Skip to main content
GullySystem
AI for Customer Support · Solutions

Walk customers through multi-step problems. Hand over before they lose patience.

Some customer requests take a dozen messages, such as moving a delivery slot or working through a fault checklist. A conversational assistant keeps the thread, works out what the customer wants, and follows the steps your team wrote. Rules that you set bring a person in when the topic, the customer or the mood of the chat calls for one.

What it is

Some questions do not fit one reply. Say a customer wants to change a booking. Then the assistant must find the booking, read the rule, offer the dates that are free and confirm. Each step depends on the one before it.

A conversational assistant keeps that thread. It asks one question at a time, reads the answers in ordinary words, and fills in the form behind the scenes. Chat, messaging and phone calls can share the same steps, and a customer who switches channel is not asked to begin again. Staff confirm anything that costs money. Every conversation can be read afterwards.

Where it goes wrong today

The problems this is meant to ease.

Multi-step requests eat agent time

A rebooking needs a lookup, a rule check, a date offer and a confirmation. An agent repeats that sequence all day for dozens of customers, and no step needs her judgement.

Customers repeat themselves

A caller explains the problem on the website chat, again to the first agent, and a third time after the transfer. Each retelling costs patience, and the details shift a little with every version.

Channels keep separate histories

What a customer said on messaging is invisible to the person answering the phone. The CRM holds only what somebody remembered to type, so continuity depends on memory and goodwill.

Escalation happens too late

An upset customer stays in an automated flow or a long queue. The complaint arrives with a human only after the third attempt, when the account is already at risk.

Steps are followed unevenly

Two agents handle the same return in different ways. Checks get skipped on a busy afternoon. Without a written sequence, quality varies by who is on shift and how tired they are.

Nobody can see where chats stall

A lead cannot tell which step makes customers give up. The transcripts exist but nobody has time to read them, so a confusing question sits in the flow for months.

Inside the module

What AI conversational AI does.

Guided steps in ordinary language

It follows the sequence your own team uses for a return, a reschedule or a fault check. It asks one thing at a time and moves on when the step is complete. Going back one step is allowed.

Memory inside the conversation

Someone who gave an order number at the start is not asked for it again near the end. If the request changes halfway, the assistant follows the new one and keeps the details already given.

Understanding what is being asked

Customers do not use your menu words. They write that the thing arrived broken, or that they were charged twice. Their wording is mapped to a request type, and the assistant asks when it is unsure.

Website chat, messaging and voice

One set of steps can serve the website widget, a messaging number and, through your phone system, calls. Voice needs a telephony link and speech recognition tested on your callers, so it comes last.

Customer records and knowledge base

With a link to your CRM or customer database, the assistant reads the account that matches the person writing. It also quotes your knowledge base, so a policy is stated as your team wrote it.

Identity checks before details

Account details appear only after the questions you choose have been answered correctly. A failed check sends the chat to a person. There is no second guess.

Escalation and a short summary

Hand over when the topic is a complaint, when the customer asks twice for a person, or when the account is marked key. The agent sees three lines on top, with the full chat below.

Proposals that wait for approval

Where a step changes something, such as a refund or a new delivery date, the assistant drafts the change into the ticket. Staff approve it or send it back with a note.

Where it gets used

What this looks like in a working week.

A courier moving a delivery slot

The customer gives a tracking number, the assistant finds the booking, offers the open slots and drafts the change. A dispatcher approves anything inside the cut-off window, and the customer gets the confirmation.

A clinic rescheduling a visit

Patients ask to move an appointment. It checks the diary the clinic exposes, offers free times and records the change request. Reception confirms and anything clinical goes to a nurse.

A machinery dealer’s fault checklist

Someone reports a stopped compressor. The checks the service team wrote are walked through one at a time. The model and serial number are collected, and a job card is opened for the workshop.

A retailer handling returns

Order, reason and photographs of damage are requested, then compiled into a return request. The store manager decides on the refund, and the customer sees the status afterwards.

A school office answering parents

Parents ask about fee dues, forms and holiday dates across several messages. Replies come from the notices and from the dues the system holds. Disputes are referred to the accounts clerk.

Who touches it

The people who open this every day.

Support agent

Picks up escalated chats with the summary on screen. Corrects the assistant’s mistakes and flags the exchange, so the same slip is not repeated for the next customer.

Support team lead

Writes the step sequences and escalation rules. Reads a sample of finished chats each week and edits any sequence where customers got stuck or confused.

Operations manager

Decides which processes the assistant may open at all. A booking change might be allowed. A change to a signed contract usually is not, and stays with staff.

