Skip to content
Web Katalyst Book 15 min15 min

AI agents & automation

Custom AI agents, connected to your business.

We design agents around your tasks, your rules and the tools you already use. They prepare the work: a reply to a customer, a quote, a selection of prospects, a sales file. Your team keeps the decision.

Illustration Scattered requests become an up-to-date record and a ready reply, which your team approves.

Beyond ChatGPT

An assistant answers a question. A custom agent moves the work forward.

Used on their own, ChatGPT or Claude know neither your customers, nor your prices, nor your tools: you have to explain everything again, then copy the answer somewhere else. We connect them to your business so they prepare a complete piece of work your team only has to review.

A general-purpose assistant, on its own

  • Does not know your data
  • Cannot act in your tools
  • Does not apply your rules
  • Keeps no record of the work

An agent designed for your business

  • Connected to your toolsIt reads and writes where you work: email, CRM, store, spreadsheet, invoicing.

  • Fed with your contextIt relies on your customers, orders, prices and documents, not on generalities.

  • Bound by your rulesMargins, lead times, terms, tone of reply: it applies what you have decided.

  • Able to actIt creates the record, prepares the quote, files the document, schedules the follow-up.

  • Tracked over timeEvery job is logged: what was done, from which sources, what is waiting.

  • Approved by youNothing important goes out without the agreement of the person in charge.

Four examples

From the simplest to the most advanced: what an agent can prepare for you.

Four situations we come across in small and mid-sized businesses. For each one: where it starts, what we connect, what the agent does and what your team gets.

Simple

Customer service: a reply prepared with the customer’s file at hand

1Where it starts

A customer asks where their order is or how to return it. The answer exists, but it is spread across the store, the carrier and your terms of sale.

2What we connect

  • Online store
  • Carrier tracking
  • Terms of sale
  • Previous conversations

3What the agent does

  1. Finds the order and its tracking
  2. Checks what your terms provide
  3. Writes the reply in your tone

Sensitive case or missing information: it hands over to an advisor, with the summary.

4What your team gets

Reply ready to review

Hello Mrs Lenoir, your parcel was handed to the carrier yesterday; delivery is expected on Thursday. If the size is not right, returns are free for 30 days.

  • Order no. 4821
  • Carrier tracking
  • Returns policy

Your advisors review and send; they keep their time for the cases that need judgement.

Before starting: written terms of sale and access to order tracking.

  • Sources cited
  • Rules visible
  • Human approval

Intermediate

Quotes: a price prepared from the request, your price list and your rules

1Where it starts

Every request means rereading the need, finding the prices, checking the constraints and working out the margin before you can reply. Requests pile up, and some go unanswered.

2What we connect

  • Incoming request
  • Price list
  • Constraints and options
  • Margin rules

3What the agent does

  1. Reads the request and spots what is missing
  2. Applies your prices and constraints
  3. Checks the margin against your rule
  4. Suggests a price or a range

Request outside the framework: it flags it and asks the salesperson.

4What your team gets

Quote preparedTo approve

  • 12 workstations€6,200
  • Two glass-walled rooms€2,400
  • Delivery and fitting€900

Suggested range€9,200 to €9,800 excl. tax

Margin within your rule

Requested lead time: to confirm with the workshop

Your salesperson starts from a costed, justified quote to approve or correct before sending.

Before starting: prices and margin rules put in writing. That is where we begin.

  • Sources cited
  • Rules visible
  • Human approval

Intermediate

Prospecting: a selection of companies to contact, with the reason for each choice

1Where it starts

Finding the right companies takes hours of research. For lack of time, the same messages go to everyone, and the best accounts are not handled first.

2What we connect

  • Your target criteria
  • Company directories
  • Public signals
  • Your CRM

3What the agent does

  1. Looks for matching companies
  2. Sets aside poor fits and says why
  3. Completes the record of selected companies
  4. Drafts a tailored first message

No message goes out without the sales manager’s approval.

4What your team gets

This week’s selection

  • Cabinet OrvaneMoving in January
  • Groupe SéraneHiring 15 people
  • Néologis TransportOutside the area

2 records and 2 drafts to approve

Your team starts its week with accounts sorted, explained and ready to be contacted.

Before starting: a precisely described target, and the list of what to exclude.

