Case Study
Ultra-Soft.co.uk
An AI Sales Agent That Can't Make Things Up
We built Ultra Soft, a UK water treatment company, a custom AI agent that runs their entire catalogue in live chat: sizing softeners, routing to the right filters, quoting live prices, and capturing commercial leads. Its defining feature is that every price, product code, stock level, and link comes from a real catalogue lookup, enforced by code, not by prompt instructions.
The Challenge
Generic AI chat invents prices, recommends discontinued products, and confirms policies that do not exist
A deeply technical catalogue of 550+ products: softeners, filtration, reverse osmosis, UV, salt, resin, and spare parts
Correct sizing is an engineering question: household occupancy, pipe size, boiler type, and local water hardness
Prices, stock, and sale offers change in WooCommerce daily, so any answer written by hand goes stale
Commercial enquiries from hotels and factories arrived in the same chat as small salt orders, with no routing
Zero tolerance for a wrong price or a fabricated product: one bad answer costs a sale and the client’s trust
Our Solution
A full-catalogue sales agent covering every category: 550+ products, sized and matched to each customer’s situation
Live WooCommerce pricing, stock, and sale prices overlaid on every answer, with the saving called out
Deterministic sizing logic in tested Python: occupancy limits, install-kit selection, contaminant-to-filter matrix
Commercial lead capture for hotels, laundrettes, and factories, straight into a pipeline with an admin view
Human handoff inside Zoho SalesIQ, the live-chat tool their team already uses, with the full transcript carried over
A second AI from a different model family reviews every answer before it reaches the customer
01 / How It Works
Facts It Can't Fake
The architecture is the product. Three layers stand between the model and the customer, so a wrong price or an invented product simply cannot reach the chat.
Layer 01
Facts come from tools, not memory
The conversational model can only get prices, SKUs, stock, specs, and links by calling catalogue tools during that same turn. It never sees the full catalogue and never states a fact it did not fetch. Grounded in tool output, or it does not ship.
Layer 02
Business rules live in tested code
Every rule was ported from prompt text into deterministic Python with unit tests pinning each clause, plus a plain-English spec the client can read 1:1. Prose rules degrade as catalogues grow. Tested code does not.
Layer 03
A different model audits every answer
A Claude validator from a different model family checks each response before the customer sees it: every price must trace to a tool result, every link must be on an allowlist of real URLs, and prices a customer asserts get flagged, not accepted. A different family means no shared blind spots.
Under the hood
- Conversational AI
- OpenAI GPT-4o in a tool-calling agent loop (model swappable by config)
- Response validation
- Anthropic Claude reviewer (cross-model-family) plus server-side SKU, price, and URL checks
- Backend
- Python / Flask, PostgreSQL
- Commerce
- WooCommerce (CSV bulk sync + REST live pricing), live sitemap URL resolution
- Chat surface
- Zoho SalesIQ webhook bot with native operator handoff, plus a standalone web demo
- Testing
- pytest (1,035 tests), rule-derived oracle gate, live scenario batches, adversarial audits
02 / The Proof
Proof, Not Promises
Every number here is backed by the repo's own test suite and reports. Nothing is a claim we cannot run on demand.
1,035
automated tests passing across rules, adapters, guardrails, and regression pins
100%
on an 89-case accuracy gate built from the client’s own business rules, run on every change
144
customer sizing profiles verified end to end against the client’s own business rules
550+
products with prices, stock, and links pulled live from the catalogue on every answer, never typed by hand
136
live conversation scenarios run end to end, with 32 fixes shipped from what they surfaced
0
broken product links, verified against the live site and pinned so it cannot regress
03 / Admin Dashboard
Command Centre
Custom-built admin dashboard for managing 1,049 products, tracking 2,896 AI conversations, and monitoring agent performance.
Overview Dashboard
Real-time stats: 1,049 products synced, 2,896 AI conversations handled.
Product Management
Full inventory control with SKUs, pricing, and category management.
Conversations
Review AI interactions with detailed product recommendations.
04 / The Engineering
Built to Be Doubted
Trust in production is not a prompt instruction. It is architecture, automated exams, and adversarial pressure applied relentlessly before launch.
A rival model checks its work
Every answer is audited by a second AI from a different model family before the customer sees it. Prices must trace to a live catalogue lookup, links must be on an allowlist of real URLs, and anything unverified is blocked, not shipped. Two model families means no shared blind spots.
An exam on every change
We turned the client’s business rules into an 89-case accuracy exam that runs on every single code change, and the agent needs 100% to ship. Underneath it, 1,035 automated tests pin every rule, guardrail, and past fix so nothing quietly regresses.
Red-teamed before launch
We attacked our own agent with adversarial audits and 136 live scenarios built to make it lie: invented policies, fake discounts, wrong prices asserted with total confidence. Thirty-two fixes shipped before a single customer ever saw it.
05 / Safe Delivery
Shipped Without Risking Production
The client's live operation never paused while we built. Deployment discipline is part of the product.
Isolated environments
Everything is built in its own isolated environment with separate routes, database tables, and ports. Nothing touches the client’s live operation until it has earned the right to.
Gates before go-live
New behaviour has to pass hundreds of pinned test cases before cutover. If a change alters a single answer, the suite catches it before a customer ever could.
Additive, never destructive
Proven data and logic are never edited in place. Changes land alongside what already works, so there is always a known-good state to stand on.
Reversible cutover
Go-live is a documented two-step plan with a rollback that takes minutes: one environment variable flips the system back to its previous state.
06 / Interactive Demo
Try UltraAgent Yourself
Ask about household size, water hardness, or a specific product. This is the real agent, not a video: every price, product, and link in its answers is fetched live from the catalogue as you chat.
Every price and link fetched live from the catalogue · validated before display
07 / AI Search Visibility
Getting Ultra Soft Cited in AI Search
Beyond the AI agent, we're positioning Ultra Soft's content so ChatGPT, Google AI Overviews, and Copilot recommend them when users ask about water softeners.
What we're doing
Why it matters for Ultra Soft
When someone asks ChatGPT "What's the best water softener for a 4-person household?", Ultra Soft should be the brand that gets cited. That's a customer with high purchase intent, ready to buy.
With 550+ products and deep domain expertise already powering UltraAgent, the knowledge base is there. We structure it so AI platforms can find, trust, and reference it.
08 / Any Industry
We Build This for Any Industry
The pattern, not the plumbing, is the product. Any business with a catalogue, price list, or service menu has the same failure mode: generic AI chat will misquote it. The fix is architectural, and it travels.
Insurance
Eligibility rules turned into an automated exam the agent has to pass on every change.
Med-spa
Treatment matching grounded in a real menu, never the model’s imagination.
Industrial parts
Deep catalogues where a wrong part number means a costly return.
Property lettings
Live availability and pricing pulled from the system of record, never guessed.
Whatever your source of truth is today, an ERP, a legacy database, or a hand-tuned spreadsheet, we connect to it, overlay live data, and turn your business rules into an automated exam the agent has to pass on every single change. That is what makes it safe to put in front of your customers.