It is ten past eight. A ward sister is on the corridor with a phone in her hand, and a head of department is in the car park before the staff meeting. Both have the same thing waiting: leave requests from their team, two or three of them, each one a login and a few clicks away in the attendance system. By lunchtime they will have been asked twice whether the leave is sorted.
With an AI assistant connected to AMS, the attendance and leave tool in Orbiz, that morning is a short conversation: list what is waiting, read each request, say yes or no. This post walks through it and then draws the lines carefully: what the assistant can read, the one thing it can change, what it cannot touch, why it must never decide for you, and what the record shows afterwards. We wrote about the wider set of attendance questions in ask AI about employee attendance; this one is only about approvals.
A manager’s morning with an assistant connected to AMS
The conversation starts with a read. The manager types “Show me the leave requests waiting on my approval.” and the assistant returns the requests that are waiting on them, as AMS holds them: who asked, which leave type, the dates, how many days, and the reason they gave. On a ward the question is usually narrower: “Which leave requests from the ward team are waiting on me for next week?” In a school, before the autumn term: “Show me the staff leave requests waiting on my approval before term starts.”
Then the manager decides, one request at a time, in their own words: “Approve Bilal’s leave request for the 14th.” The assistant calls the one AMS action that records a decision, and AMS stores the outcome as that manager’s, with the time. A rejection in AMS carries a reason, and the applicant is emailed it, so say why in the same sentence rather than leaving the assistant to invent one.
That is the whole routine. Three or four messages, no screens opened, and the applicant gets the same notification they would have got from the approvals page. The rest of this post is about why the routine is safe to use, and where it stops.
What the assistant can read before you decide
AMS gives a connected assistant 10 actions, 7 that read and 3 that change something. Every one of them runs as the person who connected the assistant, with that person’s AMS permissions. For an approver, the reads that matter on a Monday morning are three:
- The team’s leave requests. The list of requests waiting on you, with the detail AMS recorded for each: leave type, dates, days, the reason given, and its status. This is the read behind every prompt above.
- A team attendance summary. Who was in and how the week looks, for someone whose AMS role already shows their team’s attendance. Useful context before a decision, though it is a summary of attendance, not a leave calendar.
- Your own leave balance. “How much annual leave do I have left?” reads your balance. There is no assistant action that reads a team member’s remaining balance; that figure is on their request in AMS, where you check it before an approval that would take someone past their entitlement.
An approver who does not approve leave in AMS is not offered the team actions at all. The assistant lists only the tools the person’s role can use, so a member who asks for the ward’s requests without that role gets their own requests and nothing else.
What it can do: one action, as you
There is exactly one assistant action that changes a request on the approver’s side: approve or reject leave. It takes a decision and, for a rejection, the note the applicant will read. It does not take a shortcut round AMS: the same permission check that guards the Approve button on the web runs for the assistant, and the request ends up in the same state, decided by the same person, at the time it was decided.
That is the sentence to hold on to: the assistant acts as you, with your permissions, nothing more. A connection belongs to one member in one organisation. The consent page names both, and says the assistant can only see and do what you can on the web, in the areas you allow. If an admin removes your approval permission at nine o’clock, the assistant loses it at nine o’clock too.
Two things have to be true before any of this works. When you connect the assistant it sends you to an Orbiz consent page, where access is listed tool by tool with reading separate from changing; tick changes for AMS, or it can list requests but not decide them. And the assistant side has its own gate. Claude’s help pages say members of a Team or Enterprise plan connect to a custom connector individually, so Claude “can only access tools and data that the individual user has access to”, and the Free plan is limited to one connector. OpenAI’s guide says write actions in ChatGPT require confirmation by default and tells you to review the tool input before confirming. Both read as a good description of how this should work, and both are worth checking against your own plan. The steps for each assistant are written out in Settings → Connected AI apps inside Orbiz, and the general method is on the Orbiz MCP page.
