Your team answers the same 40 questions every week.
Stop paying them to.
I turn your policies, product docs and four years of
resolved tickets into instant, cited answers —
inside Gorgias, Zendesk or Front, where your
agents already work.
No new tool to learn, no chatbot in front of your
customers unless you want one.
We agree the accuracy bar before I
start.
If the pilot misses it, you don't pay for it.
That's in writing, not a sales line.
Gorgias · Ticket #4821Live example
SMSarah M.2 min ago
—
Suggested reply
—
—
SendEditNot right
Agent stays in control — nothing sends
itself
Built for the stack you already runGorgiasZendeskFrontIntercomHelp ScoutShopify
Sound familiar?
The answer already exists. Finding it is the job.
Your policies are written. Your best agent has answered
this exact question 200 times. None of that is available
at the moment someone needs it.
The hunt before the reply
Notion, the shared drive, that pinned Slack
message, and finally asking someone. Four to
twelve minutes per ticket, and most of it isn't
typing.
"I know we have a policy on this. I just don't
know where."
Every agent answers differently
Three agents, three interpretations of the same
returns policy. One of them is quietly costing
you refunds you didn't owe.
"Wait, we give free return shipping on that?
Since when?"
Onboarding takes six weeks
New agents are slow because the knowledge lives
in senior heads. So your best people spend their
day being a search engine for everyone else.
"Just ask Maya, she'll know."
Q4 breaks everything
Volume triples, you hire temps, and the temps
know nothing. Quality drops in exactly the month
it matters most.
"We'll fix the docs after peak." (You won't.)
How it works
Three weeks from your messy docs to answers your team
trusts
You don't reorganise anything first. Messy is the normal
starting condition — sorting it out is part of what
you're paying for.
1
I read everything you have
Policy docs, help centre, product data, Slack
threads, and your last few thousand resolved
tickets — which are the most valuable thing you
own, because they contain answers a human
already approved.
Week 1
2
We agree what "right" looks like
Your lead agent and I write 100 real questions
with the correct answer beside each. This is the
accuracy bar in the guarantee — you set it, and
it's how we both know whether this worked.
Week 1–2
3
It shows up in your helpdesk
Suggested replies appear beside each ticket with
links to the exact source paragraph. Agents
send, edit, or reject — and every rejection
makes the next one better.
Week 3
Plugs into what you already use:
Gorgias
Zendesk
Front
Intercom
Help Scout
Shopify order data
Notion & Google Drive
Results
Three teams, and what actually changed
Including the parts that didn't work. If a case study
has no disappointing number in it, someone edited it.
Nordwell HomeDTC furniture · 14 agents ·
Gorgias
Delivery and assembly questions were eating
a third of the queue
Bulky furniture generates a specific kind of
ticket: where is it, who carries it
upstairs, what happens if it doesn't fit.
The answers lived across a carrier matrix, a
returns policy and about 900 previously
resolved tickets.
We started with 100 test questions written
by their support lead. First measured pass
was 68% — not good enough to ship. Most
failures traced to a carrier terms PDF that
contradicted the help centre. Fixing that
one document moved it to 91%.
The thing I didn't expect was finding out
our own policies disagreed with each other.
We'd have kept answering it three different
ways forever.
— Support Lead, Nordwell Home
6.4 → 2.9Minutes average handle time
91%Accuracy on their own 100-question
test set
2 wksUntil agents used it without being
asked
What didn't work: anything requiring
a live carrier lookup still goes to a human.
That's roughly 12% of delivery tickets and
we chose not to automate it.
Halo SkincareBeauty DTC · 6 agents · Zendesk
Ingredient and sensitivity questions needed
to be right, not fast
Their volume was manageable. The problem was
risk: customers asking whether a product was
safe with a specific condition, allergy or
prescription, and agents improvising from a
product page.
Speed was never the goal here. We built it
to answer strictly from approved copy and to
refuse — visibly — when the question went
beyond it. It now declines about 15% of the
questions it sees, and that number is the
feature.
I care much more that it says "escalate
this" than that it answers fast. It knows
what it doesn't know, which is more than I
could say for our onboarding.
— Founder, Halo Skincare
0Unapproved claims since launch,
audited weekly
15%Of questions deliberately refused
and escalated
-40%Escalations to the founder
What didn't work: the first version
was too cautious and refused 34% of
questions, including easy ones. Took another
week of tuning to find the right line.
Trailhead SupplyOutdoor gear · 9 agents + seasonal
temps · Front
Peak season doubled the team and halved the
quality
Every November they take on six temporary
agents who need to be productive in days,
not weeks. Warranty terms on technical gear
are genuinely complicated, and getting them
wrong is expensive in both directions.
We aimed this squarely at ramp-up rather
than handle time. Temps got the same
suggested answers as senior agents, with the
source paragraph attached so they were
learning while they worked.
Our seasonal hires were answering warranty
questions correctly on day two. Last year
that took a month and a lot of my time.
