AI Agents for Operations: Start With Customer Prep Work

It's 8:40, and your account manager has a 9:00 call with a property manager who runs twelve of your sites. Before dialing, someone opens the CRM for the last conversation, the field service app for the last three visits, QuickBooks for the open balance, and an inbox for the quote that went out two weeks ago. Fifteen minutes later the brief exists in one person's head and nowhere else.

AI agents for operations are software that read from your business systems, assemble what a person needs for the next decision, and propose changes for a human to approve. The best first job for one is customer prep: pull the account history, draft the call brief or quote follow up, flag overdue invoices, and write back a task. A wrong draft costs a reviewer a few minutes. A wrong write to an invoice costs money.

Our take, from building data pipelines, integrations and approval flows: the agent's value comes from two things. It needs read access to clean, joined customer data, and a short list of write actions that each pass through an approval step. The model is the easy part.

This guide covers what the agent reads, what it may change, how approvals work, and how to estimate the hours it gives back.

What an AI agent for operations does

Anthropic separates two designs in Building effective agents. A workflow runs a model through steps you coded in advance. An agent decides which tools to call and in what order. Prep work usually needs a mostly fixed workflow with one agentic step: deciding what in the history matters for this call. Anthropic's advice is to find "the simplest solution possible" and add complexity when the task demands it.

Four terms we use through the rest of this guide:

• Read tool: a function the agent calls to fetch data, such as the open invoices for one customer.
• Write action: a function that creates or changes a record, such as a CRM task or an email draft.
• Approval step: the screen where a person sees a proposed write with its evidence, then approves, edits or rejects it.
• Touchpoint: any scheduled customer contact, from a renewal call to a site walk or a collections call.

OpenAI's practical guide to building agents sorts tools the same way. Data tools retrieve context. Action tools change systems. Keeping those two lists apart is the most useful design decision you will make.

Why customer prep is the right first job

The time cost of hunting for information is well documented. McKinsey Global Institute found interaction workers spend nearly 20 percent of the workweek looking for internal information or tracking down colleagues who can help (The social economy). A 2023 Gartner survey of 4,861 workers found 47% struggle to find the information they need to do their jobs, and the average desk worker uses 11 applications (IDM report on the Gartner survey). Salesforce's 2024 sales research put work outside selling, including admin and meeting prep, at 70% of a rep's time (Salesforce State of Sales statistics).

Prep work is where that searching piles up, right before a customer conversation. It also has the right risk profile. OpenAI's guide recommends rating every tool low, medium or high risk by four factors: read or write access, reversibility, the permissions it needs, and financial impact. The same guide says teams "typically achieve greater success with an incremental approach." Prep sits at the low end of every factor. The output is a draft a person reads before the call, so a misread note gets caught in the brief.

Compare that with an agent that issues credits or changes prices in its first week. In June 2025, Gartner predicted over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs and unclear business value (Reuters via The Star). A prep agent answers the value question fast, because you can time prep before and after.

What the agent reads: the prep packet

A customer prep agent for a service business usually pulls from four places.

The output is a one page brief with five parts:

1. Account snapshot: sites, contract value, owner, last contact date.
2. Service history: recent visits, open jobs, anything that went wrong.
3. Money: open balance, oldest unpaid invoice, any dispute.
4. Open loops: quotes waiting on an answer, promises from the last call.
5. Suggested agenda plus proposed write actions: the follow up task, the quote email draft, the overdue invoice flag.

Every line in the brief links to its source record. A reviewer who doubts a number clicks through to the invoice and checks it in seconds.

Joined data comes before the agent

Most AI agent guides skip this step, and it decides whether the brief is right. The same customer is "Harbor Point Property Mgmt" in the CRM, "Harbor Point PM Bldg 4" in the field tool and "Harbor Point Properties LLC" in QuickBooks. Give an agent three search tools and it will guess at the match. Some days it guesses wrong, and the brief shows another customer's overdue invoice.

The fix is a data job. Build a customer key table that maps each system's ID to one account ID, with a person confirming the matches the rules can't settle. Then give the agent one read tool, something like get_account_snapshot(account_id), that returns the joined record. Anthropic's guide to writing tools for agents argues for this shape: consolidate several operations into one tool and return high signal fields such as names in place of raw IDs. Twelve thin wrappers around twelve API endpoints give the model twelve chances to pick the wrong one.

What we check before any agent work starts:

• Can every active customer be matched across the CRM, field tool and accounting system? What share needs a human to confirm the match?
• Does the open balance in the brief equal the number finance sees today?
• How fresh is each source? A nightly sync works for visit history. Balances may need a live call.
• Who owns the mapping table when sales adds a new customer?

