a commercial decision nobody made.
The better prompt: 'Turn these notes into a numbered SOP. Use only the steps in the notes. Where a step seems to be missing, list it as a question for the process owner instead of adding it.' The output kept the five steps, gave the 500-unit approval to the production manager, and listed three questions: who checks the artwork, what happens when paper is out of stock, and how the customer approves the proof. The operations lead answered them, and only then was the SOP published.
Want to see how this connects to a role?
Explore the planned team trial ↗How do you check AI when it compares supplier quotes?
Recalculate the comparison yourself on the same basis, because models are prone to comparing headline prices and missing surcharges. Arithmetic in fluent prose is easy to accept without checking.
Fictional quotes for the same parcel service: Supplier A at £4.10 per parcel plus a 12% fuel surcharge; Supplier B at £4.40 per parcel, all-in; Supplier C at £3.95 per parcel plus a £0.60 handling fee. An assistant asked 'Which is cheapest?' answered Supplier C, on the £3.95 headline.
On a like-for-like basis: A is 4.10 plus 12%, which is £4.592, about £4.59; B is £4.40; C is 3.95 plus 0.60, which is £4.55. Supplier B is cheapest per parcel, and the order changes once the surcharges are included. A real decision would also need minimum volumes, contract length and service levels, which the quotes may state differently. Ask the assistant to build the comparison table, then check every cell against the quote documents and do the totals in a spreadsheet.
How do you start using AI in operations?
Start with one written process, one owner and one low-risk document task, and make the checking step explicit.
Choose a process with an owner
Pick something that runs often and that one named person understands well enough to spot an error.
Write down the current state
If the process exists only in someone's head, the first AI task is turning their explanation into a draft for them to correct.
Pick a document task
SOP drafts, checklists and quote comparisons are good starting points because a person can check them against a source.
Set the check
Decide who reviews the output and against what: the notes, the quote documents, the policy.
Agree the data rules
Decide which operational data, such as supplier terms or customer addresses, may go into which approved tool.
Review after a few weeks
Ask the owner whether the drafts needed less correction over time, and only then consider automating any step.
What are the main risks of AI in operations?
The risks are errors that spread because a process repeats them, data shared with the wrong tool, and staff deferring to output that looks authoritative. An invented step in one email affects one customer; an invented step in an SOP affects every order processed under it.
The NIST Generative AI Profile names automation bias, an excessive deference to automated systems, as a risk that can make others, such as confabulation, worse. Operations work is particularly exposed because much of it is about following procedures. The broader NIST AI Risk Management Framework, which is intended for voluntary use, is a useful reference if your organisation wants a structured way to think about these risks.
When is this guide, or Tosheo.ai, not the right answer?
If your goal is to automate processes across systems, select an ERP, implement robotic process automation or run process mining, you need a systems project with technical specialists, not a guide to everyday AI use. Those projects still benefit from staff who can describe and check a process, but they are a different investment.
Tosheo.ai is being built to train people, not to automate processes. The planned Operations role path covers tasks like the SOP and quote examples above through short lessons and practice tasks graded against a visible rubric. It does not connect to your systems or run workflows for you.
Sources and scope
NIST AI 600-1, Generative Artificial Intelligence Profile, July 2024 ↗
NIST, AI Risk Management Framework, released 26 January 2023 ↗