What it cannot do
This is the list people skim and should not. An assistant connected to AMS has no action for:
- Clocking in or out, and the attendance timer
- Viewing or taking screenshots
- Attendance corrections and shift settings
- A team member’s leave balance, as above
- The leave policy, the entitlements per type, or the holiday calendar
- Deleting a request, or anything else: nothing in any Orbiz tool can be deleted through an assistant
None of these is a setting someone has to remember to switch off. The actions do not exist in the catalogue, so a clumsy request has nothing to trigger. The full list for AMS, with the read and write split, is on the AMS assistant page.
Why an assistant should never auto-approve leave
It is tempting to go one step further: “approve anything under three days if the balance covers it”. Do not. Three reasons, in order of how often they bite.
- The decision is yours in the record. AMS stores who decided and when. If the assistant decided on a rule you set in a chat, the record still says you, and the ward that is short on Thursday will ask you why. A rule in a conversation is not a policy; it is a habit nobody else can see.
- The assistant cannot see cover. AMS has no “who else is off that week” view, and no assistant action for one. A three-day request from one nurse is fine; the same request when two colleagues are already off is the gap found on the day. That check is a person looking at the roster, and it stays that way.
- Language is loose. “The 14th” is a date in two months. “Bilal” may be two people. The right response to that is to read the list, name the person and the dates, and confirm each decision, not to let the assistant guess and carry on.
The assistant vendors say the same in their own words. OpenAI’s guide notes that ChatGPT can remember an approve choice for a tool for the rest of a conversation, and advises allowing that only for an application you trust to make further writes without asking. For leave, do not allow it: the confirmation prompt is the point. Claude’s help pages ask you to review the permissions a connector requests and to limit scopes where you can, which in Orbiz means ticking reading for AMS first and adding changes only when you want to decide.
What the record shows afterwards
Two records are written. The AMS one is the leave request itself: approved or rejected, by whom, at what time, with the note on a rejection. It is the same record the approvals page writes, so reports, the member’s balance and the timesheet that payroll reads all see the decision at once. The Orbiz one is the connector’s own log. An activity log shows each action with the assistant, the outcome — done, not allowed or failed — and the time. You see your own; admins see the organisation’s and can filter by member.
Two more controls sit above both records. An admin can turn AI assistants off for everyone. That stops every assistant and blocks new connections; existing ones work again when the switch goes back on. And any member can disconnect an assistant from Settings → Connected AI apps, and it loses access at once. Admins can disconnect anyone’s. If a connected laptop goes missing on a Friday, the connection is gone before the weekend.
Honest limits: one policy, one calendar, one step
AMS is deliberately simple about leave, and an assistant inherits that simplicity rather than hiding it.
- One leave policy per organisation, with fixed types: sick, casual, annual and probation leave, each with an entitlement set once for everyone, plus unpaid leave if the organisation allows it. A school that wants different allowances for teaching and support staff runs each group as its own organisation; per-group policies are on the roadmap, not in the product.
- One holiday calendar per organisation. Closure days and public holidays are entered once and count as non-working days in every balance. There is no second calendar for a second site.
- One decision per request. A request waits on one approver and is approved or rejected once. There is no second signature, no chain of approvers and no assistant action that could add one.
- No overlap view. Neither AMS nor the assistant will tell you who else on the ward is off that week. If cover matters, and in healthcare and education it always does, the roster is still the place you check.
If your approvals need two signatures, or a per-department policy, an assistant will not add either, and neither will AMS today. How AMS fits a clinic’s roster and a school’s term calendar is on the AMS for healthcare and AMS for education pages.
Prompts to copy
Every line below maps to an AMS action the assistant actually has. Swap in your own names and dates.
- “Show me the leave requests waiting on my approval.”
- “Which leave requests from the ward team are waiting on me for next week?”
- “Show me the staff leave requests waiting on my approval before term starts.”
- “Approve Bilal’s leave request for the 14th.”
- “How much annual leave do I have left?”
More, for every tool, are in the prompt library, each marked as reading or changing something. The phrasing rules are short: name the person and the dates, read before you change, and say the leave type the way AMS names it.
Sources, read 8 October 2026: OpenAI, “Add custom MCP server” (developer documentation); Claude Help Center, “Get started with custom connectors using remote MCP”. What an assistant can do in AMS is taken from the Orbiz MCP catalogue as published on the AMS assistant page.
Built for your industry