This matching and modeling work is what our unified data platform service covers, and it pays off well beyond the agent.

The short list of write actions

The read side makes the brief useful. The write side is where an agent saves the most work, and where it can do damage. Keep the list short and rate each action.

The MCP tools specification builds this into the protocol. It says a person should always be in the loop, able to deny a tool call before it runs. The MCP schema adds annotations such as readOnlyHint and destructiveHint so a client can tell a lookup from a change. Treat those hints as labels. The spec tells clients to consider annotations untrusted unless they come from trusted servers. Real permission lives in the API credentials the agent holds. Make the accounting connection read only at the token level, and no prompt can talk the agent into editing an invoice.

Each write action should fit in one line. For a HubSpot follow up task, the agent fills five fields from the HubSpot Tasks API: hs_task_subject, hs_task_body, hs_timestamp (the due date), hs_task_priority and hubspot_owner_id, plus an association to the contact or company. The approval screen shows those five values. A reviewer reads them in a few seconds.

What an approval step looks like

Approval is a screen, and its design decides whether people use it. The version we design for shows:

• The proposed change, field by field, with the current value beside the new one.
• The evidence: the source records behind the change, linked.
• Three buttons: approve, edit, reject. Edit opens the draft in place.
• A log entry for every decision, with who approved it and when.

The MCP security guidance asks clients to show tool inputs to the user before calling the server and to log tool usage for audit. Both belong in an operations agent from the first release.

When we built a pricing portal for a commercial roofing contractor, contractors generated and submitted quotes while admins managed pricing, labor, margins, approvals, users and workflows. The same split carries over to agents. The person who drafts a change and the person who approves it get different screens and different permissions. An agent is one more drafter in that model.

We also designed a lead reactivation bot for an insurance advisor: a conversational, automated system to revive inactive CRM leads. The questions we design for in a system like that are the ones in this guide. What may it read from the CRM? What may it send to a lead? Which actions wait for the advisor?

How many hours come back

Run the math on your own numbers. Time ten prep sessions with a stopwatch. Then time ten reviews of an agent brief during a pilot.

Hours saved per week = touchpoints per week × (manual prep minutes minus review minutes) ÷ 60

With example inputs of 40 touchpoints a week, 15 minutes of manual prep and 3 minutes to review a brief, that is 8 hours a week. Compare the two bars below with your own stopwatch numbers.

Review time is the number to watch. If reviewers spend 10 minutes checking each brief, the data underneath needs work, usually the customer matching from the earlier section.

A rollout plan for your first operations agent

1. Pick one touchpoint type. Renewal calls or quote follow ups work well because they repeat every week.
2. Map the systems and build the customer key table. Fix matching before writing a prompt.
3. Build one read tool that returns the joined account snapshot.
4. Generate briefs in shadow mode. The account manager preps as usual and compares.
5. Turn on the first write action, the CRM follow up task, behind approval.
6. Track three numbers: review minutes per brief, edits per brief and rejected write actions.
7. Add the next write action once reviewers stop editing the current ones.

This is how we scope AI automation and agents work: one measurable workflow, read access first, writes added one at a time.

Start with the brief, then earn the writes

An operations agent pays off when your team walks into every customer call knowing the history, the open jobs and the money, and walks out with the follow up already logged. That comes from joined data the agent can read and a few write actions a person approves. Prep work is the safest place to prove it, because every mistake shows up in a draft.

If your customer history lives across a CRM, a field tool and QuickBooks, we can help you map it and scope a first prep agent. Book a workflow call and bring the list of systems involved.

Related reading: how to design the review step, building a single customer view (the joined data an agent reads), and our AI financial intake agent case study. Planning a first operations agent? Get in touch.

Sources

• Building effective agents, Anthropic, December 19, 2024
• Writing effective tools for AI agents, Anthropic, September 11, 2025
• A practical guide to building agents, OpenAI, April 2025
• Tools, Model Context Protocol specification, Model Context Protocol, June 18, 2025
• Schema reference, Model Context Protocol specification, Model Context Protocol, June 18, 2025
• The social economy: Unlocking value and productivity through social technologies, McKinsey Global Institute, July 1, 2012
• 47% of Digital Workers Struggle to Find Information: Gartner, IDM (reporting a Gartner survey), May 18, 2023
• Sales AI statistics 2024, Salesforce, July 25, 2024
• Over 40% of agentic AI projects will be scrapped by 2027, Gartner says, Reuters via The Star, June 25, 2025
• Tasks API guide, HubSpot Developers, accessed November 3, 2026