Skip to content
Sign inRequest access

AI triage

AI in uMarker reads a report’s text and proposes the tedious parts of triage: a short title, a summary, a label and a priority, a likely duplicate, a ticket body, a reply draft. It never writes anything by itself: a person accepts, edits or dismisses every proposal.

Three switches, and all three must say yes

Section titled “Three switches, and all three must say yes”

Nothing reaches a model unless every one of these agrees. They are independent on purpose, and none of them substitutes for another.

1. The platform

An operator can switch the whole feature (or one individual capability) off for everybody, and that switch is read before anything else. There also has to be a model to run on — the provider an operator configured, or your own key. With neither, nothing runs.

2. Your account

You have to turn AI on for your account. It is off by default, it lives on its own screen, and the change is recorded as a privacy decision rather than as an ordinary settings edit.

3. Your plan

AI is a Professional (249 ⃁) and up capability, with a monthly credit allowance: 3,000 credits on Professional, 12,000 on Agency. Free and Starter do not include it.

When it is on, four things run on each incoming report without anybody asking, because they are the ones that make an inbox readable at a glance:

What What it does Credits
Label & priority Files the report under one of your labels and estimates how urgent it is. 1
Title Names the report in a few words, so the board can be scanned. 1
Summary Turns three angry paragraphs into one line you can act on. 2
Duplicate Notices when a new report is the same problem as one already open. 2

A nightly pass also proposes groups: it reads a project’s recent open backlog and suggests which reports are one issue. Where duplicate detection asks “is A the same as B” one pair at a time, clustering proposes the whole group at once, so fourteen reports of one outage arrive as one thing with a merge them all button on the report page, rather than as thirteen separate chances to notice.

These never run by themselves: pressing the button is the consent to spend. From an open report:

  • Translate: between Arabic and English, in both directions, without leaving the report page.
  • Ticket body: the steps, the expected and the actual, ready for the tracker.
  • Reply draft: a reply to the reporter — written in your language, not theirs: English when that is the language you are reading the product in, otherwise your account’s. If the reporter wrote in the other one, translate the draft before you send it.

And from a project: the week’s themes, a changelog draft from what you shipped, and survey questions drawn from what your reports keep repeating.

A proposal appears in a dashed frame, next to the control it would change: the title proposal under the heading it would replace, the label and priority inside the triage card, the duplicate beside the duplicate field. Nothing else in the product is dashed, so “provisional” needs no legend.

  1. Read the proposal. The report has not changed; accepting is what changes it, and the frame says so.

  2. Accept: the value is written onto the report. Accepting a duplicate merges the two reports.

  3. Dismiss: it is gone for good.

Each proposal can be resolved once. If a colleague took or rejected it while your page was open, you are told that in your own language and the screen refreshes rather than quietly disagreeing with the one they are looking at.

You can accept or dismiss from the report page and from the inbox list, a row at a time. The board shows the chip read-only: every pixel of a board card is a drag handle, and a 20-pixel accept target inside one competes with the surface’s own verb.

From a report, three fields and only three: the reporter’s comment, the console lines (last 20, truncated — present only on a Technical-mode report), and the page URL. That is enforced by the type the provider is handed, not by a rule somebody has to remember: a provider adapter has no reachable path to a screenshot, a recording, an email address, device metadata, breadcrumbs or failed requests.

A request is not always one report, though, and what rides alongside it is yours too. Label & priority sends your own label names and your priority list, plus up to six of your team’s recent accept/dismiss decisions as worked examples. A duplicate check sends up to 25 other open reports, each as an id and its first 200 characters — and only ones whose own text has already gone to a model, so a report that has stayed home never travels as somebody else’s comparison. The project capabilities you ask for read up to 60 recent reports, each contributing the same three fields; the nightly grouping pass reads up to 40 from the last 14 days.

One more kind of text leaves under the same consent switch, and it is not a report: when our own support staff draft or summarise a reply in a thread you opened with us, that thread’s subject and last 20 human messages go to a model, on our key and our bill. Because the bill is ours, this is the one path your plan does not gate — your consent switch still does, on any plan, including one with no AI of your own. Attachments, names and email addresses do not go — each message is labelled only “customer” or “support”.

