What to hand to AI first.

The manual work to hand to AI first is work that is repetitive, already written down and easy to check, and that touches customers, money or personal data as little as possible. Manic Edge plans an AI rollout in three stations, in this order: a knowledge base the AI tools draw on, then automations between the tools, then an orchestrator that coordinates them. Each station needs the one before it.

By Colin Moroney, co-founder of Manic Edge · Published 15 October 2026

Start with where the hours go

Most businesses that are told to use AI start with the tool. The more useful first step is a count of the hours spent on repetitive manual work, and of what kind of work fills them. Those hours are usually spread across several people, which is why nobody has added them up.

Nearly all of it falls into five kinds of work:

Moving information between tools

Typing the same details from one system into another: orders into the accounting tool, form answers into the CRM, updates into a spreadsheet.

Answering the same questions

From the team or from customers. The answer usually exists somewhere, and someone spends the day finding and retyping it.

Producing documents that follow a pattern

Reports, quotes, proposals and summaries built the same way each time from the same kinds of data.

Scheduling and chasing

Reminders, follow-ups and status requests, often sent by hand from someone’s inbox.

Checking and approving

Reviewing other people’s work against a standard before it moves on.

The total shapes the plan. Under five hours a week, the case for automation is thin and the knowledge base usually matters more. Five to twenty hours is usually two or three hand-offs repeated many times. Past twenty, the manual work is roughly a full-time role in its own right, and the order of the handoffs matters more than which tools get chosen.

Rank the work by hours and risk

Every manual process goes on one list, ranked on two things: the hours it takes each week and how safe it is to hand off. A process scores well on safety when it is written down, done the same way every time, and easy to check after the fact.

Work that touches customers, money or personal data ranks lower on safety, however many hours it takes. The early automations go on the internal steps around that work, with a person reviewing anything that reaches a customer or moves money. Those internal steps build the trust and the audit trail that the customer-facing steps need later.

The rollout order

The ranked list says which work to hand off. The rollout order says when. Manic Edge plans it in three stations, and each one depends on the one before it.

01The knowledge base

AI works from what is written down. A business where the answers live in people’s heads, or are scattered across drives and old threads, gives an AI tool nothing reliable to draw on, and an automation built on top learns whichever version of the work it was shown.

02The first automations

Once the knowledge is written down and the common work is done the same way, the repetitive steps can be automated. The tools get connected, the busiest hand-offs between them go first, and every automation has a named owner before it is built.

03Orchestration

With automations running and owned, the open question is whether a multi-agent orchestrator earns its place: one layer that passes work between the automations, the AI agents and the people, and keeps a record of what it did.

The first two stations need no AI at all, and for some businesses they are most of the plan. AI comes in on top of a maintained knowledge base and connected tools.

The first handoff, by kind of work

The kind of work that takes the most hours decides the first handoff, and how much of a person stays in the loop.

Five things to check first

Five questions place a business in the rollout order. The lowest answer among the ones a station depends on is the one to work on first.

The concrete first steps are small. Write down the answers to the twenty questions people ask most, in one place everyone can reach. Write down the steps of the most common task the way the person who does it best actually does it. Name an owner for every automation before it is built. Pick the one hand-off people type most often, connect those two tools, and measure the hours before and after.

A rollout in practice

A support team of 30+ agents checked quality by hand. A lead pulled a sample of resolved cases each month, read the replies and scored them on a form. That covered a small fraction of the conversations customers actually had, and the scores depended on who did the reading.

Station one was writing the standard down. Before anything was automated, the review form became four written rubrics: grammar, whether the reply solved the problem, empathy, and how the conversation was closed. A rubric is the knowledge an AI scorer works from, so it had to be specific enough that two people would score the same reply the same way. The second version was checked against about 10k past cases before it went live. This took several weeks and was the slowest part of the rollout.

Station two was the plumbing. A nightly run pulled the day’s resolved cases out of the helpdesk, cleaned up the message text and published it to a shared review list. None of this used AI. It replaced the manual pulling and pasting, and it gave every case a record that could be traced later.

Station three was the orchestrator. One coordinating agent sent each message to four scoring agents, one per rubric, running in parallel. It combined their scores and passed only the uncertain cases to a person, who could override any score where it sat. At month end the results rolled up into a scorecard for each agent.

What changed: In its first full month, the system scored about 6,000 cases and 8,900 messages, which is every agent-to-customer message, not a sample. Through additional calibration, the share of messages sent to a person for review fell from 46% at the start to under 15% by the final week.

Choosing the tools

The order matters more than the brand. Most teams already pay for tools that can do the first two stations, so the first question at each station is what’s already in the building.

For the knowledge base, look for one place everyone can search, a named owner for each section, and an easy way to see what’s out of date. The AI tools you add later need to be able to read from it, so check that it connects to them and that permissions carry over. If the team won’t keep it current, no AI tool will make up for that.

For automations, look for a platform your own people can maintain without a developer, with run logs you can read when something fails. Check that it connects to the tools you use most, and look at the price per run once volumes grow. Every automation needs an owner, so the platform should make it obvious who built what.

For an orchestrator, look for a clear record of every step it took, a built-in point where a person reviews or overrides, and a way to change one agent without rebuilding the rest. If it can’t show what it did last week and who checked it, it isn’t ready to sit between your tools and your customers.

Where to start

The free AI Readiness Check takes about ninety seconds. It asks eight questions about where the hours go and how the business runs today, and says which station the business is at, which of the five lenses to move first, and what the first handoff is. It runs in the browser and stores nothing.

The AI rollout plan from Manic Edge turns that into a document: every manual process ranked by hours and risk, the knowledge base structure, the first automations and where an orchestrator fits, and a 90-day plan in the order to do it. It is a single fixed fee, and the details are on the AI readiness page.

Take the AI Readiness Check About AI readiness