Getting Started

Where Should a Company Start With AI? How to Find the Workflows Worth Automating

Mitori Team•October 16, 2026•10 min read

Use this article with AI

Open a buyer-ready brief in your preferred assistant.

ChatGPT, Claude, Perplexity and Grok open with the prompt attached. Gemini copies the prompt to your clipboard first.

Start with the workflows that are high-volume, rule-heavy and full of handoffs between people and tools, not the use cases that demo best. Score candidates on volume, repetition, handoffs, error cost, data availability and risk, and base the scores on observed work rather than opinions, because much of that work never reaches a system of record.

Mitori is an operational intelligence platform that observes how work actually happens across your people, tools and handoffs, then tells you what to automate first, in what order and under what controls — and builds and monitors the AI agents that do it.

Why do most companies start with AI in the wrong place?

Most companies start with a tool or a brainstorm, so they pick the use cases that are easiest to imagine and present. Those are usually visible, front-office ideas, not the repetitive operational work where automation pays back.

There are two common starting points. The first is tool-first: someone buys a copilot or an agent platform and the team goes looking for things to point it at. The second is a workshop: leaders list ideas on a whiteboard, vote, and fund the ones that sound most strategic. Both feel productive. Both select for what is memorable rather than what is valuable.

The research backs this up. According to MIT NANDA's "The GenAI Divide: State of AI in Business 2025", about half of GenAI budgets go to sales and marketing, while back-office automation often yields better ROI. The authors put the bias down to easier metric attribution rather than actual value, and note that the highest-ROI opportunities in back-office functions remain underfunded. The same report found that 95% of organizations were getting zero return from GenAI, and that the divide seemed to be driven by approach, not model quality.

Picking the wrong first workflow is expensive in a quiet way. The pilot works in a demo, nobody's week actually changes, and the program loses credibility before it reaches the work that matters. We cover that failure pattern in detail in our piece on why AI pilots fail.

What makes a workflow worth automating first?

A strong first workflow happens often, follows mostly stable rules, moves information between several tools and people, and has a cost when it goes wrong. It also has the data it needs in digital form, and a failure that can be caught before it does damage.

The best first candidates are rarely glamorous. They look like intake, scheduling, credential and document checks, status updates, reconciliations, order or referral routing, and the follow-ups that keep all of those moving. They tend to sit between systems rather than inside one.

A lot of this is what Asana calls "work about work": communicating about work, searching for information, switching between apps, managing shifting priorities and chasing status. According to Asana's Anatomy of Work research, 60% of a person's time at work goes on work about work rather than skilled work. Microsoft's 2025 Work Trend Index found that employees are interrupted every 2 minutes during the core workday by meetings, emails or pings, and that four in five people say they lack enough time or energy to do their work. That coordination load is where repetitive, rule-heavy handoffs hide.

High volume: it runs dozens or hundreds of times a week, not once a quarter.
Rule-heavy: most instances follow the same decision logic, with a known set of exceptions.
Cross-tool: someone copies, checks or re-keys information between email, spreadsheets, chat and a business system.
Handoff-heavy: the work waits on people passing it along, and things get lost in the gaps.
Measurable: you can establish today's baseline for time, volume and error rate before you change anything.

How do you score AI opportunities? A six-factor rubric

Score each candidate workflow from 1 to 5 on volume, repetition, handoffs, error cost, data availability and risk. Add the first five for a value-and-feasibility score, then use risk to decide which wave the workflow belongs in and what controls it needs.

The rubric is deliberately simple. The hard part is not the arithmetic, it is the inputs. If the volume and handoff numbers come from a workshop, you are scoring people's memory of the process. Write down the evidence behind each score, and mark how confident you are in it. A 5 based on two weeks of observed work and a 5 based on a manager's estimate are not the same number.

For the finance side of this, including how to build an ROI case before you commit budget, see how to measure AI ROI before you buy, and our guide to how CFOs should decide what to automate first.

1Volume (1–5): How many times does this workflow run per week, and how many people touch it? Score 1 for a handful of instances a month, 5 for a steady daily stream across a team.
2Repetition (1–5): How much of each instance follows the same steps and rules? Score 1 if every case needs fresh judgment, 5 if most cases are routine and exceptions are recognizable.
3Handoffs (1–5): How many tools and people does one instance pass through? Score 1 for a single person in a single system, 5 for work that crosses several tools and waits on several people.
4Error cost (1–5): What does a mistake in today's manual version cost in rework, delay, lost revenue or exposure? Score higher where inconsistency is expensive, because consistent execution is where automation earns its keep.
5Data availability (1–5): Is the information the workflow needs digital, accessible and reasonably clean? Score 1 if it lives in people's heads or on paper, 5 if it already sits in systems and documents you can connect to.
6Risk (low, medium, high): What happens if an automated step is wrong, and could you catch it in time? Don't subtract this from the total. Use it to set sequencing and controls: low-risk workflows go in the first wave, higher-risk ones follow with human approval and escalation built in.
7Rank by the combined score, then sanity-check the top five against two questions: can we measure a baseline today, and is there a clear owner who wants this fixed?

What are the signs of a good first candidate, and a bad one?

Good first candidates are frequent, routine, measurable and recoverable when something goes wrong. Bad first candidates are rare, judgment-heavy, poorly documented or high-stakes with no way to catch errors before they land.

A poor first candidate is not a permanently bad idea. A high-stakes, judgment-heavy workflow can still be automated in part, later, once the organization has working controls and a track record from earlier waves. It just shouldn't be the place you learn.

Be wary of a chatbot as a default first project, too. If you are unsure whether a candidate needs an agent, conventional automation or a chat interface, our comparison of AI agents, automation and chatbots walks through the difference.

Signal	Good first candidate	Poor first candidate
Frequency	Runs daily or weekly at steady volume	Runs a few times a quarter, or only during a crisis
Decision logic	Mostly stable rules with a known set of exceptions	Every case needs senior judgment or negotiation
Tools touched	Several tools, with copy-paste and re-keying between them	One system that already has good built-in automation
Inputs	Emails, forms, documents and records that are already digital	Tribal knowledge, phone calls nobody logs, or paper
Cost of a wrong step	Caught by a review or approval before it reaches a customer	Irreversible or regulated, with no checkpoint in place
Baseline	Volume, cycle time and error rate can be measured today	Nobody can say how long it takes or how often it fails
Ownership	A named owner who feels the pain and wants it fixed	Owned by a committee, or by nobody

Signal

Frequency

Good first candidate

Runs daily or weekly at steady volume

Poor first candidate

Runs a few times a quarter, or only during a crisis

Signal

Decision logic

Good first candidate

Mostly stable rules with a known set of exceptions

Poor first candidate

Every case needs senior judgment or negotiation

Signal

Tools touched

Good first candidate

Several tools, with copy-paste and re-keying between them

Poor first candidate

One system that already has good built-in automation

Signal

Inputs

Good first candidate

Emails, forms, documents and records that are already digital

Poor first candidate

Tribal knowledge, phone calls nobody logs, or paper

Signal

Cost of a wrong step

Good first candidate

Caught by a review or approval before it reaches a customer

Poor first candidate

Irreversible or regulated, with no checkpoint in place

Signal

Baseline

Good first candidate

Volume, cycle time and error rate can be measured today

Poor first candidate

Nobody can say how long it takes or how often it fails

Signal

Ownership

Good first candidate

A named owner who feels the pain and wants it fixed

Poor first candidate

Owned by a committee, or by nobody

Why can't interviews or system data find these workflows?

Interviews capture the process people believe they follow, and system logs capture only what lands in a system of record. The handoffs, chasing and copy-paste that make a workflow worth automating mostly happen in between.

Ask a team lead how intake works and you'll get the designed process: the form comes in, it gets checked, it goes to scheduling. What you won't hear about is the inbox someone watches for missing documents, the spreadsheet that reconciles two systems that never quite agree, or the three Slack messages it takes to find out whether something was approved. People stop noticing that work because it is just how the day goes.

System data has the opposite blind spot. Process mining and analytics tools read event logs from systems of record, which makes them precise about what happens inside one system and largely silent about the work around it. The comparison of process mining and operational intelligence goes deeper on that distinction.

This is why Mitori is built on observation rather than interviews. The audit watches how work actually moves across the tools teams already use, with consent, so the volume, repetition and handoff scores come from evidence rather than recollection.

How does Mitori's observed audit surface and rank workflows?

Mitori observes real work for 30, 60 or 90 days, then scores each automation candidate on savings, effort and confidence and sequences the strongest into deployment waves with governance checkpoints. Mitori then builds the agents and monitors them against the observed baseline.

Mitori runs on a four-part operating model: Audit, Roadmap, Build, Control. The Audit observes real work across email, spreadsheets, chat and business systems, with consent. Nothing depends on someone remembering to mention a step in a workshop.

The Roadmap is the ranking. Each candidate is scored on savings, effort and confidence, which lines up with the rubric above: volume and repetition drive savings, data availability and handoff complexity drive effort, and the depth of observed evidence drives confidence. Candidates are then sequenced into deployment waves, with governance checkpoints such as approvals and escalation attached to the workflows that need them.

In Build, the agents start from the observed workflows rather than a generic template. In Control, Mitori logs every agent action and tracks drift against the observed baseline, so you can see whether the workflow is behaving the way the roadmap assumed. Mitori's agents already run in healthcare operations, including IVF clinics and healthcare staffing agencies, where intake, scheduling and credential handoffs are exactly the kind of work this rubric surfaces.

What should you do first, this quarter?

Pick two or three teams with heavy operational load, list their recurring workflows, score them with the rubric using real evidence, and commit to one or two low-risk candidates with a measured baseline.

If you want a quick read on where your organization stands before you begin, the readiness assessment takes a few minutes. If you are weighing who should run this work, a consultant, an audit or your own team, we compare the options in AI consultants vs an AI audit vs doing it yourself.

1Choose teams where coordination load is obvious: operations, finance, intake, scheduling or compliance-heavy back-office work.
2List every recurring workflow those teams run, including the informal ones that live in inboxes and spreadsheets.
3Gather evidence on volume, handoffs and errors, ideally by observing the work rather than asking about it.
4Score each workflow on the six factors and write down the evidence and confidence behind each score.
5Select one or two low-risk, high-score workflows for the first wave, and record today's baseline before anything changes.
6Define the controls up front: who approves, what escalates, and how you will spot drift once an agent is live.

Frequently asked questions

Start with processes that run often, follow mostly stable rules, move information between several tools and people, and have a measurable cost when they go wrong. Intake, scheduling, document checks, reconciliations, routing and status follow-ups are common examples. Avoid rare, judgment-heavy or high-stakes workflows with no review step as your first project. They can come later, once controls and a track record exist.

List the recurring workflows in your most operationally loaded teams, then gather evidence on how often each runs, how many tools and handoffs it involves, and where errors happen. Observing the work is more reliable than interviews, because much of the effort is coordination that people don't mention and systems don't log. Score each candidate with a consistent rubric so you compare like with like.

Usually not as the core of your first AI project. General-purpose assistants help individuals, but they rarely change how an operational workflow runs end to end. MIT NANDA's 2025 research found that widely adopted tools mainly enhance individual productivity rather than P&L performance. A better first project targets one specific, high-volume workflow with a clear baseline, an owner and defined controls.

One or two for the first wave is usually enough. Fewer workflows means you can measure a clean baseline, put proper approvals and escalation in place, and learn how agents behave in your environment. Once the first wave is stable and measured, the next workflows on your ranked list can follow in planned waves rather than all at once.

A workshop can produce a list in a day, but that list reflects opinions. An evidence-based view takes longer because it needs to see real volume and handoffs. Mitori's audit runs for 30, 60 or 90 days of observed work, depending on scope, and ends with a sequenced, confidence-scored roadmap of what to automate first and under what controls.

Next step

Book an audit scoping call

Find out which of your workflows are worth automating first, ranked from observed work rather than a brainstorm.

Book an audit scoping call →Review pricing

Product Audit Proof Trust Pricing Readiness Assessment

Sources

How to Deploy AI in Your Business: A Step-by-Step Guide (2026)

How to deploy AI in your business in 7 steps: observe real work, pick workflows by evidence, set governance, build agents with approvals, then scale in waves.

AI ROI

How to Measure AI ROI Before You Buy Any Tools

How to measure AI ROI before you buy: capture an observed baseline, model savings as ranges with confidence, price effort and risk, then track realized value.

AI Strategy

Why AI Pilots Fail (and How to Get to Production)

Why AI pilots fail: the research points to workflow fit, missing baselines and late governance, not model quality. Seven failure modes and the fix for each.

CFO Decision Guide

How CFOs Should Decide What to Automate First

The real CFO question is not whether AI matters. It is which workflow deserves capital first, what confidence exists in the ROI, and what should wait.