Skip to content
New: CogLake, company knowledge and AI agents in one place.See CogLake
AI agents

AI agents for bounded tasks with human control where it matters.

We turn vague AI ideas into specific agents: clear scope, connected tools, measurable outcomes and escalation paths when a decision needs a person.

Workflow: support inbox request
  1. Agent reads the requestTopic and urgency detected
  2. Answer from company knowledgeManual and past tickets as sources
  3. Draft sent for approvalGoodwill case, a person decides
  4. Reply sentTicket updated automatically

How an agent works with us

An agent gets a clearly scoped task, reads the context it needs, uses approved tools and submits its result for approval. You decide what it may do without asking.

Task

New request in the support inbox

AI agent

  1. Read contextCustomer history and manual from CogLake
  2. Use a toolLook up the order status in the ERP
  3. Draft a replyWith sources

Approval

Your team reviews and sends

Done

Reply sent, ticket documented

The agent may on its own

  • Read documents, tickets and inboxes
  • Write drafts and summaries
  • Sort and assign cases
  • Add internal notes

Only with approval

  • Send emails to customers
  • Change data in the ERP or CRM
  • Trigger credits or payments
  • Commit to dates or deadlines

What separates a useful AI agent from a chatbot?

A chatbot answers questions. An agent completes a defined piece of work: it reads an incoming request, looks up the relevant information in your systems, prepares a reply or a record and hands it to the right person. The value comes from the narrow job description, not from the language model. That is why we start by writing down what the agent should do and what it must never do.

Agents only receive the permissions their task requires. Reading a CRM entry is a different risk from changing an order or emailing a customer. Actions with an external effect are prepared by the agent and approved by a person until the results have proven reliable over time. Every tool call is logged, so you can trace afterwards why the agent acted the way it did.

Agents need context: product information, internal guidelines, past cases. With CogLake this knowledge is connected once and made available to agents with sources and access rights. Before launch and after every change, we test the agent against a collection of real examples. For publicly reachable agents we can add automated attack tests with AI Security.

Sound familiar?

  • Support and internal inbox triage
  • Research and summarization
  • Operational assistants that use tools
  • Human-in-the-loop decisions

What changes

  • AI that acts inside defined boundaries
  • Human review for sensitive decisions
  • Traceable tool use and decision logs

Examples from projects like this

01

Support inbox triage

The agent reads incoming support emails, identifies the customer in HubSpot, assigns a category and urgency and drafts a reply based on the knowledge base. A support employee checks the draft and sends it with one click.

02

Preparing tender bids

For each new public tender, the agent extracts requirements, deadlines and evaluation criteria, compares them with references and certificates stored in SharePoint and prepares a checklist for the bid team's decision on whether to take part.

03

Internal IT and HR questions

Employees ask in Microsoft Teams how to request vacation, set up VPN access or claim travel expenses. The agent answers from the current guidelines with a link to the source and opens a ticket when it cannot help.

04

Phone intake outside office hours

A voice agent takes calls in the evening, records name, concern and callback number and answers simple questions such as opening hours. In the morning, the team finds structured call notes in the CRM instead of voicemails.

Before and after

Today

  • A chatbot answers broadly but cannot finish work
  • Tool access is risky or hard to audit
  • Teams do not know when to trust the output

With centerbit

  • The agent has a bounded job and clearly defined tools
  • Risky actions require review and are logged
  • Output is tested against realistic examples

How we work

What we deliver

  • Agent scope and risk map
  • Conversation, tool and guardrail design
  • Human handoff and review workflow
  • Evaluation set for regression testing
  1. 1

    Scope

    We define the job, its boundaries and the failures the agent must not make.

  2. 2

    Connect

    We connect the tools, data and context the agent needs to be useful.

  3. 3

    Control

    We add review, escalation and logging around sensitive actions.

  4. 4

    Evaluate

    We test outputs continuously and refine how the agent is run.

Frequently asked questions

Can an agent use tools safely?

Yes, if tool access is scoped, logged and routed through human review for sensitive actions.

Do you build voice or chat agents?

We design around the workflow first. The interface can be chat, voice, inbox, API or a combination.

How do we measure quality?

We define task-specific evaluation examples and test every change against them over time.

How do you prevent an agent from acting wrongly?

With several layers. The agent only gets the tools and data its task needs, actions with an external effect require approval, and inputs it cannot classify are handed to a person. Before every new version we test it against realistic and deliberately misleading examples, and during operation we review the logs regularly.

Which language models do you use, and where is our data processed?

That depends on your requirements. We work with models from several providers, including options operated in the EU, and with CogLake also on your own servers if needed. Which data the agent may send to which model is decided together with you and documented for the people responsible for data protection.

Find out which task an agent could take on.

Describe a recurring task in which your team mostly reads, looks things up and prepares. In a free initial call we assess whether an agent fits and where approvals belong.