Describe the process without using the word “somehow”.
Who receives the information? What do they check? Which decision do they make? What happens when something is missing? If nobody can answer, the first job is understanding the work. A form, a shared template or removing an unnecessary approval may beat AI. Varied documents, messages and connected systems are where interpretation can become useful. Buying a tool does not count as answering these questions.
Six jobs. None of them “make us an AI company”.
The enquiry with three attachments and half the facts.
A manufacturer receives a request by email. Specifications are scattered across the message and attachments; something essential is missing. A workflow assembles a summary, flags the gaps and drafts follow-up questions. The salesperson checks it and decides what to offer. Measure preparation time and missed questions before quoting. The example is a workflow to test, not a promise that every attachment will behave.
The PDF somebody retypes every Thursday.
A distributor copies product codes, quantities and delivery dates from supplier documents. Extract the agreed fields into a review table, mark uncertain or conflicting values and let a colleague approve the import. Product matching and totals follow explicit rules. Count correction time and accuracy as well as typing time. A fast transcription error is still an error.
The question answered before. And before that.
A hotel or service business answers familiar questions about arrival, facilities and how things work. An assistant can draft a reply from approved information and show its sources. A person checks and sends it. Live availability, current prices and booking changes need the right system connection and permissions. Confidently inventing a spare room is not hospitality.
The catalogue that is always almost ready.
An ecommerce team has specifications in one place, photographs in another and supplier copy everywhere. Assemble a product-page draft, flag missing facts and prepare language variants. A person checks claims and translations before publication. If your source material says nothing useful, the model cannot responsibly invent a product advantage to cover for you. Measure the effort to finish the draft, not how many words it spits out.
The weekly report that eats part of the week.
Ordinary automation collects exports and calculates the figures. AI can draft a short explanation of changes, with each figure traceable to its source. The team checks unusual movements and decides what they mean. Keep arithmetic in deterministic code. A language model should not be invited to have a creative relationship with your totals.
The meeting that forgot to become work.
Turn meeting or site-visit notes into proposed actions, owners and dates. Mark unclear commitments instead of inventing them. A person approves the handover before tasks are assigned. Useful when information gets lost between a conversation and the next working day. If the meeting made no decisions, the summary is allowed to reveal that. Blaming the AI will not make one appear.
Your preferred buzzword does not choose the tool.
- A clear rule: ordinary automation. A submitted form creating a task does not require artificial intelligence to reflect on the human condition.
- Information to interpret: AI can help read, sort or draft. Review belongs where a mistake would matter.
- Several connected steps: an agent may select from approved actions, use tools and check progress towards a defined result.
- A sensitive or unusual decision: keep the person who understands the context in charge.
An agent is software that can use connected tools across several steps. It needs a defined result, usable information and limits. “Figure it out” is not a process specification. It is the phrase you use before discovering how differently two parties understood the assignment.
The pilot may tell you your idea is bad.
- Show the actual work. Recent examples, awkward exceptions, frequency, time spent and mistakes. Invite the person doing it, not only the person announcing the transformation.
- Define one small test. A result, required connections, access, running costs and what stays with the team. “Everything” is not a first phase.
- Count all the effort. Include review, corrections and exceptions. Saving ten minutes while creating twenty minutes of supervision is not a win.
- Keep what earns its place. A working setup, instructions, clear responsibilities and a way to pause it or return to the previous process. Cancelling a bad idea is an available outcome.
A spectacular demo can still be a useless expense.
Total effort saved after checks, corrections, exceptions and running costs is the useful number. You may prefer the answer that justifies the purchase you already imagined. The test does not owe you that answer. A simpler change, a narrower task or leaving the work with a person can all be better outcomes. There is no consolation prize for using the most AI.
The questions the demo usually skips.
Will it send messages or change records by itself?
Only within explicitly agreed permissions. A first version will often prepare drafts and proposed changes for human approval. Sending, publishing, ordering and deleting each need rules. If you want to skip the controls because the demo looked clever, we will have an awkward conversation before any customer receives an experimental email.
Does it work with the tools we already use?
That is checked before scoping. Available connections, permissions and data quality determine the options. A document set or export may be enough for a first test. If you cannot get access to a system, neither enthusiasm nor calling it an AI project creates that access.
What happens when it does not know?
It should flag missing information, stop or hand the task to a person at defined points. Those cases get tested alongside the easy ones. “It sounded certain” is not an acceptance criterion. Neither is your insistence that uncertainty would look bad in the presentation.
Where does our information go?
That depends on the chosen tools and model provider. Before implementation, the data flow and necessary access are explained and agreed: which systems receive which information, for what purpose. The word “private” in a sales presentation is not a technical description.
Bring a task. Leave the prophecy.
What arrives? What does someone do? What should come out? Show a recent example, including the annoying exception. If you cannot yet describe the job, that is where the conversation starts.