Skip to content

A chatbot that answers
from your own documentation.

AI chatbot development services for support and sales: assistants grounded in the docs, policies and past tickets you already have. Every answer shows its source, and anything uncertain goes to a human instead of a guess. Built on your helpdesk, with a working pilot in 14 days.

Pilot scoped in 48 hours · Human handover built in

  • Hello Peter
  • Gruber Logistics
  • Delhivery
  • Thomson Reuters
  • Qatar Airways
  • Grundfos
  • Save
  • BERD
  • Yale University
  • Kuwait Police
  • Dubai Police
  • Panasonic
  • Infosys
  • Kia
  • Hitachi
  • Orange Business Services
In one answer

PixelCrayons builds support and sales chatbots grounded in your own documentation, on the stack you already run. Answers cite their sources, and low-confidence questions escalate to a human by design. Guardrails and an evaluation suite ship before launch, not after. The first working pilot lands in 14 days on your real docs. Direct for brands, white-label for agencies.

The operating record

Judge the record,
not the adjectives.

Outcomes tied to real engagements, not averages.

21 yrs
Years in continuous delivery
100+
Agency partnerships
2,500+
Projects delivered
30+
Countries served
2+ yrs
Average partner retention
14 days
NDA to first deliverable
340%
Revenue growth · 7 months
Client outcome: eCommerce
+127%
Organic traffic · 5 months
Client outcome: SaaS
85%
Faster delivery · zero churn
Client outcome: via agency partner
Clutch — 4.8 / 5 ratingGoodFirms — 4.7 / 5 rating
Google Partner
Meta Business Partner
Shopify Partner
Where the work happens
ShopifyWooCommerceMagentoWordPressWebflowKlaviyoGoogle AdsMeta AdsGA4Next.js
  • Answers grounded in your documentation
  • Sources shown on every answer
  • Human handover built into every assistant
  • Working pilot in 14 days
  • Proposals itemised in 48 hours
  • You own the code, prompts and pipelines
  • Evaluation before launch, logs after
What it can actually answer

Every demo answers tier one.
You are buying tier three.

Chatbot questions come in four tiers, and each one costs something different to serve. The tier a vendor demos is almost never the tier that decides whether the project was worth doing. This is the ladder we scope against before anyone writes a prompt.

  1. 01

    Answerable from your documents

    Assistant

    “What is your refund window?”

    Requires

    An index of the docs you already have

    No systems touched, nothing to authenticate. This is the tier in every demo, and it is genuinely useful. It is just not the hard part.

  2. Step change: everything below needs your systems
    02

    Needs this customer’s own record

    Assistant

    “When does my plan renew?”

    Requires

    Identity verification, then read access to the system of record

    The real step change. You now need a login boundary the assistant respects, and a rule for what it does when identity is uncertain.

  3. 03

    Changes something

    Assistant, on confirmation

    “Cancel my renewal before it bills.”

    Requires

    Write access, an audit trail, and a reversal path

    Write actions get an explicit confirmation step and a logged record: what was changed, by whom and on whose instruction. A wrong write is not a bad answer. It is a support ticket with money attached.

  4. 04

    Judgement, exception or an upset customer

    Human

    “This is the third time this has happened.”

    Requires

    A person, with the conversation already summarised

    Handed over deliberately, not as a failure. The measure of this tier is how much context arrives with the customer: nobody should have to repeat themselves to a human after explaining it to a bot.

The jump in difficulty is not between tiers three and four. It is between one and two: the moment the assistant needs to know who it is talking to.

Illustrative tiers: question examples are generic, not client data

In every build

What a chatbot build
actually includes.

01

Grounded in your documentation

The chatbot answers from your own help centre, policies, manuals and past tickets. It finds the relevant passages first and writes the answer only from them; the industry calls this retrieval-augmented generation, or RAG, and it is the same layer under our knowledge systems. Every answer links its source. No source, no answer.

The foundation
02

Escalation to a human, by design

Confidence thresholds decide when the assistant stops. Refunds outside policy, legal questions, angry customers and anything ambiguous route to your team with the full conversation and the relevant document attached. A chatbot that never hands over isn't confident. It's reckless.

Every build
03

Guardrails & evaluation before launch

Before a single customer sees it, the assistant runs against an evaluation suite built from your real tickets: grounding checks, tone checks, resistance to users talking it out of its rules (prompt injection) and the topics it must refuse. You see the scores and sign off on launch. The suite re-runs whenever your docs change.

Before launch
04

On your stack, in your channels

The assistant plugs into the helpdesk, CRM and website you already run, and the channels your customers actually use, rather than forcing a platform migration. We stay model-agnostic too. The model underneath is chosen on fit, cost and data residency, and structured to be swappable without rework.

Included
05

Handover, logs and ownership

Every conversation is logged and auditable, so you can see exactly what was said, what source it used and why it escalated. At handover you get the runbook, admin training and outright ownership of the code, prompts and pipelines. Updating the knowledge base is your team's job, not a change request to us.

At handover
Pilot in 14 days

From your documentation
to a working assistant.

01
Days 1 to 3

Discovery & doc audit

We map the queue the chatbot will face, audit the documentation it must answer from, and agree in writing what a correct answer (and a mandatory escalation) looks like.

