Can you use our documents and past conversations?

We first review what information is suitable and permitted for this purpose. Depending on the task, retrieval can supply relevant source material at the moment the agent needs it — that is not the same as training or fine-tuning a model on your data. Fine-tuning is a separate option, not a requirement for every project.

Can an agent make mistakes?

Yes. We define limits, test representative cases and agree where human review is needed. No model or workflow makes every answer correct.

Will information stay on our own infrastructure?

Hosting and model-provider choices affect where information is processed. We document that data flow and agree the deployment approach for the project.

Can it send messages or update our CRM?

That can be part of the scope when the integration supports it. We agree permissions, approval requirements and error handling before enabling those actions. For CRM work specifically, see HubSpot and CRM.

What is the difference between an agent and an automation?

A rule-based automation follows a fixed sequence and is usually the better choice when the steps do not vary. An agent is worth the extra cost when the work requires interpreting text or choosing between a small set of actions.

How long does it take to build an agent?

It depends on how many systems the agent touches, what it is allowed to change, and how much evaluation the task needs. We agree a delivery plan after the scope is defined.

AI agents

AI agent development for tasks your business needs done.

We build AI agents that use your company’s information and connect to the tools involved in a specific job. Start with a defined task — handling a support request, preparing information, or updating a record — and decide where a person should review the result.

Discuss your project See our support-agent project

Start with a free 30-minute call to define the first scope.

Agent · Lead intake

illustrative example

Example

Checked the inbound enquirystep 1

Looked up the company detailsstep 2

Booked the callstep 3

Updating the CRM recordstep 4

Pricing question → passed to a personhandoff

Illustrative example, not client dataone step routed to a person

Agent or chatbot

Two different purchased scopes.

Both use the same underlying models, and a chatbot can take actions too. The difference is what you are buying: a conversation on your website, or a task carried out across your systems.

—A website chatbot

—Answers questions in a conversation, on your site or a messaging channel
—Usually scoped to the information it can retrieve
—Can also trigger actions, when that is part of the agreed scope

An AI agent

Owns a defined task and the steps that task takes
Reads from and writes to the systems agreed in the scope
Escalates to a person at the points you define

If you mainly need answers on your website, start with an AI chatbot. If the work is a repeatable sequence of rules, a rule-based automation is often the simpler and cheaper option.

Inside the agent

It runs a loop: perceive, reason, act, observe.

The agent reads the situation, decides the next step, takes an action in the systems it has been given access to, checks the result, and repeats until the task is finished or handed to a person.

Repeats until the task is finished

Perceive

Reads the incoming request and the current context.

Reason

Decides the next step and which tool to use.

Act · call a tool

Runs a real action in your systems.

CRMCalendarHelpdesk

Observe

Checks the result, then loops or finishes the task.

Hand off to a person

Any step can pass the task to a person, with the context, when it falls outside the agreed scope.

Possible scopes

Three example scopes, from trigger to outcome.

The same loop, applied to three kinds of task. These are illustrative scopes to discuss, not features of a delivered project.

TriggerAgent step · your toolsDecisionOutcome

Sales agent

Inbound lead → booked call

Web formLinkedInInbound email

Qualify & score

Enrich contact

Route to owner

HotWarmNurture

Call booked

Sequence

Newsletter

Support agent

Ticket → resolution

EmailWhatsAppLive chat

Understand intent

Fetch context

Draft resolution

AutoAssistEscalate

Reply sent

Suggestion

Human handoff

Ops agent

Trigger → systems synced

ScheduleWebhookNew record

Orchestrate

Process data

Sync & notify

RealtimeBatchedOn-demand

Records synced

Daily report

Team pinged

In each scope, the handoff points are agreed before the build: the agent passes the task to a person, with the context, wherever you decide it should.

Start with a defined responsibility

Give the agent one job to own.

These are possible scopes, not a claim that every feature exists in the support-agent project shown below.

Support requests

Retrieve relevant answers from approved information, prepare a response, and hand the conversation to a person when it needs judgment.

Knowledge search

Help staff find answers across business documents, with references they can check and access controls appropriate to the information.

Workflow assistance

Extract information, classify a request or prepare an update for another system. Approvals can be required before the agent sends a message or changes a record.

What we define before building

Agreed before the build starts.

Six decisions that set the boundaries of the agent, its access and its running costs.

Step 01

The task

What the agent owns, and what it must leave to people

Step 02

The access

The documents and systems it can reach

Step 03

The approvals

The actions that require a person to confirm

Step 04

The failure paths

Incomplete information, errors and unavailable tools

Step 05

The evaluation

Representative examples for checking answers and actions

We also agree how usage costs, logs and future updates will be managed.

Connected to the work

An agent is useful when it fits the process around it.

Review the integrations

We look at what access the surrounding tools actually expose, and what that allows an agent to read, propose or change.

Agree the boundary

Some tasks need a simple rule-based automation. Others benefit from AI interpreting text or choosing between a small set of actions. We separate the two before building.

Build and test the scope

The agreed task, the approval points and the failure paths are built and checked against representative examples.

Run it and review

We review real runs against the agreed examples, adjust the scope, and document what the agent does and does not handle.

Related work

An AI support agent connected to Chatwoot.

Our support-agent project connects Chatwoot with an organization’s support knowledge. The case study explains the information sources, the conversation flow and the human handoff.

Explore the support-agent build

Chatwootanswering inside their live inbox.

3 yearsof past support conversations used as source material.

Retrievalover that history — not a fine-tuned model.

Handoffto a person, with the conversation context.

Built with the tools you already run

The model and the integrations are chosen per project, based on the task and where the information is allowed to be processed.

ClaudeGPT-4HubSpotTwilioIntercomWhatsAppMakeSupabaseand others, where the API access exists

FAQ

Which task would you like an agent to handle?

Bring a real example, the information it needs, and the tools involved. We’ll discuss a practical first scope.

Discuss your projectNo obligation.