Category
The 90-Day AI Implementation Plan
Last updated: June 4, 2026
If you are the operator who just got the green light on AI and is expected to land week one on Monday, this plan is for you. Generic 90-day templates fall apart at week three: the team picks three workflows, the security review hits in week eight with a data-residency clause nobody priced in, and day 90 closes with a vendor short-list and nothing in production. In this guide, you'll get the three 30-day blocks (Assess, Build, Operate), the deliverables that have to ship each week, the gate questions that decide whether you advance, and the four anti-patterns to refuse on day one.
A scoped single-workflow agent reaches production in 6 to 10 weeks of build, or 8 to 12 weeks when the deployment is private; that operating range is what makes a 90-day calendar realistic instead of optimistic. We use what we sell, including the agents that run Arkeo itself.
Quick Answer
• What it is: Three 30-day blocks: Assess, Build, Operate. Each block has deliverables, a gate question, and a failure mode.
• What ships at day 90: One scoped workflow agent in production behind a human-in-the-loop owner. Not a fleet, not a platform, not a strategy deck.
• Timeline reality: A scoped single-workflow agent reaches production in 6 to 10 weeks, or 8 to 12 weeks private or enterprise-grade.
• Why it matters: Most pilots stall because no block had a clear gate and no owner was named before the build started.
Before day one, run the upstream version of Assess. A free AI Assessment compresses two weeks of current-state mapping into one session, so the team walks into week one knowing the workflow, the data path, and the owner.
Why do generic 90-day AI implementation templates fail?
Open any popular "90-day AI plan" and you will find the same outline: weeks 1-4 stakeholder interviews, weeks 5-8 vendor selection, weeks 9-12 pilot. It looks structured, but it treats AI implementation as a procurement exercise instead of a workflow-and-data exercise. By week 12, the team has chosen a vendor and shipped nothing.
The failure rate confirms it. The Deloitte State of Generative AI Wave 4 study found more than two-thirds of enterprises expect 30 percent or fewer of their generative AI experiments to fully scale in the next three to six months. The real first question is not "which vendor?" It is "which workflow, on which data, owned by which human?" Answer those three in days 1-30 and the next 60 days are a build problem; answer them wrong and the next 60 days are a meeting problem.
What does the three-block AI implementation plan actually look like?
The plan is three 30-day blocks. Each block has a deliverable list, a gate question, and a failure mode. The gates are not bureaucratic checkpoints; they are the moments when the program either earns the next block or pauses to fix what the last one missed.
THREE 30-DAY BLOCKS
Assess, Build, Operate
Each block has deliverables, a gate question, and the failure mode that halts progress.
DAYS 1-30
Assess
Ships: a one-page current-state map, a shortlist of three workflows with the top one chosen, the data-path decision (public-cloud-safe vs private), and one named operational owner.
GATE QUESTION
Do you have one workflow, one owner, and one data path agreed in writing?
FAILURE MODE
The team picks three workflows and tries to build all of them in parallel.
DAYS 31-60
Build
Ships: the workflow agent running end-to-end in a controlled environment, a passed security and access review, an integration plan against systems of record, and a small library of test cases.
GATE QUESTION
Does the agent run end-to-end on real data, with security signed off?
FAILURE MODE
Security review surfaces a data-residency clause that forces a re-platform.
DAYS 61-90
Operate
Ships: the agent in production behind a human-in-the-loop checkpoint, the operator trained and on-call, the operating metrics defined, and the first weekly review run.
GATE QUESTION
Is the agent in production with a named owner and a weekly drift check?
FAILURE MODE
Launch happens, then nobody is on the hook. The agent silently degrades.
Skipping a gate is faster than fixing what skipping it produces.
The grid is the shape. The next sections are what each block actually ships, in operator language.
What ships in the first 30 days of an AI implementation?
Days 1-30 are the Assess block. Four things ship: a one-page current-state map, a shortlist of three workflow candidates with the top one chosen, a data-path decision, and one named operational owner. No vendor evaluation, no platform selection. The block answers three questions: which workflow, on which data, owned by which human.
The right first workflow is repetitive, high-volume, and rides on data you can already get to. Picture a 220-person specialty insurer in week two of Assess: three candidates on the whiteboard, the claims-summarization workflow looked sexiest, and the answer turned out to be the underwriting submission triage workflow because the data already lived in one system and the operations lead actively wanted the agent to work. Alignment beats model selection.
The data-path decision, skipped, kills the build phase. Decide in days 1-30 whether the workflow is safe on a hosted public model or whether contracts, sensitive data, or audit posture force a private deployment. This is Arkeo's wheelhouse: a private AI workforce where the data never leaves the building. The IBM Cost of a Data Breach 2025 report found 97 percent of organizations that suffered a breach of an AI model or application lacked proper AI access controls, with shadow AI use adding $670,000 per incident. The data path is a financial decision, not an IT detail.
The named owner is the deliverable most often forgotten. Not a sponsor, not a champion: an operational owner who runs the agent after launch, named in writing, with explicit authority to halt at any gate. If no name fits the box at day 30, the program halts there until one does.
WEEK-BY-WEEK
Week 1: Current-state interviews with the operators who run the candidate workflows today, plus a tool inventory of every system the work touches.
Week 2: Shortlist four to six workflow candidates and score each on pain level and data accessibility.
Week 3: Data-path decision for the top candidate, including a written sensitivity classification (public-cloud-safe or private).
Week 4: Name the workflow, appoint the human owner in writing, and prepare the day-31 build kickoff.
What happens between days 31 and 60 in the Build block?