02
Week 1

Pilot scoped & priced

A written scope lands: the questions in bounds, the sources it grounds in, the escalation rules and a demo date. You approve the itemised price before code ships.

03
Day 14

Working chatbot, your docs

By day 14 you're chatting with a working pilot grounded in your own documentation: answering, citing sources and escalating on the rules you set.

04
Weeks 3 to 6

Evaluate, harden, launch

The pilot runs against real tickets from your queue, guardrails tighten, the helpdesk integration goes live and your team learns the runbook. Then it meets customers.

Inside Prism

Your engagement, week to week,
in one workspace.

Your AI build runs in a Prism workspace you log into: requests, approvals, the task list and the weekly review, with each decision and its expected result written down.

  • 01

    Requests and approvals

    One queue, one owner, one due date.

  • 02

    Weekly review

    Each decision recorded with its expected result.

  • 03

    Actions checked against outcome

    What we did and what happened, side by side.

How we report it, in Prism

The eval scorecard,
before the pilot goes near a customer.

What you receive in month one

AI work is reported against an evaluation set built before any model is chosen. In the first month you receive the task definition with the inputs and acceptable outputs written down, the eval set of real cases with expected answers, and the scorecard showing accuracy, refusal behaviour and cost per run. The go/no-go checklist names what must hold before the pilot touches live traffic.

  • 01Task definition: inputs, acceptable outputs, and the cases the system must decline
  • 02Eval set: real cases with expected answers, reviewed by your team
  • 03Scorecard: accuracy, refusal behaviour, latency and cost per run
  • 04Go/no-go checklist: the thresholds the pilot must meet before launch
  • 05Failure log: each wrong answer, its cause, and the fix applied

Related workB2B SaaS: from invisible to answer-engine cited.SaaS · UK · A different discipline, so it is linked here rather than presented as proof of this one.

The rest of the AI bench

A chatbot is one door in.

The retrieval layer under a chatbot is a knowledge system. The escalation logic is workflow automation. Start wherever the pain is, or start at the hub. To add chatbot developers to your own team instead, hire them by the month.

Questions

Frequently
asked.

It depends on the queue it faces and the systems it touches, which is why we don't quote from a rate card. Every engagement opens with a scoped pilot: one queue, your own documentation, a working assistant in 14 days, priced as a small, defined piece of work. The proposal you receive within 48 hours itemises what ships, including the questions in scope, the integrations wired, and the guardrails and evaluation included. You compare deliverables, not day rates. Scaling to more channels or queues is priced only after the pilot has earned it.

Any language model can hallucinate, and no vendor can promise zero. Our job is to make errors rare, visible and safe. The assistant answers only from your own documentation and shows the source under each answer. When retrieval comes back thin or confidence is low, it says so and hands over to a human rather than improvising. Before launch it runs against an evaluation suite built from your real tickets, and after launch every conversation is logged so you can audit exactly what was said and why. Honest answer: not zero errors, but engineered, measured and contained ones.

No, and we'd rather tell you that now than in a post-mortem. What a grounded chatbot genuinely does is clear the repetitive layer of the queue: the questions your documentation already answers, asked hundreds of times over. What it hands back to your team is everything that needs judgement (exceptions, complaints, negotiations, anything ambiguous) with full context attached. That's a better use of skilled people than copy-pasting the same reply. Teams change shape around that. What changes is where human attention goes, not headcount, and we won't put an invented percentage on it, because your queue isn't an average.

It grounds in the documentation you give it (help centre, policies, manuals, past tickets) by finding the relevant passages first (the RAG layer described above). Your data is used to answer your customers and for nothing else. Work runs under NDA as standard, we use enterprise API terms under which providers don't train on your data, and where you need it we deploy inside your own cloud so nothing leaves your infrastructure. Retention, residency and access rules are agreed in the scope document upfront, and at handover you own the code, the prompts and the pipeline outright.

The ones you already run: that's the point. Builds integrate with your existing helpdesk and CRM, your website, and messaging channels your customers actually use, rather than forcing a new platform on your team. We're deliberately model-agnostic too. The model underneath is chosen in discovery on fit, cost and data residency, and structured to be swapped without rework as better or cheaper options ship. If an off-the-shelf tool would serve you better than custom chatbot development, discovery ends with us saying so in writing.

Yes: the same grounding discipline applies, pointed at revenue. A sales assistant answers product and pricing questions from your real documentation, qualifies visitors with questions you approve, and books meetings or routes hot leads to a human. Handing a ready-to-buy prospect to your team fast is the one job it must never fumble. What we don't do is let it improvise discounts, invent comparisons with competitors, or promise features that don't exist. The guardrails and evaluation cover the sales scripts just as strictly as the support ones.

Put a grounded chatbot
on your busiest queue.

Scope a pilot: one queue, your own documentation, a working assistant in 14 days, with sources shown, escalation built in, and guardrails tested before a customer ever sees it. Tell us which tier your questions sit in; the pilot is scoped to that tier.

48-hour proposal · NDA standard · You own what we build

Last updated

May we run analytics (Google Analytics via Google Tag Manager) to see which pages are useful? Nothing loads unless you accept, and declining means no analytics script runs at all. No advertising cookies either way. Cookie policy · Privacy policy