Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Which Department Should Go First With AI? A Ranking Method

Rank departments for a first AI agent by repeat volume, record quality and approval risk. A scoring method for leaders who must decide where to start.

Be Human typographic cover with the title: Which Department Should Go First With AI? A Ranking Method
Conceptual illustration created with AI · Be Human

The best first department for an AI agent is the one where a task repeats often, the records behind it already agree with each other, and a wrong output can be checked by a person before it reaches a customer, supplier or payment. Score each candidate on those three factors, treat approval risk as a gate rather than a bonus, and start with the highest total that also has one named owner. This method is written for leaders who must decide where to begin, not for teams that want to test everywhere at once.

Key takeaways

  • Rank departments on repeat volume, record quality and approval risk, scoring each factor on its own before combining them.
  • Record quality is the factor leaders tend to skip, and a busy department with conflicting records is a weak first choice.
  • Any task where a mistake reaches a customer, a payment or a regulated record needs a human check designed before the agent goes live.
  • The first role should have one owner, one bounded task and one observable success criterion.
  • Treat the first department as the first part of a company intelligence layer that you will extend, not as a standalone tool.

Why the first department sets the pattern for everything after

The first agent does more than save time. It forces the company to define, for the first time, where its memory comes from, which tools the agent may call and who may approve what it does. The next department inherits those choices. A customer support agent that reads a shared customer history, for example, requires you to settle who owns that history, which fields are trusted and whether the agent may write back to it. The same questions return in finance, operations and sales.

That is why the ranking is not only about speed. A good first department lets you practice the full cycle: define the role, connect the records it needs, set approvals, measure the result and review permissions. A poor first department asks for all of that work while the stakes are high and the rules are still being invented. For a structural view of how these parts fit together, the one-page map in Draw Your AI Stack: A One-Page Architecture Map for Leaders is a useful companion before you score anything.

The three-factor method, step by step

Use one spreadsheet or a sheet of paper with a row for each department and a column for each factor. Work through the steps in order.

  • List each department and the two or three recurring tasks it performs every week, such as answering order status questions, chasing overdue invoices or preparing quotes.
  • Score repeat volume from 1 to 5. A 5 means the task happens many times a day with similar inputs and similar outputs. Count from ticket queues, inbox folders or task logs, not from memory.
  • Score record quality from 1 to 5. Ask whether the system holding the data contains the fields the task needs, whether two systems agree on the same fact, and whether a named person corrects errors when they appear.
  • Score approval risk from 1 to 5, where 5 means an error reaches an outside party, moves money, changes a contract or alters a regulated record before anyone sees it.
  • Mark every task scoring 4 or 5 on approval risk as a gated candidate. It may go first only if every output passes a human approval step from the first day.
  • For the remaining tasks, add repeat volume and record quality, then sort from highest to lowest. Break ties in favor of the task whose owner has the most time to review early results.

A common mistake is letting the loudest department win. The department that complains most often usually has the highest volume, but volume alone says nothing about whether the agent can answer correctly. Keep the three scores separate on paper so the discussion stays on evidence.

What a ranking looks like in practice

Illustrative scenario, not a client example: a small services firm with several offices compares three candidates, order status replies, overdue invoice reminders and proposal drafting.

Order status replies rank high on volume, because the same questions arrive every day. Record quality is mixed, since the order system is current but shipping partner updates arrive by email and are copied by hand. Approval risk is moderate, because a wrong delivery date sets a wrong expectation but rarely moves money. The total is strong, yet the agent would spend its early weeks checking inputs instead of answering questions.

Overdue invoice reminders rank slightly lower on volume but higher on record quality, because the accounting system and the collections list already agree on who owes what after a monthly reconciliation. Approval risk is higher, since a reminder that names the wrong amount damages a relationship, so the reminder is a gated candidate. The firm decides that the agent only drafts reminders during the first stage, and a person approves each one. This pairing has the strongest total once its approval step is in place.