The Build block is short, narrow, and unforgiving. Four deliverables: the workflow agent running end-to-end in a controlled environment, a passed security and access review, an integration plan against the production systems of record, and a small library of test cases the operator can run on a schedule.
The trap in Build is scope expansion. The first agent is meant to do one workflow well, not three workflows passably. The PwC AI Agent Survey of 300 senior US executives found 66 percent of organizations adopting AI agents report measurable productivity gains; the gain comes from one well-scoped agent shipping.
The IBM IBV CEO Study of 2,000 CEOs across 33 countries cited "lack of expertise" as the top barrier to AI innovation. Mid-market teams almost never have a full AI engineering bench, so Build is typically staffed by a senior internal lead plus a build partner who carries the operator role through the first 90 days. In Arkeo's build experience, a scoped single-workflow agent runs about $15,000 to $40,000 and 6 to 10 weeks to production, or 8 to 12 weeks when the deployment is private. Arkeo runs its own operation on the same private agents it deploys for clients; we use what we sell.
Security review is the gate that catches a data-path mistake. Picture a Build engagement where the week-seven security review surfaces a permission scope on a hosted model that was set too wide during the controlled-environment scaffolding; the team has to unwind the scope, re-run the access audit, and lose four days before re-entering the gate. A halt at day 50 is recoverable. A halt at day 110, after the workflow is live, is not.
WEEK-BY-WEEK
Week 5: Controlled-environment scaffolding plus a first integration test against the systems of record.
Week 6: First agent prototype running on real data, with an internal review against the test cases written in Assess.
Week 7: Security review and access-control audit, including data residency and permission scopes.
Week 8: Acceptance test against the gate question (end-to-end on real data, security signed off) and a fix list for any items the test surfaces.
Build also sets the boundary on what the 90-day plan will not attempt. The four anti-patterns below kill 90-day plans; they belong in months four through twelve.
DO NOT ATTEMPT IN THE FIRST 90 DAYS
Four anti-patterns that kill the plan
Each looks ambitious. Each blows past day 90 with nothing in production.
ANTI-PATTERN 1
Replacing a major system
Swapping the ERP or CRM as part of the AI rollout. The integration alone is a multi-quarter project; bundling it with the first agent guarantees neither ships on time.
ANTI-PATTERN 2
Agentic workflows across 5+ teams
Cross-functional agents that touch five or more departments. The change-management coordination overruns the build calendar before any production run completes.
ANTI-PATTERN 3
Custom model training from scratch
Training a proprietary foundation model when a fine-tune or a well-prompted off-the-shelf model would do. The first 90 days are for deploying value, not for research.
ANTI-PATTERN 4
Public-facing AI features
Shipping an external chatbot, customer-facing agent, or marketing-site assistant in the first 90 days. The brand risk lives at zero internal QA hours. Operate internally first.
These belong in months 4-12, not days 1-90.
Pressure-test the plan before you commit a quarter to itThe free AI Assessment is a 60-minute working session that pre-runs the data-path decision, the security-review surface area, and the human-in-the-loop owner question, so the failure modes above hit the spreadsheet on day zero, not day 50.
Book Your Free AI Assessment →
What does the Operate block actually deliver by day 90?
Operate is the block most plans skip. The slide deck ends at "launch." The deployed agent then drifts: models update, data shapes change, a feeder workflow gets re-platformed, and the agent quietly produces worse outputs for weeks before a customer notices.
Four deliverables ship in days 61-90: the agent in production behind a human-in-the-loop checkpoint, the operator trained and on-call, the operating metrics defined, and the first weekly review run. Production does not mean autonomous; it means the workflow runs every day with a human reviewing outputs against the test cases built in Build. This is the Manage half of the Assess to Deploy to Manage model Arkeo runs against.
Picture a regional logistics operator at day 75: the triage agent is running, three escalations a week, the operations director has the dashboard open every morning, and one test case quietly fails on a Tuesday because a feeder system updated its date format. The director sees it on the weekly review, the build partner ships a fix, and the program never loses a day. That is what the Operate block buys.
WEEK-BY-WEEK
Week 9: Cutover plan plus human-in-the-loop training for the named operator.
Week 10: Production turn-on with the human reviewing every agent action against the test cases.
Week 11: Operating metrics live (latency, accuracy, drift) and the first weekly review run on real production data.
Week 12: Month-twelve readiness review: does the workflow survive without the build partner in the room?
BCG's Where's the Value in AI? (October 2024) found 74 percent of companies struggle to capture value from AI even as adoption climbs. The 26 percent who do share one trait: they ship the Operate block. The deeper detail lives on AI implementation challenges, and the year-long arc beyond day 90 lives on the 12-month AI roadmap.
What does day 90 honestly look like?
At day 90, the program owns one workflow agent in production. One workflow, one named operator, one weekly review rhythm, one documented failure-mode list, one set of operating metrics. It sounds small. It is the foundation everything else gets built on.
Context makes the discipline matter more, not less. The Stanford HAI 2025 AI Index reports 78 percent of organizations used AI in 2024, up from 55 percent the year before, the largest single-year jump in the Index's history. The plans that survive that pressure shipped one production agent at day 90 and earned the political room for the next two by month six. The plans that did not tried to ship five and shipped zero. This 90-day cadence sits inside a broader enterprise AI strategy that sequences the work beyond day 90.
Walk into day one already knowing the workflowA 60-minute free AI Assessment with an operator who has shipped this plan turns the Assess block into a single session, so the 90-day clock starts with the gate already cleared.
Book Your Free AI Assessment →