Screenshots are never sent under any configuration. See PII redaction for what is blurred in the visitor’s own browser, long before any of this.

The unit is credits, not “AI actions”, because the calls are not the same size: a duplicate check reads twenty-five candidate reports, a five-word title reads one sentence. Pricing them identically would mean either the title subsidises the comparison, or the plan is sized for the worst case and nobody can afford the cheap ones.

The costs above are the defaults; an operator can re-price a capability or switch it off entirely. Your account’s screen shows this month’s usage broken down by capability, what each one cost, and, for the capabilities where it can be observed at all, how often your team accepted what it proposed. Six of them produce a draft that leaves through the clipboard, where nobody can honestly observe whether it was used; those are marked as not measurable rather than guessed at.

At 80% of the monthly allowance, the people who can act on it are told. Past 100%, the capabilities stop and say so; nothing else about your account is affected.

Instead of ours, you can paste your own provider key: Claude, OpenAI, Gemini, DeepSeek, xAI, Mistral, or anything else speaking a compatible endpoint. It need not be the vendor the platform runs on.

Two things change. Your vendor bills you directly, so there is no monthly ceiling and we stop counting credits. And the key is verified at the moment you paste it. A key that does not work is refused on the screen in front of you, not at 3am inside a queue.

Your key replaces ours; it does not replace the plan. AI is still a Professional-and-up capability, and the key screen asks that first: on Free or Starter it refuses the key and names the plan that unlocks AI, before a single request reaches your vendor — rather than showing a key as connected that nothing would ever use.

Every capability call is still recorded on your usage ledger, because an account paying a vendor directly is the one most entitled to see what was spent on its behalf. The verification call above is the exception: it is made with your key and writes no ledger row.

A capability that cannot run stays visible and explains itself, rather than disappearing and teaching you the product cannot do it. Seven reasons are checked, and they produce six sentences: the operator’s global switch and a single disabled capability read the same on screen.

What you read What it means
Turned off for everyone right now. Either the operator’s global switch, or that one capability disabled. The two are not told apart.
AI is off for this account. Nobody has turned on consent.
You turned this one off. Your own per-capability toggle.
Not included in your plan. Professional and up. On the AI settings screen this one answers for the whole page, with a link to your plan, rather than sitting on each capability.
This month’s credits are used up. The allowance is spent; it resets next month.
No model is set up on this installation. This installation has no provider and you have not brought a key.

A report that predates your consent window is refused separately, with its own sentence: not on record as filed after AI was turned on, so it is never sent to AI.

No AI failure ever blocks ordinary triage. If a call times out, is rate-limited or comes back unusable, the report and the inbox stay exactly as usable as they were.

This part is for whoever runs the uMarker platform, not for tenants. AI providers are configured in one place only: the AI page of the operator console. (There is no AI group under Platform settings any more; older key and model settings were moved into the chain once and are no longer read.)

The page shows the providers as an ordered chain, in the order they are tried. uMarker asks the first working provider; if it fails or is busy, it moves to the next. Each row shows the provider, its plan (free, trial credit, your own server, paid — with a warning on plans that may train on what they receive), its status in words (working, resting until a time, needs a key, failing and why), its model, and today’s usage against its limits. From the row you can move it up or down, switch it on or off, test, edit or remove it.

Each provider has one model, used for every AI feature. If you want, an optional, collapsed Use a stronger model for long tasks sets a second model for the longer jobs; empty means the same model. A provider can be added once: a vendor already in the chain shows Added in the catalogue until you remove it. (An OpenAI-compatible endpoint can be added again only with a different address.)

  1. Press Add a provider and pick one from the catalogue, grouped Free · Trial credit · Your own server · Paid, each with its country, its limits and a Get a key link.

  2. Paste the API key.

  3. Press Load models. The model list is fetched live from the provider with the catalogue’s default already selected; you can also type a model name.

  4. Press Test, then Add provider.

A plan that may train on the data it receives can be saved and tested, but is marked Test only and the chain never sends it customer text.