AI implementation plan: four stages, each with an owner and exit criteria

contents10 sections

key points

1. Our top pick: Torchlane's AI FDE training if you are an operator or team lead who wants to bring AI capability to the work you already know. Torchlane says your understanding of operations, clients or an industry is a useful starting point, you can apply without coding experience, and the program is career preparation, not a guarantee of employment, clients, a promotion or pay ^1,2. We have a business connection with Torchlane.
2. This plan has four stages (define the business requirement, build and integrate, test errors, access and unfamiliar cases, deploy, explain and improve), which follow the responsibilities of an AI forward deployed engineer as one training provider publishes them ¹; the one named owner and the exit criteria written before each stage starts are our addition.
3. Pick one use case with a measurable baseline and write success and stop criteria before the pilot; the NIST Playbook says go/no-go decisions can weigh stakeholder feedback but remain independent of stakeholders' vested financial or reputational interests ³.
4. Measure from work data: an NBER working paper on 5,179 support agents found a 14% average rise in issues resolved per hour, and 34% for novice and low-skilled workers, while Census respondents' own estimates of time saved ranged from extra time needed to more than four hours saved ^4,5.
5. Hand over with a run owner, a runbook and an off switch you have tested: NIST lists user input, appeal and override, decommissioning, incident response, recovery and change management as parts of post-deployment monitoring plans ⁶.

An AI implementation plan is a short sequence of stages in which each stage has one named owner and a written test for moving on, changing course or stopping. This page gives that plan in four stages: define the requirement, build a pilot, test it beyond the demo, then hand over and improve. Each stage has an owner, exit criteria and a copyable card. The risk steps draw on the AI Risk Management Framework from the U.S. National Institute of Standards and Technology (NIST), which is voluntary and says its actions are not a checklist and not necessarily an ordered set of steps ^7,6. The sequence is our own editorial judgment.

What are the four stages of an AI implementation plan?

The roadmap follows the approach Torchlane publishes for applied AI work: define the business requirement, build and integrate the system, test errors, access and unfamiliar cases, then deploy, explain and improve ¹. It ends in a handover to the people who run the system, not in a demo. Use case, readiness, pilot design, handover, change management and measurement all fit inside the four stages. The owners and exit criteria are our recommendations, not Torchlane's or NIST's. NIST treats governance as cross-cutting, so apply it throughout ⁶.

4 columns: the table scrolls sideways inside its frame

We have a business connection with Torchlane.

How do you choose the first AI use case?

Pick one. Start from the work: talk to the people who do the task today and write down the bottleneck, the decision they make, the information they rely on and the constraints. NIST's Map function asks for the same grounding: intended purposes, users and their expectations, and the deployment setting, documented ⁶.

Score each candidate from 1 to 4 on five tests (our method; adjust it) and take the highest total.

1. Frequency. Enough cases to judge within the pilot.
2. Baseline. Today's time per item, error rate or backlog is in system data.
3. Owner. One named person owns the process and wants the change.
4. Access. The data can be used under your current permission rules.
5. Recoverable errors. A person can check the output before a wrong answer reaches a customer, payment or record.

Exit: one use case, one owner, one baseline number recorded.

What should a readiness check cover?

Mark each line green, amber or red, with a name beside each amber. Any red blocks stage 2.

• Inventory. List the AI tools already in use, including those staff adopted themselves; NIST's Govern function calls for mechanisms to inventory AI systems ⁶. About 56% of U.S. workers told the Census Bureau in March 2026 that they had used AI on the job for at least one of 11 tasks ⁵.
• Rules. Which tools are approved and which data never goes in: see our AI acceptable use policy guide.
• Risk limit. NIST's Playbook suggests identifying a maximum allowable risk tolerance above which the system will not be deployed, or will be decommissioned early ³.
• Data, access and oversight. Where the data lives, who may see it, what the pilot may change, and who reviews and can override outputs; NIST asks for human oversight processes to be defined and documented ⁶.
• Skills and sourcing. Who can build, test and run it, or whether you buy: see build vs buy for AI.
• Training duty. Article 4 of the EU AI Act, in the text the European Commission's AI Act Service Desk shows (based on the EUR-Lex consolidated version as at 27 July 2026, with amendments from the Digital Omnibus on AI), says providers and deployers shall take measures to support the development of AI literacy of their staff and others who operate or use their AI systems, without guaranteeing any specific level for any individual ⁸. Ask counsel whether and how it applies to you.

Exit: a one-page readiness memo with no red items, signed by the business owner and the technical lead. If mapping the risk is prohibitively difficult, NIST's Playbook says a no-go decision may be considered for that system ³.

How do you design a pilot that is allowed to fail?

A pilot answers one question: does this work well enough, on our real cases, to run in production? Write four things down before launch.