  • Sources cited
  • Rules visible
  • Human approval

Advanced

Tenders: a response file assembled and checked point by point

1Where it starts

Responding means matching specifications that run to dozens of pages with your suppliers’ offers, your documents and your costs, then having each part approved. One overlooked requirement can be enough to rule the bid out.

2What we connect

  • Specifications
  • Supplier offers
  • References and certificates
  • Cost table
  • People to consult

3What the agent does

  1. Lists the requirements in the specifications
  2. Matches each one with your documents and the offers received
  3. Flags what is missing or inconsistent
  4. Prepares the costing and the draft response

Each part is submitted to its owner: technical, finance, management.

4What your team gets

Response file

18 of 21 requirements covered

  • Up-to-date insurance certificateto provide
  • Lot 2: supplier B’s offer selectedchecked
  • Technicalapproved
  • Financeapproved
  • Managementpending

At any time you know what is covered, what is missing and who still has to approve.

Before starting: up-to-date company documents and a named owner for each part.

  • Sources cited
  • Rules visible
  • Human approval

Illustrative examples with fictional data: they describe solutions to be scoped with you, not results obtained for a client. An agent can make mistakes; that is why it shows its sources and leaves the decision to you.

Under the hood

Behind every agent, a workflow connected to your tools.

An agent is not a black box: it is a sequence of steps we build with you. Here is a simplified version, for an incoming request.

What triggers it

  • Website form

  • Mailbox

  • Store or CRM

AI agentreads, searches, prepares

  • AI modelClaude or GPT, depending on the task

  • Memoryyour customers, your exchanges

  • Toolsprices, calendar, documents

Your rulesmargins, lead times, cases to hand over

Expected case and complete information?

Yes

Draft preparedreply, quote or record

Human approvalthe person in charge reviews

No

Handed to an advisorwith the case summary

Sending and updatereply sent, CRM up to date

Logevery step is recorded

Simplified diagram. We build these workflows in n8n or an equivalent tool, depending on what you already use.

The technologies we use

  • n8nWorkflow automation

  • ClaudeAI model by Anthropic

  • ChatGPTAI model by OpenAI

We choose the tool according to the task, your constraints and your data. These brands belong to their owners; Web Katalyst is not an official partner of any of them.

Sources, rules, human control

An agent that knows the context, and knows when to hand over.

To prepare a relevant reply, the agent relies on what your business already knows about the customer. When information is missing or the decision is yours, it hands the case to an advisor, with everything it has found.

What it looks at

  • The customercontact details, purchase history
  • Their ordercontents, status, delivery tracking
  • Previous conversationsemails, tickets, notes
  • Your terms and ruleslead times, returns, goodwill gestures allowed

Every reply shows what it is based on

Sources consulted
Order no. 10482 Carrier tracking Returns policy
Rule applied
Return accepted within the period set in your terms
Approval point
Refund: an advisor’s agreement is required
Illustrative example

When it calls in an advisor

  • Information is missingThe order cannot be found or the request is incomplete.
  • The situation is ambiguousTwo rules apply, or the request falls outside the expected cases.
  • A refund needs authorisationThe agent prepares the case; an advisor decides.
  • The subject is sensitiveComplaint, dispute, unhappy customer: a person replies.

An agent can make mistakes. That is why every reply states its sources and the rules applied, and why your advisors keep the approval of decisions.

Our method

Four commitments, from the first task to follow-up.

We start with a single task: you see the result on your own requests before deciding to go further.

  1. Choose a task and its expected result together

    We agree with you on one specific task: what the agent must produce, for whom, and how you will judge it a success.

    You receiveA one-page scoping note: the task, the data, the rules, the success criterion.

  2. Test on your real cases

    We run the agent on real requests, reviewed by your team, before anything goes live.

    You receiveThe results obtained on your own requests, reviewed with you.

  3. Connect your tools and define approvals

    We connect the agent to your tools and agree with you on who approves what.

    You receiveThe agent in service, and the written list of who approves what.

  4. Monitor and improve

    We follow usage with you and adjust the agent over time.

    You receiveA regular review, and the record of adjustments made.

First conversation

Which task would you like to hand over to an agent?

Tell us about a task that keeps coming back. In fifteen minutes, we will tell you whether an agent can prepare it, and how we would start.