How it is set up

From first conversation to daily use.

  1. Discovery of the journeys

    We list the requests that customers repeat and that your team already handles in a fixed order. Tickets, call notes and chat history show which journeys have steps clear enough to write down.

  2. Preparing records and steps

    Each chosen journey is written as numbered steps with its rules. The CRM fields, order records and knowledge base articles the assistant will read are checked for gaps and duplicate customers.

  3. A trial on real conversations

    It plays through a few hundred past chats and call transcripts, with nothing sent to customers. We compare where it would have asked, finished or stopped against what your agents did.

  4. Agents and leads review the misses

    Agents and the team lead read the stalled and wrongly handled cases. They correct steps, add escalation rules in plain words, and mark journeys the assistant must never open.

  5. Going live on one journey

    One journey opens on one channel, with a visible route to a person. Staff watch the first conversations, and the assistant can be paused from the lead’s screen at any moment.

  6. Reviewing after a few weeks

    The lead reads finished chats, escalations and ratings together. Steps that confuse customers are rewritten, and further journeys or channels are added only once the first is settled.

Measuring it

How you would know it is working.

Record where you stand before it goes live, so the change is measured against your own figures.

  • Task completion is the share of started journeys that reach their last step, compared with the same journeys handled by agents before go-live.
  • Customer satisfaction comes from the rating or survey you already send after a contact. Results from the months before go-live give the starting figure.
  • Conversation containment is the share of chats that end without a person, counted for the same journeys in the helpdesk before the assistant starts.
  • Escalation reasons are tallied by rule, so you see which topic hands over most. The first month of tallies serves as the baseline.
  • Repeat contacts within a week are counted from ticket history before go-live and followed after it.
Used responsibly

Where a person stays in charge.

Money and contracts need a decision-maker

Facts are gathered and a proposal is drafted. Someone with the authority approves a refund, a credit or a change to a signed agreement. Nothing is paid or promised first.

Secure handling of customer data

It reads only the CRM fields you allow and asks for identity checks first. Logs are kept for the period your adviser sets. Data protection duties vary by country. The software supplies no legal opinion.

It follows written steps, not judgement

Journeys that depend on a manager’s call are left with staff. It is weakest on vague, emotional or unusual cases, and the escalation rules exist to catch them early.

Voice and languages are tested before use

Accents, noisy lines and each added language are tried on your own recordings first. Where it fails twice to understand, the call or chat passes to a person.

Customers know what they are talking to

The first line says it is an automated assistant. Each country sets its own duties on disclosure and call recording. Your lawyer reviews the wording and the consent text.

How it fits

It is one part of a system, not an island.

A module earns its place by what it passes to the next one. These are the connections we set up most often.

  • Ticket management and helpdesk
  • Order and booking records
  • Knowledge base
  • WhatsApp Business and web chat
  • Customer database
Common questions

Questions owners ask about AI conversational AI.

How does this differ from a plain FAQ bot?

An FAQ bot answers one question at a time from an article. This one holds a thread across a dozen messages and follows a process your team defined in advance. That takes more setup, since each process must be written step by step.

Where should a business begin?

Choose a journey your team already performs the same way every time, such as a rebooking or a standard return. Processes that depend on a manager’s judgement come later, and some never will.

May it decide a refund or credit?

No. It collects the facts and drafts the proposal for a person with the authority to decide. Nothing is paid or promised first.

What does it need from our systems?

Read access to the customer, order or booking records it must quote, normally through an integration your software supports. We check what yours allows before any work starts. Without a link it can still answer from articles.

How accurate is it, and what are mistakes like?

It follows clear steps reliably and handles vague requests less well. Slips usually show as a wrong request type or a missed detail. Staff read sampled chats weekly and correct the steps behind each one.

What happens if the customer writes in another language?

That depends on the languages chosen at setup and tested with your own staff. Outside those, the chat goes straight to a person. Each language is checked before use.

Is customer data private?

Only the fields you allow are read, identity checks come first, and named staff can see the logs. Retention and consent vary by country. Your own adviser settles them, not the software.

When is the effort not worth it?

Skip it when your cases are rare, mostly one-off, or too varied to write as steps. The quote follows discovery and reflects how many journeys, channels and links are involved. Trained agents serve small teams better.

Talk to us

Tell us how you handle AI conversational AI today.

A spreadsheet, a register or another system: say which, and we will tell you plainly what is worth changing.

  • No obligation
  • A reply within one business day
  • Your details stay private

Start with your requirements

Seven short steps that tell us what you need. An engineer reads them before getting in touch.

Share your requirements