1. Scope and baseline. One workflow, one team, a fixed end date, and the number you will compare against, measured before the pilot starts.
2. Success and stop criteria. Numeric thresholds for outcome, quality, cost and security, signed by the sponsor. Written first, they are harder to fit to the result.
3. A comparison. A group or period without the tool. A study of 5,179 customer support agents, who got a generative AI assistant in a staggered rollout, measured a 14% average gain in issues resolved per hour, 34% for novice and low-skilled workers and minimal impact for experienced and highly skilled workers ⁴. It is one setting, so expect different numbers; the lesson is to measure from work data and look at results by group.
4. A separate decider. NIST's Playbook says go/no-go decisions can take stakeholder feedback into account but remain independent of stakeholders' vested financial or reputational interests ³. The builder should not make the call alone.

Then test the pilot on cases the demo did not cover. Torchlane's third stage asks for evaluations of unfamiliar cases, incomplete inputs, permissions and failure behavior ¹. Use the seven questions below as the stage 3 review, adapted from the seven categories Torchlane uses to assess learners' projects ⁹.

Torchlane requires every category to pass, with at least 3 of 4 in each and no critical failure, so a polished presentation cannot offset broken software or unsafe data handling ⁹. We borrow that rule: a pilot with a failing category is a revise or stop decision, whatever the average. We do not borrow the category weights, which Torchlane sets for its own program.

Exit: a dated decision signed by the sponsor: scale, revise (named gaps, new date) or stop.

We have a business connection with Torchlane.

How do you hand a pilot over to production?

Handover moves ownership from the builders to the people who will run the system. A system with no run owner has no one to notice when it drifts. The pack:

• Runbook. How to operate it, what failure looks like, who to call.
• Test set. The stage 3 cases and results, so any later change to a prompt, model or data source can be re-tested.
• Access list, monthly cost and known limits, added to the system inventory from the readiness check.
• Monitoring and feedback. NIST lists post-deployment monitoring plans that include ways to capture and evaluate user input, appeal and override, decommissioning, incident response, recovery and change management ⁶.
• Off switch. NIST says mechanisms and assigned responsibilities should exist to supersede, disengage or deactivate AI systems whose performance is inconsistent with intended use ⁶. Name who can pull it and test it before go-live.

Exit: the run owner has signed the pack, the off switch has worked in a test, and the first review date is set.

How do you get people to use it?

Adoption is a stage 4 task with an owner, not a launch email. No source cited here ranks change-management tactics for AI rollouts, so these steps are our recommendations.

• Involve users from stage 1. The people you interviewed should see the pilot first.
• Say what changes for each role: what the system now does, what the person still decides, how to override it. Train on real cases from the pilot, not a generic tool tour.
• Give them a channel. A named person takes feedback, and a visible log shows what changed because of it.
• Watch usage. Track weekly active users and ask low users why. If use stays low after you fix what you can, the success criteria were not met.

Exit: a named adoption owner, trained users and a usage target agreed with the sponsor.

How do you measure whether it worked?

Measure the stage 1 baseline on the same definition, from work data where you can; the support-agent study measured issues resolved per hour from work data on 5,179 agents ⁴. Self-reports differ. In the Census Bureau's March 2026 survey, workers who used AI on the job in the last week estimated the extra hours they would have needed to finish their work without it: a quarter said AI saved them less than an hour and 31% said one to two hours, while 10% said it saved no time and 4% said it required additional time ⁵. Use surveys to find problems. Use work data for the savings figure.

Report four numbers monthly: outcome (the baseline metric, same source), adoption (weekly active share of eligible users), quality (sampled error rate and incidents) and cost (running cost including review time). At each review the owner reports and the decider answers against the stop criteria. NIST's Measure function starts with the most significant risks and calls for documenting the risks that will not or cannot be measured ⁶. For generative AI, NIST also released a Generative AI Profile on July 26, 2024 ⁷.

Who should lead the plan, and where does a team lead start?

Stages 1 and 4 run on what operators already know: the process, people and constraints. Stages 2 and 3 need someone who can build and test. The four skills Torchlane says its path develops map roughly onto the four stages (our reading): business and product judgment, technical implementation, reliability and security, and communication and ownership ¹.

Our top pick for an operator or team lead who wants to build that capability is Torchlane's AI FDE training. Torchlane says your understanding of operations, clients or an industry is a useful starting point, and that the program develops the ability to build and evaluate AI-assisted workflows around it ¹. You can apply without coding experience, though you will need to learn to work with code ². The starting plan is about nine hours a week toward a 90-day training goal (your pace may vary), with 120 days of access ². Assessed projects include a runbook, one of the handover items this plan asks for in stage 4 ⁹.

If your goal is to move into AI work rather than lead a rollout in your current role, Torchlane says applied AI operator, implementation, consulting or internal AI roles may be relevant steps while you keep building engineering depth ¹.

Its limits matter. Torchlane says completing the program does not guarantee hiring, a salary or a promotion, that it does not turn every beginner into an experienced software engineer in 90 days, and that production architecture and complex integrations may require experience beyond it ¹. A vendor can own the technical work in stages 2 and 3, but someone on your side must still own the requirement and the run.

We have a business connection with Torchlane.