Calls and the AI assistant

Agent training

Teach the AI agent how your pharmacy wants calls handled, see exactly what it works from, and correct it from real calls. For owners and managers (who change the training) and the whole team (who can suggest corrections from call transcripts).

What it does

The agent answers from a small, fixed set of sources — nothing else. Agent training shows you every one of them and lets you add to the part that is yours: the training guidance.

Training comes in three layers, applied top down:

1. Platform — set by PharmaLineAI for every pharmacy: safety rules and the standard test scenarios (prescriptions, stock, services, appointments, Pharmacy First, emergencies and the "false positive" traps where a routine-sounding call is really clinical).
2. Group-wide — set by a pharmacy group's head office for all its branches.
3. This branch — set by the branch (or a standalone pharmacy) for its own line.

A lower layer can add entries, replace an inherited one with its own version, or switch one off — unless the layer above locked it. Safety rules, such as emergency handling, are locked for everyone.

Each entry is one of five kinds:

Changes are used from the next call.

What one pharmacy teaches its agent can help every pharmacy's. PharmaLineAI can share a pharmacy's corrections, exemplary replies, scenarios and rules with every pharmacy — rewritten to make them general, without naming the pharmacy. Owners can turn this off (see below).

Where to find it

• Navigation: Training in the header.
• URL: /dashboard/training (tabs: ?tab=sources, guidance, review, transcripts)
• A call's transcript: Calls → click the caller's name → /dashboard/calls/<call>

Step by step

See what the agent works from

1. Open Training. The first tab, What the agent uses, lists each source: the built-in safety rules, your pharmacy details, services, prescription records, greeting and transfers, and the training guidance.
2. Each card says how the agent gets it, what it currently holds, and where it is Maintained in, with an Open button to go there. Anything missing is flagged (for example "Not set — the agent will say it doesn't know").
3. What the agent is told shows the training guidance exactly as the agent receives it. Copy copies it.
4. When the agent gets something wrong tells you where to fix each kind of mistake.

Add an entry (owners and managers)

1. Training → Guidance → Add entry.
2. Choose the Kind. The fields change to suit it; each says how the agent uses it.
3. Give it a Title, and optionally a Category and Safety priority (when a call has several needs, the higher priority wins).
4. For a scenario, fill in What the caller says, What they mean, What the agent should do, and optionally What the agent should say or do, What it must not do and Don't mistake it for (false positive).
5. At group level you can tick Lock for the layers below.
6. Click Add.

Replace or switch off an inherited entry for your branch

1. Guidance → find the entry (its Layer shows Platform or your group's name).
2. Open the row menu (⋯) → Replace for _your branch_ to write your own version, or Switch off for _your branch_ to stop the agent using it here.
3. The original stays as it is for everyone else, now marked Replaced below or Switched off below. To undo a switch-off, open its row menu → Switch back on.
4. A locked entry offers neither option.

Review a call and mark replies

1. Calls → click the caller's name, or Training → Transcripts → click a call.
2. Read the Transcript line by line. Where the call was recorded, press play — or click a line's time to jump there. The line being spoken is highlighted.
3. Under each agent reply, Based on: shows what it had to go on for that reply — for example Pharmacy details & services (loaded at the start of the call), which the agent has from the start of every call, or Prescription records (after identity check) — each linked to where that data is kept. Otherwise it answered from its instructions alone.
4. Click Exemplary to hold the reply up as the standard, or Correct it to record what was wrong and What it should have said.
5. Check the masked text (dates, postcodes, phone and NHS numbers are masked automatically) and remove any names. Choose where it Applies to, then Save — or Send for review if you're a team member.

Approve suggestions (owners and managers)

1. Training → Review (the tab shows how many are waiting).
2. For each suggestion: Approve to put it in force, Edit to tidy the wording first, or Discard.

