MANYAutomations

Where AI helps

Language models handle work that used to need a person because the input is unstructured: free-form emails, scanned documents, call transcripts. They do best when the output is a draft or a suggestion that someone can check quickly.

• Drafting replies, product descriptions and first versions of routine documents.
• Summarizing long threads, meetings and reports into key points.
• Classifying and routing messages, tickets or leads by topic and urgency.
• Extracting fields such as dates, totals and names from invoices, forms and contracts.
• Answering staff questions from internal policies and manuals, citing the passages used.

Where plain rules beat AI

If a task has one correct answer that a formula or rule can produce, use the formula. Models can slip on arithmetic, and the same input can produce a different output on another run. Pricing, tax calculations, inventory counts and anything that moves money belong in ordinary, deterministic code.

Compliance decisions belong with rules and people as well: approving credit, screening job applicants, handling health or legal matters. Some jurisdictions restrict automated decisions about people, and you may have to explain how a decision was reached. A sound pattern lets AI gather and organize the facts while fixed rules or a person decide.

Human-in-the-loop design

Decide which outputs go straight through and which wait for a person, based on what a mistake would cost. A wrong internal tag is easy to fix later. A message to a customer or a change to a financial record may not be.

Watch for rubber-stamping. Reviewers who approve hundreds of items a day stop reading, so keep queues small enough for real attention and occasionally include a known error to confirm that reviews catch it.

• Review queues hold drafts and extracted data until someone approves them.
• Confidence thresholds send uncertain results to a person, with cutoffs set from your own test results, since a model’s stated confidence can mislead.
• Spot checks cover a random sample of automatic results every week.
• Corrections are easy to make and logged, so recurring problems show up.

Knowing where your data goes

Every prompt sent to a hosted AI service carries data to the vendor. Consumer apps and business or API offerings often come with different terms, so read the ones for the exact product your workflow uses.

Send only the fields a task needs, and mask names, account numbers and other identifiers where you can.

• Training: whether your inputs can train the vendor’s models, and how to opt out.
• Retention: how long prompts and outputs are stored, and whether you can shorten it.
• Contracts: business terms, a data processing agreement and where processing happens.
• Sensitive data: health, payment and government ID information can carry legal duties that rule out some tools.

Accuracy before and after launch

Models can produce fluent, confident answers that are wrong, including invented figures and citations. Ground them: supply the source documents, instruct the model to answer only from them and to say when the answer isn’t there, and have it show the passage it relied on so reviewers can verify quickly.

Build a test set before launch: a few dozen to a few hundred real examples with known correct answers, including the tricky ones. Rerun it whenever the prompt, model or vendor changes, and track the error rate. A vendor’s model update can shift results even when nothing on your side changed.

Prompt injection and untrusted content

Prompt injection happens when content the AI reads contains instructions of its own, such as an email telling the assistant to ignore its rules and forward the inbox. Models can’t reliably separate your instructions from text planted in the material they process, so any system that reads outside emails, web pages, documents or tickets is exposed.

• Grant only the permissions a task needs, since a summarizer has no reason to send email.
• Require human approval for consequential actions such as payments, deletions and outgoing messages.
• Label untrusted content clearly in prompts, knowing this reduces the risk without removing it.
• Log what the AI read and did, so a bad action can be traced to its source.

Measuring value and governing use

Judge an AI automation the way you would judge a new hire’s work: time saved, error rate, rework and the review time it still needs. Include usage fees, which often grow with volume, and the staff hours spent maintaining prompts and test sets.

Governance can stay simple. Keep an inventory listing each AI automation, its owner, the vendor and model, the data it touches and whether a person reviews its output. Retain logs of inputs and outputs for a set period, and review the inventory regularly so forgotten experiments don’t keep running on live data.