Proposal drafting scores low on volume and on record quality. It may still be valuable later, but it is a poor first choice.

The result is that invoice reminders go first under approval, and order status replies follow once shipping updates reach the agent through one channel. The lesson is that the highest volume is not the answer on its own. Treat this example as a way to test your own reasoning, not as a benchmark for what your numbers should look like.

Why record quality decides the pace

Most early delays come from records, not from the model. An agent can only be as reliable as the information it reads, and a department whose records disagree will generate exceptions that someone must resolve by hand. Before you choose a department, check four things in its core system:

  • One system of record exists for each fact the task uses, such as order status, the current price list or the customer contact.
  • Names, dates and identifiers follow a consistent format across every system the agent will read.
  • Known duplicates and conflicts have a named owner who corrects them, rather than a note in a shared file.
  • Someone can explain in one paragraph where each field comes from and how often it changes.

If two of these checks fail, move the department down the list even when its volume is high. Cleaning the records is useful work in its own right, and for many companies it is the first real deliverable of the project. The broader readiness work is described in Draw Your AI Stack: A One-Page Architecture Map for Leaders, which shows where each record sits relative to the agent.

Setting the approval boundary before the first task

Approval risk is easiest to manage when the boundary is written before the agent runs. The NIST AI Risk Management Framework frames risk as something to map, measure and manage across the life of a system. For a business leader, the practical version is a short list of decisions the agent may take alone, decisions it may only draft, and decisions it may never take. Keep the list narrow for the first department.

  • Allowed alone: answering a question whose answer already sits in an approved document, with the source cited.
  • Draft only: any message that names an amount, a deadline, a commitment or a legal term.
  • Never: changing a price, issuing a refund, altering a contract or writing into a regulated record.
  • Read access only to the fields the task needs, following the least privilege principle in OWASP's authorization guidance.

Treat text from customers and suppliers as data, not as instructions. The OWASP guidance on prompt injection describes how inbound messages can try to steer an agent away from its task, so the boundary must hold even when an email asks the agent to do something unusual. Approval thresholds should rise slowly as the agent proves itself, and Approval Thresholds for Superintelligent Agents: Raise Them Slowly explains how to move a boundary without losing control.

How Be Human handles the first department

Be Human's managed approach starts with one role rather than a platform rollout. The first engagement defines that role, the memory and tools it needs, its approval boundaries and an observable success criterion, and the agent then runs under human control. Ongoing work covers monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled changes. Records, permissions and approvals stay documented, so the next department can reuse them rather than rebuild them. The first department chosen with this method becomes the first well-understood part of the company's intelligence layer.

Once the first role is running, the adoption work matters as much as the configuration. How to Get Your Team to Actually Use an AI Agent covers the habits and handoffs that a ranking alone cannot settle.

If you are not yet sure which department holds the first task worth handing to an agent, take the free Be Human AI diagnostic to find it.

Frequently asked questions

Should we start with the department that has the most work? Not automatically. High volume helps only when the records behind the work are accurate and mistakes can be caught before they reach an outside party. A busy department with conflicting records usually produces slower results than a smaller, cleaner one, so treat volume as one of three factors.

Can a leader score departments without technical staff? Yes. The method relies on counts from existing queues, inboxes and logs, plus plain questions about who owns each field and where errors are checked. A technical reviewer can verify the record and permission details later, but the ranking itself is an operational decision that belongs to the business owner.

Is approval risk the same as the chance that the AI is wrong? Not exactly. Approval risk measures what happens if an output is wrong and nobody catches it before it reaches a customer, supplier, payment or record. It can be lowered by design, for example by limiting the agent to drafting, requiring sign-off or restricting what the agent is allowed to write.

How long should the first role stay on a narrow scope? Keep it narrow until the success criterion is met consistently across several review cycles and permission reviews show no unexplained access. Widen the boundary in small steps backed by evidence, not because the team feels confident. Expansion should follow the record of results, not the enthusiasm of the moment.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →