Where to start with AI in operations: the first-90-days guide

Most teams ask where to start with AI in operations after they have seen a demo. That order causes the trouble. This guide gives you four decisions in a 90-day sequence, a timeline you can hand to your team, and a short list of choices you can make this week. It is written for an operations leader who is not a technical specialist.

01What does "starting with AI in operations" actually mean?

Starting means choosing two real processes, deciding how you will judge the result, and deciding who stays in control. It does not mean buying a platform or writing a company-wide AI strategy.

We use a few terms the same way throughout, so here they are once. A proactive AI agent is software that watches a process, notices when something needs doing, and either acts or asks a person, without waiting to be prompted. Process automation with AI is the wider practice of having software carry out the steps of a defined business process, using AI where a step needs judgment, such as reading a free-text message.

Human in the loop means a person reviews or approves the AI's action before it takes effect. An exception path is the route a case takes when the AI cannot or should not handle it; it always ends with a named person. A pilot-to-production gate is a written yes-or-no review at the end of a pilot, held against the success criteria agreed at the start, with a named decision maker.

The sequence has four steps, and the rest of the article takes them in order:

1. Pick two processes.
2. Agree success criteria for each.
3. Set the oversight level.
4. Set the pilot-to-production gate.

02How do you pick the two processes to start with?

Pick two processes that are frequent, writable, owned, recoverable and already recorded in a system. Choose one that is steady and low-risk, and one that stretches a little. Two gives you a comparison. One gives you nothing to compare against, and five spreads your team too thin to learn from any of them.

Run this five-part test on every candidate. A good starting process needs a clear yes on all five.

1. Frequent. It happens often enough that you will see many cases within a few weeks.
2. Writable. You can describe the normal steps and the usual exceptions in a page or two.
3. Owned. One person is responsible for it today and can be reached during the pilot. That person is your first reviewer.
4. Recoverable. A wrong result can be caught and corrected without lasting harm to a customer, a payment or a legal position.
5. Recorded. The inputs and outcomes already live in a system, such as a ticketing tool, a spreadsheet people really maintain, or an inbox with consistent labels. If the work only exists in people's heads, the agent has nothing to read.

Now choose the pair. The first process should pass all five and be unglamorous: sorting incoming requests, checking forms for missing fields, chasing a routine approval. The second can stretch one criterion, for example a process with more exceptions or one that touches another team.

Adding software does not repair a broken process, so its problems carry over into the new setup. If your first candidate has unclear steps, tidy them before any AI is involved, and do that before you settle the pair.

03What success criteria should you agree before the pilot starts?

Agree four things for each process: one measure, a baseline recorded before Day 1, a target, and a stop condition. Write them on one page and have the process owner and your sponsor initial it. Without this page, the pilot ends in opinions, and the loudest opinion wins.

1. One measure. Pick a single measure of time or errors. For request sorting it might be the time from arrival to the right queue; for a form check, the number of forms sent back for missing fields.
2. A baseline. Measure the current process before the pilot, over a period long enough to include a normal mix of cases. Without a baseline you cannot say whether anything improved. Do this in the first two weeks.
3. A target. Say what result would make you want to continue. Set it from your own baseline, not from a figure you read somewhere; we give no benchmark because any number would be invented for you.
4. A stop condition. Say what result or event ends the pilot early: a wrong result in a category you named, a customer complaint, or a measure that gets worse for a set stretch.

A criterion is only as good as the boundary around the work it judges. Many teams settle that boundary in a single working session with the process owner, an operator and the sponsor.

04How much oversight should the AI have at the start?

At the start, the AI should ask first. That is the lowest of three oversight levels, and it is the right place to begin for every process on your list. You raise the level only when results justify it, and you write down who decides that and when.

Starting at "asks first" does three useful things. It protects you while you do not yet know how the system behaves on your real cases. It produces a log of what the AI proposed and what the person changed, which is your best evidence at the gate. And it shows your team that people stay in charge.

Two details make this level work. First, name the exception path. Decide what the AI does when it is unsure, when information is missing, or when a case falls outside the process as you described it. The answer is never "guess". It is "stop and send to this person, with the reason".

Second, name who answers. An agent that asks and gets no reply is a stalled process, and people soon learn to ignore it. For each process, name a primary and a backup, and say how fast they are expected to reply.

05What is a pilot-to-production gate, and who decides?

A pilot-to-production gate is a written yes-or-no review at the end of the pilot, held against your success criteria, with one named decision maker. It stops a pilot drifting into production by default.

Set the gate in the first weeks, not at the end, while you are still neutral about the result. Fix the date in days from the start and say who sits in the room. Day 60 and Day 90 are both reasonable.

Bring the same short list to every review:

1. Results against the criteria. The measure, baseline and target, and what happened, process by process.
2. Exceptions seen. What went down the exception path, how often and why.
3. Oversight for production. The level you propose, with evidence from the log of corrections.
4. Owner after launch. The person accountable for the process and the agent once the pilot team steps back.
5. Rollback. How you return to the old way of working, how long it takes, and who can order it.

The decision maker has three honest options: go to production as proposed, go with named and dated changes, or stop. Stopping is a legitimate outcome. You spent a few weeks and learned something specific about your process.

The decision maker should not be the person who built the pilot. Choose someone with authority over the process and no stake in the pilot looking good, typically the process owner's manager or the operations lead.

06What does the first 90 days look like, week by week?

The first 90 days run in four phases: choose and measure in Days 1-14, set criteria, oversight and access in Days 15-30, pilot at "asks first" in Days 31-60, and review, gate and decide in Days 61-90. The timeline uses day numbers only, because your start date is yours to choose.

Composite example (illustrative, not a client). Picture an operations lead at a mid-sized distribution business. The details below are invented for illustration. No real company stands behind them, and no result is implied.

Days 1-14: choose and measure. The lead lists the recurring tasks and runs the five-part test on each. Two pass: sorting incoming customer emails into the right queue, and checking supplier invoices for missing purchase-order numbers. The lead names an owner for each and starts recording a baseline: how long each case takes from arrival to the correct next step, and how many are sent back. Nothing is built yet. The output is two pages.

Days 15-30: criteria, oversight and access. The lead meets each owner and fills in the criteria page: one measure, a baseline, a target and a stop condition. Oversight is set at "asks first" for both, with a primary and a backup reviewer for each and a named person who can switch the agent off. They settle access: which mailbox and fields the agent may read, and which it may never touch. The gate is set for Day 75, with a named decision maker. Any link between two existing systems is raised now, because systems integration work often takes longer than expected and should not surprise the pilot.

Days 31-60: pilot at "asks first". The agent runs on live cases but proposes rather than acts. Each proposal is approved or corrected by the reviewer, and every correction is logged with a short reason. The owners read the measure weekly. A reviewer notices the agent keeps misreading one invoice type, so those invoices go down the exception path to a person and the pattern goes into the log.

Days 61-90: review, gate and decide. The team compiles results against the criteria, the exceptions list, a proposed oversight level, an owner after launch and a rollback plan. The gate meeting takes place on Day 75. The evidence supports moving email sorting to "acts and reports" with weekly sampling. For invoice checking, the decision maker chooses "go with changes": stay at "asks first" and fix the misread invoice type before the next review. Days 76-90 put those decisions into practice and set the next review date.

07What rules and risks should you check in your market?

Check which rules apply where you operate and where your customers are, and read them before the pilot rather than after. Two references are worth knowing as of November 2026: the voluntary NIST AI Risk Management Framework in the United States and the EU AI Act in Europe. This is general information, not legal advice.

According to the US National Institute of Standards and Technology, its AI Risk Management Framework "is intended for voluntary use" to improve the ability to build trustworthiness into the design, development, use and evaluation of AI products, services and systems. NIST released version 1.0 on January 26, 2023, and also publishes a Playbook of suggested actions, which it describes as "neither a checklist nor set of steps to be followed in its entirety". For a first pilot, the useful idea is to name who governs the AI. Your owner-after-launch and gate decisions do that.

The European Commission describes the EU AI Act, Regulation (EU) 2024/1689, as the first comprehensive legal framework on AI worldwide, with four levels of risk. As of November 2026, its page says the Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with some exceptions. Prohibited practices and AI literacy obligations applied from 2 February 2025, and rules for high-risk systems in certain sensitive areas apply from 2 December 2027 after the "AI Omnibus" changes. High-risk systems need "appropriate human oversight measures". Dates change, so read the page itself before relying on any of this.

The OECD's AI Principles, adopted in 2019 and updated in May 2024, add a third reference. They include transparency and explainability, robustness, security and safety, and accountability. They are not law, but they give you vocabulary for questions a customer or board may ask.

Three habits follow for a first pilot. Keep a record: the log from the "asks first" phase doubles as documentation of how the system behaved. Keep a person accountable: a named owner after launch is the simplest answer to "who is responsible?" And ask whether your use is high risk: employment, credit and education draw more attention, so if either process touches one of them, take advice before starting.

Following this guide does not make you compliant with anything, and nothing here says ProblemSolver holds a certification. We are pointing to public sources.

08Should you start with a partner, your own team, or wait?

Start with your own team if a process owner has time and someone can configure and monitor the tools. Start with a partner if you lack that capacity or your processes cross several systems. Wait only if you cannot yet name two processes and their owners.

Whichever you choose, the four decisions stay yours. A good partner asks for those decisions early. A poor one starts with the tool.

If you want to talk your situation through, you can Book a 30-minute call.

09What should you ask your team before Day 1?

Ask your team seven questions and write down the answers. They surface the facts that decide whether a process is a good candidate, and they are easier to ask now than in the middle of a pilot.

What to ask your team

1. Who owns each process today? Name one person per process, not a department.
2. Where is the work recorded? Which system holds the inputs and outcomes, and is the record complete?
3. What does a wrong result cost? In time, money, trust or a legal position, and how quickly would we notice?
4. Who answers when the AI asks? Name a primary and a backup, and how fast they can reply.
5. What do we measure now? If nothing, what could we record for two weeks to make a baseline?
6. What must never be automated? List the decisions, customers or data that stay with people.
7. Who can switch it off? Name the person, and confirm they know they may.

10Which decisions can you take this week?

You can take four decisions this week, and each fits on one line. None needs a budget or a vendor. Together they turn "we should do something with AI" into a plan with an end date.

1. Name your two processes. Write each one, its owner, and why it passed the five-part test.
2. Choose one measure per process. State what you will measure, and book time in the next two weeks to record a baseline.
3. Set the starting oversight level. For both processes, write "asks first", plus the reviewer, the backup and who can switch it off.
4. Set the gate. Pick Day 60 or Day 90 and name the decision maker.

If you would like a second pair of eyes on your four decisions before Day 1, you can Book a 30-minute call with us. We will go through your pair, your criteria and your gate, and you leave with a list of what to fix, whether or not you work with us afterward.

11Frequently asked questions

Sources and further reading

1. AI Risk Management Framework. NIST (US National Institute of Standards and Technology), 2023.
2. NIST AI RMF Playbook. NIST, 2023.
3. AI Act. European Commission, Shaping Europe's digital future, 2024.
4. OECD AI Principles. OECD (OECD.AI), 2024.