— Head of CX, Trailhead Supply
6 wks → 4 daysTime to productive for new
agents
88%Accuracy on warranty
questions
3.2kTickets handled during peak with
the same headcount
What didn't work: we tried it on
pre-purchase sales questions too. Accuracy
was 61% and we pulled it. Sales
conversations need context the system
doesn't have.
Pricing
Published, because you shouldn't need a call to learn a
price
Start with the pilot. Almost everyone should — it's
designed so that finding out costs less than guessing.
Knowledge audit
Not sure your docs are in good enough shape?
Find out first.
Before the pilot starts, you and I agree a number —
usually 85% correct on the 100 questions
your team wrote. Not my test set. Yours.
At the end of three weeks we run it together. Hit
the bar, you pay the $6,500. Miss it, you pay
nothing and keep the test set, the audit of your
documentation, and a written explanation of what
blocked it.
I can offer this because measuring first is how
I work anyway
— and if your content can't support good answers,
I'd rather find out in week one than take your money
for a build that disappoints in month four.
Model and hosting costs are billed directly by your
provider, typically $40–$200 a month. I don't mark
them up.
Objections, answered
What CX leads ask me first
Is this a chatbot that talks to my customers?
Not by default, and I'd usually advise
against starting there. It suggests replies
to your agents, who send, edit or reject
them. Your team stays in the loop and your
brand voice stays yours.
Once accuracy is proven on a category over a
few months, some teams let the clearest
question types auto-respond. That's a
decision you make with real data in front of
you, not a thing I ship on day one.
Gorgias and Zendesk already have AI features.
Why pay you?
Use them — they're improving, they're cheap,
and if the built-in feature solves your
problem you shouldn't hire anyone. Honestly,
for a store with a tidy help centre and
simple policies, it often does.
Where teams call me is when the answers live
outside the help centre: in a carrier terms
PDF, a contract, a spreadsheet of SKU
exceptions, four years of resolved tickets,
or in the difference between what the policy
says and what you actually do. Built-in
features read your help centre. I build over
everything you have, and I'll tell you which
situation you're in during the audit —
including if the answer is "you don't need
me."
Our documentation is a mess. Do we need to fix
it first?
No. Every team says this, and it's true of
every team. If your docs were tidy you
probably wouldn't have the problem in the
first place.
Sorting through the mess is a large part of
what the audit and pilot actually do. What
matters isn't whether it's organised — it's
whether the answers exist somewhere. Usually
the biggest find is that two documents
contradict each other and everyone has been
quietly choosing sides.
What if it gives a customer the wrong answer?
Three defences. Every suggestion cites the
exact source paragraph, so the agent can
check it in a second rather than trusting
it. Every answer carries a confidence score,
and low confidence gets flagged for review
instead of suggested. And it's built to
refuse rather than improvise — a visible "I
don't have a source for this" is a correct
answer.
Then there's measurement: you get a weekly
report on accuracy, refusals and rejected
suggestions, so quality is something you
watch rather than assume.
How much work is this for my team?
Roughly six to eight hours total, almost all
from one person — usually your support lead.
Most of it is building the 100-question test
set in week one, which is genuinely useful
on its own regardless of what happens next.
After that I need read access to your docs
and helpdesk, and about thirty minutes a
week for a check-in.
You're one person. What happens if you
disappear?
Everything runs in your infrastructure and
your accounts, with a runbook and the
evaluation harness your team can run without
me. No proprietary layer, no hosted black
box, nothing that switches off if I do.
I'm also honest about capacity: two clients
at a time. If I'm booked, I'll tell you when
I'm free rather than starting badly.
Who you'd be working with
I'm not an agency. I'm the person who writes the
code.
I'm Bahman. For eight years I built backend
systems where being wrong was expensive — an
automated trading platform placing orders on the
Texas power market, a national payment gateway I
helped break out of its monolith, Kubernetes
clusters and data pipelines underneath both.
I've also built e-commerce platforms, including
Shopify integration microservices at Vgang and a
hosted store builder at Pyango — so I know what
a messy product catalogue and an inconsistent
returns policy actually look like from the
inside.
That background is why this offer is shaped the
way it is. Measuring before deploying isn't a
sales gimmick I invented for this page; it's how
you build anything that runs unattended and has
to be trusted.
2024 — 2025Automated energy trading, ERCOT
— pipeline on GCP, health metrics and
alerting
2023 — 2024US equities trading platform —
AWS services on the trading path
2021 — 2022Pyango — hosted e-commerce store
builder
2020 — 2021Zibal — payment gateway,
monolith to microservices
Remote, UTC+3 — overlapping European hours and
US mornings. 76 technical posts on
DEV
·
LinkedIn
Next step
Tell me your three most repeated questions
That's the whole first email. I'll tell you
within a day whether they're the kind a system
can answer well — and if they're not, I'll say
so and explain why.