The first 90 days of an enterprise AI program: a week-by-week checklist

This checklist puts the first 90 days of a large company's artificial intelligence (AI) program in order, week by week, with the reason and the source for each line. Print it, or save it as a PDF with the button on this page, and tick each line. It is general information as of 26 October 2026: ask your lawyers which rules apply to your company. It is part of the guide on enterprise AI adoption.

Weeks	Work	Page with the detail
1 to 2	Inventory, owners, switch-off, training record	AI governance
2 to 4	The rules and dates that apply	AI governance
3 to 6	Two or three first uses	Choosing your first AI use cases
4 to 8	Data and older systems	Getting company data ready; connecting AI to older systems
4 to 10	Security	AI security risks
6 to 12	Build, buy or partner	Build, buy or partner
8 to 12	Measure	Measuring what AI returns
All weeks	Training	AI training for staff

Weeks 1 to 2: inventory and owners

List every AI system in use. The US National Institute of Standards and Technology (NIST) AI Risk Management Framework asks, in item GOVERN 1.6, that "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities" [1]. Include tools staff use on their own, AI features inside software the company already buys, and systems built in-house. In the US Census Bureau's AI supplement, collected 17 November 2025 to 8 February 2026, 22.9% of respondents at firms with 250 or more employees did not know whether their firm had used AI in the last two weeks [2]. Because of Executive Order 14434, signed on 29 September 2026, the Census Bureau now calls AI "Super Intelligence" or "SI" in its files [3][4]. Its questions asked about "Artificial Intelligence (AI)" until the collection that began on 19 October 2026, after this supplement closed, so this checklist says AI [4].

• One list of AI systems, with purpose, vendor and data used for each.
• A named owner for each system.

Name who can switch each system off. Item MANAGE 2.4 asks that responsibilities are "assigned and understood, to supersede, disengage, or deactivate AI systems that demonstrate performance or outcomes inconsistent with intended use" [1].

• A named person per system who can stop it, and a written rule for when.

Set up the four functions. NIST's framework has four functions, govern, map, measure and manage, and NIST says "Actions do not constitute a checklist, nor are they necessarily an ordered set of steps" [1]. Use them as headings for the program's policy.

• A short AI policy, under the four headings, approved by the executive team.

Open the training record. The European Commission says: "There is no need for a certificate. Organisations can keep an internal record of trainings and/or other guiding initiatives" [5].

• A training record started, with the date and the people for each session.

Weeks 2 to 4: the rules that apply

EU AI literacy duty. As of 26 October 2026, Article 4 of the EU AI Act requires providers and deployers to "take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf", in the text set by Regulation (EU) 2026/1744, in force from 27 July 2026 [6]. "The national market surveillance authorities will start supervising and enforcing the rules as of 2 August 2026" [5].

• A training plan for staff and contractors who operate AI for the company.

EU high-risk systems. High-risk duties apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems [6]. Annex III includes credit scoring of people and the pricing of life and health insurance [7].

• Each system on the list marked high-risk or not, with the reason.
• For high-risk systems used at work: a plan to inform workers before use, under Article 26(7) [7].
• For public bodies, providers of public services, and credit or insurance uses: a fundamental rights impact assessment planned, under Article 27 [7].

EU marking deadline. Providers of generative AI systems "that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026" [6].

• If the company provides such a system, a plan to mark its output by 2 December 2026.

Colorado. SB26-189 became law on 14 May 2026 and covers automated decision-making technology in consequential decisions about education, employment, housing, lending, insurance, health care and government services [8]. From 1 January 2027 developers must give deployers technical documentation, and both must keep compliance records "for at least 3 years" [8]. A deployer must explain an adverse decision to the consumer within 30 days [8]. This checklist is based on the bill summary; confirm the deployer start dates with counsel. A federal court order of 27 April 2026, in X. AI LLC v. Weiser, stops the attorney general from starting enforcement of the 2024 law or a law replacing it for violations that occur up to 14 days after the court rules on a preliminary injunction motion, and the court record read on 26 October 2026 showed no change to that order [20].

• Each system that makes or shapes a consequential decision about a Colorado consumer listed, with its vendor's documentation requested.

US federal suppliers. OMB memo M-25-22 binds US executive agencies buying AI [9]. A company that sells AI to them should read its contract terms.

• If the company sells AI to US agencies, M-25-22 terms reviewed.

Weeks 3 to 6: pick two or three first uses

See where large firms already use AI. Among US firms with 250 or more employees, the functions where the most firms used AI were information technology (25.9%) and finance and accounting (20.5%) [2].

Check the reasons others stop. Large EU enterprises that considered AI and did not use it most often named lack of expertise (65.11%), data protection concerns (58.76%) and unclear legal consequences (51.22%) in 2025 [10].

Test each task. In a field experiment with 758 consultants, those using AI on a task chosen to be outside AI's abilities were 19 percentage points less likely to produce correct solutions [11]. In a study of 5,172 support agents, gains were largest on moderately rare problems [12].