Import a transcript (owners and managers)

1. Training → Transcripts → Import transcript.
2. Paste the transcript or choose a .txt, .vtt or .srt file. Start each line with who is speaking — Caller:, Agent: or Staff: (Pharmacist, Receptionist and Patient work too). Timestamps like [00:12] are kept.
3. The preview shows how many lines were read and any speaker it didn't recognise.
4. Optionally add a Label, When the call was, Answered by, the Recording (MP3, WAV, M4A, OGG or WebM, up to 4 MB) and Notes. Click Import — the call opens, ready to mark.

Import scenarios from a spreadsheet (owners and managers)

1. Save the sheet as CSV. Useful columns: Kind, Title, Category, Priority, Caller says, Intent, Action, Ideal response, Do not, Incorrect. Names are matched loosely.
2. Guidance → Import spreadsheet → choose the file or paste it. The preview says how many entries are ready and which rows will be skipped and why.
3. Click Import.

Test the agent against scenarios (owners and managers)

1. Guidance → a scenario's row menu → Run against the agent, or Run all scenarios.
2. The agent answers the caller's words using your guidance and your real opening hours and services — no patient records are read and nothing is logged as a call.
3. The result is Passed, Failed or Check reply (only a person can judge the wording). Open the entry to read the reply and what it looked up.

Choose whether PharmaLineAI may share what you teach (owner)

1. Training → Guidance. The box at the top reads Help every pharmacy's agent learn from ours.
2. Tick it to allow sharing, untick it to stop. It is on unless you turn it off.
3. When it is on, someone at PharmaLineAI may adapt your corrections, exemplary replies, scenarios and rules into the training every pharmacy starts from. They rewrite each one first and take out names. Your pharmacy is never named, earlier lines of a call are never copied, and your Approved information is never shared.
4. If something of yours is shared, your own entry says Shared with every pharmacy and stays in force for you, word for word. The shared copy shows in the Platform layer as Replaced below — by yours.

A group's owner decides for the group and all of its branches. Other team members see whether sharing is on, but only the owner can change it. Turning it off stops anything more being shared; what was already shared stays, without your name.

Record calls so you can listen back (owner)

1. Settings → Phone line → tick Record AI calls for review → Save phone line.
2. Make sure your greeting tells callers the call may be recorded.

What each screen shows

• Layer cards — Platform → your group → your branch, how many entries each has in force, and whether you can edit it (You can edit / Read-only).
• Showing the agent for — group owners can look at the group-wide set or any branch.
• Guidance — every entry from every layer. Status: In force, Replaced below, Switched off below, Blocked by a lock, Switch-off, Left out (too long), Awaiting review, Archived. Current hides drafts and archived entries; Everything shows them. Click a title for the full entry, where it came from, and its last run.
• Help every pharmacy's agent learn from ours (top of Guidance, your own pharmacy) — whether PharmaLineAI may adapt your training for every pharmacy. Owners can tick or untick it.
• Shared with every pharmacy — on an entry of yours that PharmaLineAI has shared. Came from: Shared from a pharmacy marks a platform entry adapted from some pharmacy's lesson.
• Review — suggestions from the team, waiting for approval.
• Transcripts — calls with a transcript: Live call or Imported, who answered, how many replies have been marked, and whether there's a recording.
• A call's page — the call's details, the transcript with times, the recording, the feedback already given on each line, and What the agent had been told on this call (the guidance version it ran with, even if it has changed since).

Tips and good to know

• The agent says nothing beyond its sources. If it didn't know something it should, add Approved information; if it said something wrong about your hours or services, fix Settings or Services rather than adding training.
• Replying "we don't do that" when a service simply isn't listed is a common mistake — correct it, and keep your Services list complete.
• Transcripts and recordings are deleted after the retention period (30 days by default). What you mark is kept, with personal details masked.
• If the guidance grows past its size limit, the least important entries are Left out (too long) — archive what you no longer need.
• The agent can't book, cancel, take messages or check stock yet; it puts those callers through.
• If the agent tells a caller it is putting them through, it does — even if it forgot to press the button, so to speak. An offer ("Would you like me to put you through?") waits for the caller's answer.
• Training can never unlock clinical advice or patient details: the identity check is enforced by the system, whatever the training says.

Troubleshooting

Related guides

• Calls & the AI agent
• Phone line
• Services
• Settings
• Lists and tables