AI Auto-Learning (Smart Autocomplete)
Filling out large forms repeatedly can be a productivity bottleneck. To reduce it, Flowi Agentic learns from your own company's history and suggests field values in Human Tasks (User Tasks).
How it works
Every time someone in your company completes a task, the values chosen in each field are recorded in the process history. Daily, the platform revisits that history and counts, field by field, what was decided in comparable situations — taking into account what was already filled in the rest of the form. For example: when the category is "IT", the chosen vendor was "ACME" in 34 out of 38 cases.
There is no guessing in that count: these are real decisions, made by people in your company, in the same process and at the same step.
The confidence score
The confidence of a suggestion is the fraction of comparable cases in which that value was chosen. If "ACME" was the vendor in 34 out of 38 cases with category "IT", the confidence is 89% — and that is exactly the explanation, "34 out of 38 cases", that comes with the suggestion.
Three rules keep that number honest:
- Only decisions made without seeing a suggestion count. When the platform suggests a value and you accept it, that case does not raise the confidence. If it did, the AI would be confirming itself: a wrong guess would look more and more certain just by being accepted.
- A minimum number of cases is required. With fewer than 5 comparable decisions (default value), no pattern is formed and nothing is suggested.
- A field that left the form stops counting. A form changes over time: one field is renamed, another is dropped. Once the published version of the form no longer asks for that field, what was decided in it stops becoming a suggestion and stops being used to narrow the search for comparable cases. The history stays in the process — what changes is that the platform stops comparing by something nobody fills in any more.
This is about form fields, not about everything
Values the process writes outside the screen — whatever an integration or an automated step computes — keep counting as usual. The form has no way to declare that they are gone, so the platform does not assume they are.
Two kinds of suggestion
When you ask for AI suggestions in a task, two different behaviors can happen, depending on the field:
- Pattern-based fill — when a pattern with at least 80% confidence exists (default value), the still empty field is filled with the pattern's value, along with the confidence score and the count of cases supporting it. You can keep the value or change it: the final decision is always yours.
- Previous case — in free-text fields (notes, descriptions), counting occurrences makes no sense: an opinion written differently for every document would appear only once, and its confidence would always be too low. For those fields the platform returns the most frequent previous value in that same context, literally, for you to edit. It comes with no confidence score, is not treated as a learned pattern, and is traceable: it can say how many previous cases it came from.
Neither one is AI-generated text. Both are counts over real decisions, so asking for suggestions consumes no AI quota. A genuinely unique text field, with no similar previous case, simply gets no suggestion.
What you already filled in guides the suggestion
When you ask for suggestions, the platform looks at the form exactly as it stands on screen at that moment — including what you have just typed and not saved yet. Those values are part of the question: they say which situation the task is in, and they are what makes the search for comparable cases more specific. The more you have already filled in, the more targeted the answer.
And what you filled in is not replaced. Only empty fields are filled automatically. When a pattern points to a value different from the one on screen, nothing is overwritten: the divergence shows up in the Suggestions to review list, together with the suggested value and the count of cases supporting it, and it only enters the form if you click to use it.
None of this is recorded on its own. An accepted suggestion goes into the form as if you had typed it, and reaches the process when you save or complete the task — if you leave the screen without saving, it does not stay.
Why the AI sometimes stays silent
Every so often (1 in every 10 tasks, by default), the platform deliberately withholds its suggestions, even when it has a confident pattern for the field.
The reason is simple: to know whether a pattern still holds, the platform needs to observe decisions that were not influenced by any suggestion. If everyone always saw the suggested value, there would be no way left to tell whether it is still what the company would choose on its own — and a pattern that became obsolete would keep looking right forever. That occasional silence is what keeps the confidence score true over time.
If you are testing the feature, test on several tasks
The draw is per task and deterministic: if that task fell into the silence, it will never suggest anything, however many times you click. And the silence removes every suggestion for that task — counted pattern, past case and AI alike.
The message shown is the same one used for "not enough cases", so the screen cannot tell the two apart. Testing on a single task and concluding the feature does not work is the easiest mistake to make here — open three or four.
Privacy
Two different protections act here, and they do not cover the same things.
By the type declared on the form. Fields declared as CPF, CNPJ, CEP, E-mail, Phone and Password are kept out of learning entirely: they form no patterns, are never suggested, and never enter as context to suggest other fields. An excluded field is removed from what the model sees, not masked.
By the field's name, and only when the form does not declare that field — the case of variables the process creates outside the screen. There the platform looks for terms such as nome, endereco, nascimento, rg, idade, senha, password, token, secret and apiKey in the name.
Name and address are not protected by type
There is no "person name" or "address" type in the form builder, and the name check is not consulted for fields the form declares. A nomeCompleto or enderecoEntrega field declared as text is learned normally.
For those, the administrator has to list them under Fields excluded from learning, on the Auto-Learning tab of the process definition. It is the only protection that works.
What the administrator can tune
The numbers quoted above — 80% confidence, one task in ten kept silent — are the starting point, not a fixed rule. They live on the Auto-Learning tab of the process definition, and every process has its own:
| Setting | Default | What it changes |
|---|---|---|
| Time window (months) | 3 | how many months of history the counting sees — and how many it reads: narrowing it here makes the nightly read cheaper, not just the result shorter |
| Confidence threshold | 0.80 | the confidence from which a value is offered. Raising it makes the platform quieter and more accurate; lowering it does the opposite |
| Minimum sample size | 5 | how many cases decided with no suggestion on screen a pattern must gather before it counts. It guards against concluding from two or three repetitions |
| Conditioning fields | 2 | 1 compares one field at a time; 2 also crosses pairs of fields, which gives more specific patterns and demands more history |
| Holdout (%) | 10 | the share of tasks deliberately left without a suggestion, explained in Why the AI sometimes stays silent |
| Fields excluded from learning | empty | the field names kept out, described under Privacy |
A holdout of 0 freezes the confidence
Setting the holdout to 0 is allowed, and the screen says what it costs: at zero every task gets a suggestion, no uninfluenced decision is ever observed again, and the confidence of the patterns already learned stops being updated — it freezes at the score it has today. A pattern that goes stale will keep looking right.
Where the learning comes from
- Learning is per company: your company's patterns never appear to another.
- Patterns are recomputed daily over a recent window of the history, so changes in how you work show up in the suggestions within a few days.
- History that already exists counts. The counting reads every task completed inside the configured window, with no need for the process to have been prepared for it. A process with history can get suggestions on the very first nightly run; a new process has to accumulate cases.
- Not every setting waits for the night. What only the nightly run rebuilds is the set of patterns: which ones exist, over what context, at what confidence — so the time window, the minimum sample, the conditioning fields and the removal of a field from the form only show up in the result after the next run. Turning learning on or off, changing the confidence threshold, changing the holdout and adding a field to the excluded list all take effect on the very next request for suggestions: they filter what reaches the screen, not what was counted. In practice, taking a field out of learning has immediate effect; loosening the minimum sample does not.