Human on the loop
The agent takes every conversation. Your people take the ones that need a person.
Medra is not a chatbot with a fallback. It is one operating surface where AI agents and your commercial team work the same threads, with the same context, under the same rules. The agent resolves. A person decides. The customer never notices the seam.
Illustrative figures from a representative deployment. Your baseline sets the real ones.
The control surface
Eight things your team can do that a chatbot cannot let them do.
Each one is a real control in the console, not a setting buried in a dashboard. Together they make the agent something your executives run, not something they compete with.
01PAUSE
One switch. The agent stops sending. It does not stop working.
Any executive can pause the agent on a conversation with a single control. From that moment the thread belongs to a person. The agent keeps reading, keeps summarising and keeps drafting suggestions, but it does not send a word until someone resumes it.
- Scope is yours. Pause one conversation, one account, one campaign or an entire channel.
- Resume by rule or by hand. Set a time, set a condition such as “when the case is closed”, or simply flip it back.
- No silent gaps. If a paused thread receives a message and nobody answers within the SLA, the supervisor is alerted and the agent can send a holding reply you approved in advance.
- Every pause is logged. Who, when, why and for how long. It shows in the audit trail and in performance reporting.
Try the switch. The agent goes quiet, the executive becomes the owner, and the drafts keep coming.
02ESCALATION
Escalation is a rule you write, not a failure the agent has.
You define when a person must enter the conversation. The agent watches every thread against those rules and, when one fires, routes the case to the right inbox with a brief that says what happened, what the customer wants and what the agent would have done next.
- Customer asked for a person34%
- Deal value above threshold26%
- Negative sentiment18%
- Low agent confidence14%
- Compliance keyword8%
The mix tells you where to invest. A rising share of “low confidence” means the agent needs training. A rising share of “asked for a person” usually means a segment wants a relationship, not an answer.
03INBOXES
Inboxes that mirror how your team actually splits the work.
Every conversation lives in exactly one place at a time. Personal inboxes for what is yours, team inboxes for what is shared, and system inboxes for what nobody has picked up yet. Views on top let each executive slice the same queue by channel, priority, account tier or SLA state, and save it.
- Unassigned never hides. It is the first thing a supervisor sees, with the oldest case on top and its SLA clock running.
- Waiting on customer is a real state. Threads park there without counting against anyone, and come back to the top the moment the customer replies.
- Snooze with a promise. Snoozed threads return at the time you set, or earlier if the customer writes first.
- The agent’s own inbox is visible too. Your team can read what the AI is handling right now and step in at any point.
04ASSIGNMENT
Cases find the right person. Nobody has to go looking.
Assignment is a policy per inbox, not a habit per team. Cases can rotate evenly, follow the executive who already owns the account, go to whoever has room, or match on language, territory and skill. Manual pull is always there for the days when the policy is wrong.
- Capacity is respected. Each executive has a ceiling of open cases. When it is reached, the router skips them and the supervisor sees it.
- Ownership is sticky. A returning customer goes back to the person who handled them last, if they are available.
- Availability is real time. Shifts, breaks and out of office states feed the router. Nothing lands on an empty chair.
Even rotation.
Each new case goes to the next available executive in the inbox. Fair by construction, and the simplest policy to explain to a team.
To whoever has room.
The router weighs open cases, pending tasks and recent volume, and sends the case where it will be answered soonest.
Back to the person who knows them.
Returning accounts route to their last owner. The relationship survives the channel switch and the week off.
Match on what the case needs.
Portuguese speakers, enterprise pricing, disputes. Tag executives once and the router does the matching every time.
Your CRM ownership, honoured.
Accounts already assigned in your CRM go to that owner. Medra reads the mapping, it does not replace it.
Take it because you want it.
Any executive can pull from Unassigned or hand a case to a colleague with one action and a note. Both are recorded.
05TASKS
The conversation ends. The commitment does not.
Every promise made in a thread becomes a task attached to that thread, whether the agent made it or a person did. “I will send the revised quote by Thursday” is a task with an owner and a due date the moment it is written. Tasks live next to the conversation, so the context is always one click away.
- The agent creates tasks from what it hears. Call backs, documents to send, approvals to chase. The executive confirms or edits.
- The agent can also complete them. A follow up on the due date, a reminder to the customer, a status check. You decide which tasks it may execute alone.
- Overdue is loud. Late tasks surface in the owner’s inbox, the supervisor’s view and the daily digest.
- Two way sync with your CRM. Tasks and activities are mirrored so nobody maintains two lists.
06PERFORMANCE AND SLA
Measured per inbox, per executive, per minute. Not claimed.
Service levels are defined where the work happens: a first response target and a resolution target per inbox and per priority. The console shows the clock on every conversation, warns before a breach, and reports who met what, so coaching is about numbers rather than impressions.
Supervisors get a live view of every inbox, every clock and every executive’s load. Alerts go out before a breach, not after it, by console, email or WhatsApp. Weekly reports break performance down by executive, inbox, channel and campaign, and compare humans and the agent on the same scale.
07HUMAN TO HUMAN CALLS
Some deals close on a call. Make that call from the thread.
When the customer asks for a person, or the deal size says one is needed, the executive calls straight from the conversation. Over WhatsApp voice or the phone network, from the same screen, with the whole history in front of them. The call is recorded with consent, transcribed, summarised and filed in the thread the moment it ends.
- Click to call, click to schedule. Call now, or let the agent offer the customer three slots from the executive’s calendar and book one.
- The agent listens so the executive does not take notes. Key points, objections, numbers mentioned and commitments made are extracted automatically.
- Every commitment becomes a task. “I will send the contract tonight” is on the executive’s list before they hang up.
- The agent picks the thread back up. Once the call is filed, the executive can hand the conversation back and the AI continues with everything that was said.
- Martín confirms the move to 65 seats on the annual plan, payment terms stay at 30 days.
- Agreed a 12% discount on list. He asked whether a two year term would improve it.
- Wants the contract before Friday so legal can review during the week.
08EVERYTHING AROUND THE HANDOFF
The rest of what a commercial team needs to work beside an agent.
Small controls, deliberately built so that nothing about the human side of the platform feels bolted on.
One thread across channels.
A customer who writes on WhatsApp, calls on Tuesday and emails on Friday is one conversation with one owner. History, notes and tasks travel with it.
Copilot for the executive.
While a person owns the thread, the agent proposes replies in the account’s tone, pulls the data behind every claim and flags anything that contradicts policy. Send as is, edit, or ignore.
Internal notes and mentions.
Talk about the customer without the customer seeing it. Mention a colleague and the thread lands in their inbox with your note on top.
Availability and shifts.
Executives set themselves available, on break or off shift. The router, the SLA clocks and the agent’s holding replies all follow that state.
Audit trail on everything.
Every pause, assignment, edit to an AI draft, template used and call placed is recorded with who and when. Exportable, and ready for compliance reviews.
QA scoring and coaching.
The agent scores a sample of human conversations against your own rubric: tone, accuracy, next step set. Supervisors review the outliers, not the pile.
Human on the loop
Your team, with an agent that knows when to step aside.
We build the AI ourselves and we have run commercial teams ourselves. Bring us one inbox and one week of conversations. We will show you where the agent should stop and a person should start.