Write down purpose, benefits and the cost of errors. NIST asks for these under MAP 1.1, MAP 3.1 and MAP 3.2 [1].

• Two or three candidate uses scored on frequency, cost of an error, testability, data, rules, systems and owner.
• For each chosen use: purpose, benefits and the cost of errors written down.
• For each chosen use: a set of past cases with known right answers, to become tests.

Weeks 4 to 8: data and older systems

Check the data. NIST warns that the data used to build an AI system "may not be a true or appropriate representation of the context or intended use", and that "the ground truth may either not exist or not be available" [1]. Ground truth is the correct answer for each case.

• For each use: where the data lives, who may use it, and how recent it is (NIST MAP 2.3 [1]).
• Test sets and metrics documented (NIST MEASURE 2.1 [1]).

Plan for drift and provenance. The May 2025 AI Data Security sheet from the NSA, CISA, the FBI and partner agencies recommends monitoring, retraining and data cleansing for drift, and asks companies to "Employ digital signatures to authenticate trusted data revisions" [13].

• A fixed test set to run on a schedule, with a trigger for review.

Connect through your own code. In the tool-calling designs of OpenAI, Anthropic and Google, the company's application runs the calls it defines, as the page connecting AI to older systems shows. If the company uses the Model Context Protocol, "Hosts must obtain explicit user consent before invoking any tool" [16].

• For each older system: a read-only interface first, and write actions only with a person's approval.

Weeks 4 to 10: security

Design for prompt injection. Prompt injection is text in the input that changes what the model does. NIST says "current mitigations do not offer full protection against all attacker techniques" [19]. OWASP, which publishes a Top 10 list of security risks for applications built on large language models, asks for least privilege and for "human-in-the-loop controls for privileged operations to prevent unauthorized actions" [14].

Limit what agents can do. OWASP's guidance on excessive agency asks companies to "Minimize extension permissions" [15].

Protect the data supply chain. The AI Data Security sheet names three data risks: "data supply chain, maliciously modified ("poisoned") data, and data drift" [13].

• Each system given only the tools and access its task needs.
• A person approves actions that change records, move money or send messages outside the company.
• Injection test cases written and run before each release.
• Usage and spending limits set for each system.

Weeks 6 to 12: build, buy or partner

Know the common routes. In 2024, 67.23% of large EU enterprises using AI bought commercial software ready to use, 34.05% developed AI in-house and 45.76% used external providers [10].

Put the buyer terms in the contract. OMB memo M-25-22 names "requirements regarding knowledge transfer, clear data and model portability practices, clear licensing terms, and pricing transparency" [9]. It also asks buyers to "Convene a cross-functional team" [9].

Watch the EU provider rule. Under Article 25 of the EU AI Act, a deployer that puts its name on a high-risk system, makes a substantial modification to it, or changes its purpose so that it becomes high-risk is treated as its provider [7].

• A route chosen for each use, with the reason.
• Knowledge transfer, portability, licensing and pricing terms in each contract.
• The company owns its test cases and its data under every contract.

Weeks 8 to 12: measure

Take the baseline first. A field study of 5,172 support agents found an AI assistant raised issues resolved per hour by 15% on average, and its authors say the findings "capture medium-run effects in a single firm" [12]. A company's own figure can differ.

Decide what saved time is for. In a Danish study, 85% of chatbot users said they moved time saved by AI to other job tasks [17].

Count the cost of errors. NIST's MAP 3.2 covers "Potential costs, including non-monetary costs, which result from expected or realized AI errors" [1].

• Output per hour, quality, time to competence and cost per task measured before release.
• The same measures repeated after release, with a comparison group where possible.
• A written decision on what saved time is for.

Every week: training

The Commission says "having different levels of training or learning approaches could be appropriate", and that simply relying on instructions for use "might be ineffective" [5]. NIST's GOVERN 2.2 asks that personnel and partners "receive AI risk management training" [1]. A Danish study found that use and reported benefits peak in workplaces that combine encouraged use with enterprise chatbots and training [17]. A report by the OECD, Boston Consulting Group and INSEAD singles out training on real projects, using the AI systems and datasets common in a given area of business [18].

• Training by group: executives, staff using chat tools, staff in high-risk processes, engineers, contractors.
• Training on the tools and data each group will use, before each tool goes live.
• Each session recorded with its date.

Where Reveneau fits

Reveneau is an AI software development consultancy with offices in San Francisco and Vancouver. Reveneau says AI writes all of its code, and every change must pass an eval suite written from the specification, the written description of what the software must do. An eval suite is a set of written, repeatable tests of a system's answers. Reveneau says it uses AI instead of hiring more engineers and puts that saving into clients' prices, and that it takes responsibility for the whole project through production and after release. For the evals behind the measure and security lines, see why AI evals matter. To talk about a project, see Reveneau's services or contact Reveneau.