What current AI is good at, and where it struggles
Language models read, summarize, classify, extract and draft. They do it quickly and at any hour, and they are sometimes confidently wrong. That combination decides where they belong. They fit work where volume is high and each result can be checked, and they fit poorly where one unnoticed mistake is expensive.
Good fits
- Reading documents and pulling out specific fields
- Sorting incoming messages and sending each to the right place
- Drafting replies or reports for a person to review
- Answering questions from a known body of documents
- Comparing a document against a checklist or a policy
- Summarizing long threads, calls or case files
Poor fits
- Decisions with legal, medical or financial weight and no review
- Exact arithmetic or record keeping, where ordinary software is better
- Tasks whose rules nobody can explain, even after the fact
- Work that depends on information the system cannot reach
- Rare tasks, where the effort would never pay back
- Anything where being wrong once is unacceptable
List the candidate workflows
Start from the work, not from the technology. Spend an hour with each team and ask where the time goes. You are looking for queues, repetition and reading.
- 01
Follow the queues
Shared inboxes, ticket lists, document piles and approval backlogs show where work waits for a person.
- 02
Look for copy and paste
Wherever someone reads one screen and types into another, a step can probably be automated.
- 03
Ask what people would hand to a new assistant
Tasks that can be explained to a newcomer in an afternoon are good candidates.
- 04
Record volume and time
How many cases per week, and roughly how long each takes. Estimates are fine at this stage.
- 05
Write each candidate as one sentence
For example, read supplier invoices and enter them into the accounting system.
Expect a list of a dozen or more candidates. That is fine, because the next step sorts them. For ideas organized by job, our AI agent solutions catalogue lists common ones, each with its status stated.
Score each candidate on five criteria
| Criterion | The question to ask | Scores high when |
|---|---|---|
| Volume | How often does it happen, and how long does each case take? | Many cases every week, each taking minutes of a person's time |
| Clear pattern | Could an experienced person explain how they decide? | The rules can be written down, with examples of right and wrong |
| Data access | Is everything the task needs available in digital form? | Inputs are documents, messages or records the system can read |
| Safety | What does a mistake cost, and can it be undone? | Errors are cheap, visible and reversible |
| Checkability | How easily can a person confirm the result? | A glance is enough, or the system can show its source |
Scoring four workflows at a distribution company
Each workflow is scored from 1 to 5 on each criterion. The company and the scores are invented to show the method.
| Workflow | Volume | Pattern | Data | Safety | Check | Total |
|---|---|---|---|---|---|---|
| Entering supplier invoices | 5 | 4 | 4 | 4 | 5 | 22 |
| Answering order status emails | 5 | 4 | 3 | 4 | 4 | 20 |
| Drafting contract clauses | 2 | 3 | 3 | 1 | 3 | 12 |
| Setting customer credit limits | 2 | 2 | 4 | 1 | 2 | 11 |
| Highest possible | 5 | 5 | 5 | 5 | 5 | 25 |
Invoice entry comes first. It is frequent, it follows a pattern and every result can be checked against the document. Order status emails are close behind, held back only because order data sits in two systems. The last two score low on safety, and they stay with people for now.
Check the data and the systems before you commit
A workflow can score well and still be blocked by practical matters. Run through these for your top two or three candidates.
Readiness checklist
0 of 7 checkedDecide how much the AI should do alone
Suitable for AI does not mean handed over to AI. The same workflow can be supported at four levels, and the right one depends on what a mistake costs. A sensible path is to begin at the first or second level and move up as trust is earned.
| Level | What the AI does | Fits when |
|---|---|---|
| Suggest | Shows information or a recommendation beside the person doing the work | Mistakes are costly, or the workflow is new to automation |
| Draft for approval | Prepares the complete result, and a person approves or edits each one | Quality is good and each case still deserves a look |
| Act, with exceptions reviewed | Completes clear cases alone and routes uncertain ones to a person | Results are easy to verify and errors can be undone |
| Act within limits | Completes the task alone inside defined boundaries, with sampling | Volume is high, risk is low and quality has been proven over time |
Workflows to leave alone for now
- Processes that are broken in themselves. Fix the process first, or you will automate the confusion
- Tasks a simple rule or an ordinary integration would handle. They are cheaper and fully predictable
- Work where nobody agrees what a good result looks like
- Decisions that the law or your own policy says a person must make
- Anything you could not explain to a customer or a regulator afterwards
The second item deserves attention. Some automation requests turn out to need no AI at all. A form, a rule and a connection between two systems are often enough. See business process automation.
What to do next
Take your highest-scoring workflow and run a small pilot with success criteria agreed in advance. Planning an AI pilot with measurable success criteria shows how. If you will work with an outside team, read how to choose an AI development company first.
Our AI integration page describes how we add AI to systems you already run. You can also send us your scored list. We will tell you which workflow we would start with, and which we would not touch.
Common questions
Which departments usually have the best workflows for AI?
Teams that handle large volumes of documents and messages, such as customer support, finance operations, sales administration, human resources and compliance. The department matters less than the shape of the work. Look for frequent, patterned tasks on text where results can be checked.
Do we need to clean up our data first?
You need the data for the chosen workflow to be reachable and reasonably consistent. You do not need a company-wide data project before starting. Often the pilot itself shows which data problems matter, and that is cheaper than fixing everything in advance.
How is this different from ordinary automation?
Ordinary automation follows fixed rules on structured data. It is predictable and cannot deal with variation. AI handles unstructured input such as emails and documents, and tolerates variation, at the price of occasional errors. The strongest systems combine both, with AI reading and interpreting and ordinary software doing the exact steps.
What if our best candidate involves sensitive data?
It can still be suitable, with more care. Options include masking personal details before processing, choosing providers and terms that do not retain or train on your data, and running models in your own environment. Involve whoever is responsible for privacy and security from the start.
How many workflows should we start with?
One. A single workflow with a clear owner and clear measures teaches your organization how to run such projects. Starting several at once spreads attention and makes it hard to tell what worked. Add the second when the first is running steadily.
