Case studies

What a first piece of work has looked like.

One named customer and the two builds delivered for them, described without invented figures. Each covers what the business needed, what was configured, the challenges that were scoped and how the outputs were checked.

Case study 01 / Internal AI Assistant

Own knowledge,
answered and checked.

A UK supplier of packaging films and machinery. Built and refined with the team, and its layout is reproduced below with neutral branding and sample content.

What the business needed

Answers to everyday questions were spread across sales and stock exports from the ERP system, product documents, supplier correspondence and the knowledge in people’s heads. Questions such as which accounts had gone quiet, how a product family was moving across several codes, or what a document said about a product meant opening several files and asking a colleague.

What was configured

  • A company-specific assistant with a chat interface, chat history and search, and a selector for which datasets an answer should use.
  • A training area with a knowledge base, a product base, a competitor base and an email base, plus structured business data from the sales and stock exports.
  • Saved team rules and corrections, and team responsibilities for who owns which knowledge and who reviews outputs.
  • Answers that name the document, rows or rule they relied on, so anyone on the team can check them against the source.

Challenges that were scoped

  • Loading complete rows from exports and confirming the scope of each export before reading it.
  • Choosing the correct document and period when several reports could plausibly answer the same question.
  • Reconciling totals in the assistant’s output to the source report.
  • Grouping equivalent film codes only with the team’s sign-off on specifications and units.
  • Excluding recent arrivals from slow-stock flags so a fresh delivery was not mistaken for stock that does not move.

How outputs were checked

Staff compared example outputs with the source reports, including row counts, totals and the exclusions above, and corrections were saved as team rules. The reporting use case is shown as an interactive example on the internal assistant page.

The client has agreed to be named: read the full Kempner story. No staff name, figure or file from the project appears on this site; the interface reproduces the layout only.

YOUR COMPANYInternal GPT
  • New chat
  • Search chat
  • Which accounts are late against their usual interval?
  • Please analyse these sales and stock exports for Q3
  • Show me stock where the receipt was in the last 14 days
  • What is the difference between DEMO-F01 and DEMO-F01-B?
  • Do you know the 2026 price list rate for PT-48?
  • Can you give me a list of customers who ordered BX-40 this year?
  • Summarise the supplier correspondence about delivery dates
  • What causes the variance on the Q2 stock report?
Train your agentYOUR COMPANY
Use:Sales history by customer, to 31 August 2026Sales analysis 2026 (Aug 26)Sales analysis 2026 (Jul 26)Sales analysis 2026 (Jun 26)Stock movements Q3 2026Stock movements Q2 2026Price list 2026Product data sheetsSupplier correspondence (export)Team rules

Which accounts are late against their usual ordering interval this week, and what should I check before calling?

Two accounts are later than usual at the 31 August review. Customer A is 84 days on against a usual 32; Customer C is 21 days on against a usual 7. Customer B is within its interval.

Before calling, team rule 14 says to check for a partial return in the last 30 days and for an open price query. Customer A has a credit note on 12 August for two damaged units; the original order stays in the interval calculation.

  • SourceSales history by customer, to 31 August 2026 · rows 118 to 142
  • SourceSales analysis 2026 (Aug 26) · credit notes, row 7
  • RuleTeam rule 14: partial returns, saved by Sales on 2 September
Ask about your business data, documents or rules

Case study 02 / Website AI Assistant

Product questions,
answered from agreed content.

A UK supplier of packaging products. Built and demonstrated around the supplier’s own product information, with product questions and enquiry capture as the focus.

What the business needed

Visitors with practical product questions could not always find the right entry, compare similar codes, or tell when a question needed a person. Those questions arrived by phone and email and were answered from information already on the website.

What was configured

  • A website assistant configured around the supplier’s agreed product information, answering in plain words and citing the entry it relied on.
  • Handover to a person when the content did not cover a question, with the option for the visitor to leave contact details.
  • Capture of the questions asked, so the team could review what visitors wanted and where the catalogue fell short.

Challenges that were scoped

  • Which information the assistant may use, and which needs a person: suitability for a site, pricing and stock stayed with the team.
  • Keeping answers to what the content states, without turning a plausible match into a recommendation.
  • Making the handover carry the visitor’s context rather than a bare “contact us”.

Status and what is shown here

Built and demonstrated. The guided demo reproduces the style with an invented catalogue and preset questions, so nothing on this site uses the supplier’s real content.

Website assistantSample catalogue
VisitorWhich boxes do you have for loads around 20 kg, and do you stock the tape to go with them?
AssistantFor loads around 20 kg the catalogue lists BX-40 double-wall cartons (400 by 300 by 300 mm, tested to 20 kg) and BX-45 for taller items. The matching tape is PT-48 packing tape, 48 mm wide. Would you like the carton sizes side by side?
Sources: sample entries BX-40, BX-45, PT-48
ENQ-S006Cartons and tape · Answered from catalogueAnonymous visitor
Referred to a person when the catalogue is silent.

Sales Sidekick

Not a case study yet. Available to scope.

The Sales Sidekick has not been delivered as a client project, so it is not described as one. The fictional preparation sheet shows the intended output, and the pricing page explains how a build would be quoted.

A 15-minute conversation

Worth a
conversation?

If either project sounds like your situation, a 15-minute conversation with sample data is the quickest way to see whether one agreed piece of work would help.

Or email directly.