Techbanq

AI use case · support

A support agent is measured by what it is allowed to do

Deflection answers a question and leaves the customer to act. Resolution does the thing — resets the password, checks the shipment, issues the credit — against your live systems. The whole design is the list of actions the agent may take alone and what happens when a request runs past it.

Serving Saudi Arabia from our Riyadh office and our engineering hubs in Lahore and Karachi

Operators monitoring process screens in an industrial control room
  1. 01

    Market

    Saudi Arabia

  2. 02

    Written for

    COO, Head of CX

  3. 03

    Decision

    Automate tier-one support resolution, not just deflection

  4. 04

    Languages

    EN · AR

Enterprise delivery credentials

  • ISO

    ISO/IEC 27001

    ISO/IEC 27001 certified

Certifications & registrations

The action surface

What the agent may touch, and where a person takes over

Every request enters as a question and leaves as an action or a handover. The middle column is the permission set: the changes the agent is allowed to make alone. Nothing outside it is attempted.

A funnel: account lookups, reversible changes and bounded goodwill converge into a written permission set, and out to a recorded resolution or a handover to a person.

  • Account and order lookups

    Status, history and entitlement, read from the system of record rather than a stale copy.

  • Reversible changes

    Password resets, address edits, delivery reschedules — actions a person could undo in one step.

  • Bounded goodwill

    A credit or refund up to a fixed limit and count, logged and attributable, above which a person decides.

What we do

The permission set

A written list, per action, of what the agent may change alone — with the ceiling, the confirmation and the log format fixed before a single live write is enabled.

  • Resolved and recorded

    The action taken, the reason, and the before-and-after written where the next agent can see them.

  • Handed over with context

    Anything past the boundary reaches a person with the account, the request and what was already tried attached.

A support agent with a write action and no ceiling is not a faster support desk; it is an unbounded one.

Why deflection stalls

The number that moves is repeat contact

A deflection bot lowers first-contact volume and raises the thing nobody puts on the dashboard: the customer who was answered, still could not act, and came back angrier. Resolution is measured on the second contact that never happens, not the first one that was cheaper.

That is why the first work of an engagement is not the model but the systems: which of them expose an action the agent can call safely, which need a confirmation step wrapped around them, and which are so fragile that the agent only ever reads from them and routes the rest to a person.

The requests that decide it

Where a resolution agent earns or loses its place

A deployment is judged on the awkward requests, not the clean ones, and each has an answer that is a permission decision rather than a model setting.

01

The customer asks for a refund above the ceiling

A partial resolution now and a person for the rest, rather than a flat refusal.

The agent explains the limit, does what it can within it, and hands the remainder to a person with the case assembled.

    02

    Two accounts look like the same person

    One verification step, because a change made to the wrong account is the failure that ends the pilot.

    No write until identity is confirmed against a second factor the account already holds.

      03

      The system of record is down mid-request

      An honest hold, not a confirmation for something that did not happen.

      The agent stops rather than guesses, tells the customer plainly, and logs the intent so a person can complete it.

        04

        The request is a complaint, not a task

        A person and a record, which is what a complaint is owed.

        Acknowledged, logged and escalated to the standing procedure, with no attempt to close it automatically.

          Our role

          How the action list is built and proven

          1. 01

            The permission set, written first

            Before any model work, we agree the action list with your team: each action, its ceiling, its confirmation, and whether it is reversible. The list is the contract the deployment is held to.

          2. 02

            A sandbox before a live write

            The agent runs against a copy of your systems and a week of real tickets, so its resolution and its mistakes are both visible before it touches a live account.

          3. 03

            The audit line

            Every change the agent makes is logged with the reason and the before-and-after, so a supervisor can review a day of actions the way they would review a new hire’s.

          Which actions are safe to automate, and the ceilings on each, stay your decision; we build the boundary you set, prove it in a sandbox, and make the handover carry the context every time it fires.

          Common questions

          What buyers ask

          How is a resolution agent different from the chatbot we already tried?+

          The chatbot answered; a resolution agent acts. The difference is a set of authenticated actions against your systems of record — a reset that actually resets, a refund that actually posts — each scoped with a ceiling and a log. A bot that only produces text will always hand the real work back to the customer, which is why deflection numbers look good while repeat-contact numbers do not move.

          What stops it from making a change we cannot undo?+

          The permission set. Irreversible actions are either kept off the agent entirely or wrapped in a confirmation step and a ceiling, agreed with your team before any live write is enabled. In the sandbox phase we run a week of real tickets specifically to surface the actions that should never have been on the list, and remove them before it touches a live account.

          Does it work for Arabic support, not just English?+

          Yes, and Arabic is where the engineering is, not the demo. Customers move between Gulf dialects and Modern Standard Arabic, switch to English for product terms, and describe the same problem several ways. The evaluation set is built from your real Arabic tickets rather than translated English ones, because a resolution rate measured on translated text is measuring a language nobody is writing in.

          Map it to your support stack

          A support agent running against a copy of your systems with a fixed action list, resolving a week of real tickets in a sandbox before it is given a single live write.

          Start the conversation