Skip to content

automation 04

AI Agents

Eighteen agents, each with one job and a clear edge to it. They answer, extract, qualify and call, and they stop and ask a human at the point where being wrong would cost you.

Install
Scoped per agent
Timeline
About 21 days to live
Runs on
Your own n8n
01

18

agents

02

21

days to live

03

68%

handled alone

An agent is not a chatbot with a personality. It is a member of staff with one job, a clear edge to that job, and a human it knows to stop and ask.

how it runs

Watch the jobDraft modeConfidenceHandled aloneEscalatedMonthly review
Draft mode is the whole safeguard. The agent writes beside your team without sending anything, so its accuracy is known before a customer ever sees it.

the workflows

  1. 01

    Support agent

    Answers the repeat questions from your own documentation and past resolved tickets. Anything it cannot source from those goes to a person with its reasoning attached.

  2. 02

    Knowledge assistant

    Internal search that answers in sentences and cites the document it read. Built for the questions new staff ask in their first month.

  3. 03

    Document extraction

    Invoices, statements and forms turned into structured rows. Low-confidence fields are flagged for review rather than written through.

  4. 04

    Voice agents

    Inbound and outbound calls that qualify, book and hand off. The slowest of the five to release, because interruption handling has to be right before it takes a real call.

  5. 05

    Multi-agent pipelines

    Several narrow agents in sequence with a checker at the end, used where one agent doing everything would be vague at all of it.

how it holds up

  • 01

    Watch the job first

    We sit with the person doing the job now and write down every judgement call they make. That list, not a prompt, is what the agent is built from.

  • 02

    Draft mode before release

    The agent runs beside your team writing drafts nobody sends. You see its accuracy on real cases before it ever touches a customer.

  • 03

    It escalates with its reasoning

    The clear cases it handles alone. The rest arrive at a person with the working shown, so the human starts from a position rather than from scratch.

  • 04

    The escalations are the roadmap

    Monthly review of what it handed back. That list is how it gets better, and it is also how you find the places it should stay out of.

what it will not do

  • About a third of everything still reaches a person. The 68% is what we aim at, not a promise, and an agent that never escalates is one that has been told to guess. 

  • It will not invent a policy. Where your own documentation is silent or contradicts itself, it stops and asks rather than picking the more helpful answer. 

  • Voice is the slowest to release. Latency and interruption handling take longer to get right than text, so plan a longer draft-mode period there. 

  • An agent is only as good as the archive it learned from. If nobody wrote the answers down, there is nothing to build on yet. 

you supply

  • A runtime we can deploy into, yours or ours
  • Anthropic or OpenAI API key
  • Access to the knowledge it should answer from
  • A named person for escalations

running cost

Model cost only, billed by your own provider. It scales with conversation volume and with how much context each agent is given, not with seats.

Billed by your own providers, not by us. There is no platform fee and no seat count.

InstallAI?

Scoped per agent, about 21 days to live. Tell us about your setup and we will confirm it fits before taking anything.

hello@kraftline.org