The office manager has three AI assistants. One is in her email, one is in the word processor and one appeared in the accounting software last month. Each of them can summarise, draft and suggest.
The supplier invoice that arrived on Monday is still waiting for approval on Thursday.
None of the assistants did anything wrong. The invoice was not stuck because nobody could read it. It was stuck because nobody knew it was their turn.
That is the difference this article is about. An AI feature helps a person with a task. A workflow moves a piece of work from start to finish, across people and systems, without relying on someone remembering. Both are useful. They solve different problems, and many SMEs pay for the first while their real cost sits in the second.
One helps a task. The other moves the work.
Same invoice, two kinds of help. Only one of them knows whose turn it is.
“Summarise this invoice.”
- Waits to be asked
- Helps one person, one step
- Done when the window closes
- ReceivedOwner: admin
- CheckedOwner: accounts
- To approveOwner: manager
- Paid & recordedLogged
Starts on its own, shows the status, knows whose turn it is.
What the tools you already pay for are good at
Most business software now includes some form of AI. It usually works inside one screen, for one person, at the moment they ask for help.
That is genuinely valuable for work like:
- ✓drafting or rewording an email
- ✓summarising a long thread or document
- ✓turning meeting notes into a tidy list
- ✓explaining a spreadsheet formula or cleaning a column
- ✓translating a message before it goes out
If the slow part of a job is one person writing, reading or tidying, start there. It costs little or nothing extra, needs no project, and the person stays fully in control. For many small teams, simply using these features well is the right first step, and it would be poor advice to build something custom instead.
Where a feature stops
A feature waits to be asked. It works on what is in front of the person using it. When they close the window, it is done.
Now look at where time is actually lost in most SMEs. It is rarely inside one task. It is between tasks:
- •a request arrives and nobody is sure who owns it
- •a document is ready but the approver has not seen it
- •a client was promised a reply and the promise lives in one person’s head
- •information is typed again from an email into a system, then into a spreadsheet
- •the manager asks “where are we with this?” because there is nowhere to look
A writing assistant cannot fix any of these, however good it is. The problem is not the quality of one step. It is that the steps are not connected, nobody can see the status, and nothing happens unless someone remembers.
What a workflow adds
A workflow is a designed path for one type of case. It knows where work comes in, what happens next, who is responsible at each point and what “finished” looks like.
In practice, that means a few things a feature does not provide:
A trigger
The work starts when something happens, such as an email arriving, a form being submitted or a date passing, not when someone thinks to open a tool.
A status
Every case is somewhere visible: received, waiting for information, ready for approval, done.
An owner at each step
When it is someone’s turn, they know, and so does everyone else.
Rules that do not drift
Who approves what, which template applies, when a reminder goes out.
A record
What was decided, by whom and when, without anyone writing it down separately.
AI can sit inside a workflow, reading an incoming document, drafting a reply or classifying a request. But the AI is one step. The value usually comes from the path around it.
Four questions to tell them apart
For any process you are thinking about improving, ask:
Is the slow part inside one person’s task, or between people?
If one person spends too long writing or reading, a feature may be enough. If work waits between people, you need a workflow.
Does it need to happen without someone remembering?
Reminders, follow-ups and deadlines fail because they depend on memory. A feature cannot start itself. A workflow can.
Does anyone need to see the status?
If a manager, colleague or client regularly asks “where is this?”, the information exists in no shared place. That is a workflow problem.
Does the work touch more than one system?
When the same details are copied from an inbox into a CRM and then into an invoice, the cost is in the copying. Features inside each tool do not talk to each other.
If you answered “one person, no, no, no”, use the feature you already have. If two or more answers point the other way, the problem is probably a workflow.
A common mistake in both directions
The first mistake is building too early. A team pays for a custom setup to draft emails that the tool in their inbox could already draft. The project adds maintenance, an extra login and a new thing to break, for a gain a feature would have delivered.
The second mistake is expecting features to fix process problems. A business rolls out AI assistants to everyone, sees better-written emails, and is then surprised that follow-ups are still missed and approvals still sit for days. The writing improved. The process did not change.
Both mistakes come from the same place: starting with the tool rather than the problem.
Start with what you have, then look at the gaps
A sensible order for most SMEs:
- 1.Use the AI features already included in your tools for single-person tasks. Agree a few simple rules, such as what may not be pasted in and when a human must check the output.
- 2.Write down where work waits, gets chased or is typed twice. That list is usually short and specific.
- 3.Pick one of those gaps with a clear start, a clear finish and a visible cost.
- 4.Design a workflow for that one case, with AI only where it genuinely reads, drafts or classifies, and fixed rules for everything else.
This keeps spending proportionate. You only build where the tools you already pay for cannot reach.
A quick self-check
For one process in your business, write down:
where the work starts, and what triggers it
every hand-off between people or systems
where a case can wait without anyone noticing
how someone outside the task would find its current status
which step is slow because of reading or writing, and which is slow because of waiting
If most of the delay comes from reading or writing, sharpen the use of your existing AI features. If it comes from waiting, the improvement is a workflow, and the AI is only one part of it.