Skip to content
Koderead

Solution

AI Agents

AI Agents That Act Inside Your Systems

A chatbot replies. An agent reads the request, looks up the record, checks the rules, performs the action in your systems, and hands anything ambiguous to a person with the context attached. The difference is whether work gets finished.

We bought an AI tool and it just answers questions. Nothing actually happens.

How you know

You probably have this problem if…

The change

What the workflow looks like before and after.

Amber marks a step the system performs. Solid ink marks one a person still performs. The count above each column is how many steps still need someone's attention.

Today

5 human steps
  1. Request arrives

    Email, WhatsApp, web form

    Performed by Customer
  2. Someone reads it

    Often hours later

    Performed by A person
  3. Looks up the record

    Switches between 3–4 systems

    Performed by A person
  4. Re-types the details

    Into the CRM

    Performed by A person
  5. Replies

    Copy-pasted from a template

    Performed by A person
  6. Sets a reminder to follow up

    Sometimes

    Performed by A person

After

1 human step
  1. Request arrives

    Same channels

    Performed by Customer
  2. Agent reads and identifies

    Matches to the existing account

    Performed by Automated
  3. Checks rules and availability

    Against live system state

    Performed by Automated
  4. Acts or proposes

    Within its written authority

    Performed by Automated
  5. Approves the exception

    Only the ones that need judgement

    Performed by A person
  6. Records and follows up

    Automatically, on schedule

    Performed by System
  • Automated
  • A person
  • System
  • Customer

The system

What this looks like running in a business.

The same components we build with.

Scope of authority — AI sales deskSigned before build

May act unattended

  • Identify the customer
  • Read order or account state
  • Answer from cited policy
  • Book or move an appointment

Must propose for approval

  • Any price outside the band

    Margin shown at approval

  • Goodwill or refund over threshold
  • Anything contractually binding

Must never attempt

  • Close a complaint
  • Invent a policy

    Escalates if not in source

  • Move money
  • Low confidence on intent14Routed with full thread and record attached
  • Price outside approved band6Held for approval, margin shown
  • Distress or complaint language2Immediate handover, never auto-closed

22 items awaiting a person, grouped by cause so one decision clears a class rather than a case.

On a real engagement this document is signed by your operations lead before any code is written.

One enquiry, start to finishWorked example
  1. Enquiry arrives21:42

    Portal, out of hours. Nobody is in the office.

  2. Conversation opens21:42

    Budget, area, property type and timeline established.

  3. Position checked21:44

    Chain and financing status asked directly, not assumed.

  4. Scored and written to CRM21:44

    Against your thresholds, entered once, no re-keying.

  5. Viewing proposed21:45

    Agent, vendor and property availability resolved together.

  6. Agent opens a warm lead08:30

    Full context attached. First substantive contact already happened.

State

Source
Portal
Received
21:42, Sunday
Message
“Is this still available?”
Known to us
No

Follows a single record through the system, including the point where it stops and waits for a person.

The work does not disappear. It moves — and the steps left are the ones that needed a person.

Koderead · scope of authority

What we build

Four things, in this order.

01

A defined scope of authority

We write down exactly which actions the agent may take unattended, which need approval, and which it must never attempt. This is a document your ops lead signs, not a prompt we hide in code.

02

Real system access

The agent reads and writes to your CRM, ERP, ticketing, calendar and billing through their APIs. Without that, it can only talk about the work rather than do it.

03

An escalation path with context

When the agent is not confident, the human receives the conversation, the record, the reasoning and a proposed action — not a bare alert to go and investigate.

04

An audit trail

Every action the agent took, what it based that on, and who approved it. This is what makes the system defensible in a regulated review.

Boundaries

What stays with a person.

Agreed in writing before the build, not discovered afterwards.

  • Pricing outside an approved band always goes to a person
  • Anything touching a contractual commitment is proposed, never sent
  • Complaints and distress signals escalate immediately and are never auto-closed
  • The agent never invents a policy — if it is not in the source material, it says so and escalates

Integrations

We connect to what you already run.

Replacing a working system mid-transformation adds risk without adding value. If the existing tool genuinely cannot support the workflow, we say so at assessment rather than three months into a build.

More specific

Narrower services under this capability.

If you arrived looking for one of these by name, each page answers it directly — including where the honest answer is that you need something else.

Questions we get

Asked and answered.

How is this different from the chatbot we already have?

Your chatbot almost certainly has no write access to any system. It can describe your refund policy; it cannot issue the refund, update the order, or notify the warehouse. An agent is defined by what it is permitted to change, not by how well it talks.

What happens when it gets something wrong?

Two controls. First, scope: high-consequence actions are proposed for approval rather than executed. Second, reversibility: every action is logged with the reasoning and can be rolled back. We design the blast radius before we design the capability.

Do we need to replace our CRM first?

Usually not. We would rather integrate with what your team already uses than add a migration to the critical path. If the existing system genuinely cannot support the workflow, we will tell you that in the assessment rather than three months into a build.

Where this lands hardest

Industries feeling this most.

Next step

Find out whether this is actually your highest-value fix.

Eleven questions, scored in your browser, with the arithmetic shown. It may well tell you that a different area matters more.

Start the assessment