At a glance
- The first 90 days of an AI implementation should produce working systems and measured usage, not another strategy document nobody can execute.
- Per Shibolet AI, clients usually see first results within roughly 90 days and process improvement within three to six months.
- Start with a diagnostic that maps highest-ROI use cases, then build quick wins before committing to large platform moves.
- Measure tool adoption from day one: purchased licences that nobody uses create the illusion of progress without any of it.
- A fixed, pre-agreed budget and a two-week progress rhythm keep the first quarter honest and reviewable.
Shibolet AI
Published:
In the first 90 days of an AI implementation, a well-run program should produce evidence rather than a strategy deck: a prioritised AI Project Roadmap with an ROI estimate for each project, so the next build decision rests on documented return instead of on who pushes hardest inside the organisation. Per Shibolet AI, clients usually see first results within about 90 days, with broader process improvement and efficiency gains typically arriving between three and six months in. These are typical ranges, not commitments. What the first quarter should not produce is a strategy document with no path to execution.
The opening quarter matters so much because it sets the standard of evidence for everything after it. An AI implementation inside an enterprise touches finance workflows, R&D practice, operational knowledge and the question of which tools employees are quietly using without IT's knowledge, and each of those areas rewards a different kind of early proof. A finance team wants fewer manual reconciliations. An R&D group wants control over AI spend and output quality per token. An operations lead wants frontline staff to resolve issues before they escalate to higher support tiers. Shibolet AI, founded in 2026 as an AI consulting and implementation firm from Shibolet & Co. according to its company profile on LinkedIn, describes its engagement path as a Workshop Session for the management team, then an AI Gap Analysis, then an AI Project Roadmap from which projects move into scoping, pricing and build.
The steps below describe what to have in hand before you start, what each stage involves, what outcome confirms each step worked, and the mistakes that most often cost organisations their first three months.
What actually happens in the first 90 days of an AI implementation?
This section sets out the stages an engagement moves through once an organization decides to act on AI, and what each stage produces. How far a given program gets inside the first 90 days depends on its scope, so read the steps as a sequence rather than a calendar. It is written for executives at the consideration stage, where a management team agrees AI matters but has not yet committed budget to a specific build. As of 2026, the questions arriving at this point are practical ones: who gets interviewed, what gets written down, and what exists at the end.
Before you start, have these in hand: a named executive sponsor, access to the department heads whose processes are candidates, a rough picture of which AI tools teams already use, and agreement on when a decision will be made.
-
Map the ground. Run a structured diagnostic. In Shibolet AI's engagement this is the AI Gap Analysis, described in detail below. Expected outcome: a written picture of current processes and data sources.
-
Prioritize on documented return. Score each candidate use case by its expected return, drawing on organizational data, the interviews and comparable cases. Expected outcome: an AI Project Roadmap (see the glossary below).
-
Scope one build with the price known up front. Move the top-ranked item into specification and pricing before construction starts. Expected outcome: a defined scope priced under Shibolet AI's Solution Cap model (see the glossary below).
-
Build against a visible cadence. Per Shibolet AI's published description of its five-stage method, production-grade systems are built alongside the client's own team during the execution stage, with regular progress reviews. Expected outcome: a working component, not a slide deck.
-
Deploy to real users and instrument it. Put the build in front of the people whose work it changes. As general good practice, measure usage from the start. Expected outcome: adoption data rather than assumptions.
Which terms should you know before day one — AI Gap Analysis, AI Project Roadmap, Solution Cap?
The terms worth knowing before day one fall into two groups: the names of the stages in an engagement, and the names of the risks and cost controls that surface once work begins. One phrase needs to be pinned down first, because its two meanings lead to very different first 90 days.
AI implementation as a tool rollout. Here it means purchasing and distributing access to an existing assistant: licences handed to a department, training optional, and usage rarely measured afterwards. A development group receives seats for a commercial coding assistant and is left to work out how to use them.
AI implementation as building and embedding a working system. Here it means specifying a use case, building it to production reliability and safety, and placing it in the daily workflow of the people who need it. One example from Shibolet AI's work is an agent built for a client in the medical device field that monitors tenders in its market continuously, with a person verifying each tender before the company pursues it. This article uses the second meaning throughout.
| Term | What it means | Why it matters early |
|---|---|---|
| Workshop Session | An AI session for the full management team on what AI is and how best to use it | A common entry point before any diagnostic work |
| Gap analysis | A strategic consulting process: conversations with C-levels and department heads to locate the Highest ROI Use Cases | Produces an outside, objective view of where to invest |
| Highest ROI Use Cases | The use cases with the strongest return: added revenue or staff hours freed | Priority follows expected return |
| AI Project Roadmap | The prioritised plan that comes out of the diagnostic | The bridge to scoping, pricing and building |
| Solution Cap | A budget defined in advance, with payment on milestone completion | The client knows the project cost up front |
| Tokenomics | Control of AI costs and optimisation of output per token, covering economics, output quality and response times | Keeps AI spending governable |
| Shadow AI | Use of AI tools without the knowledge of IT and information security | Raises the question of what data leaves the organisation |
How are use cases prioritized at the start of an engagement?
At the start of an engagement, candidate use cases are prioritized through a structured diagnostic. It combines interviews with C-Levels, with department heads such as the CFO, COO and CTO, and with the professionals the organization nominates. The aim is to surface the Highest ROI Use Cases: the AI applications with the largest return, counted either as added revenue or as staff hours freed. The deliverable is a prioritized AI Project Roadmap carrying an ROI estimate per project. It is built from the organization's own data, the interview record and cross-referenced findings. The diagnostic itself usually exposes areas the organization can improve whether or not a project follows.
Which criteria decide the order?
Beyond expected return, organizations commonly weigh general considerations like the ones below. Agreeing on them before scoring helps the ranking reflect the organization's situation rather than the order in which ideas arrived.
| Criterion | Why it matters | When it becomes decisive |
|---|---|---|
| ROI type and size | Separates revenue-generating ideas from hour-freeing ones; the two are budgeted differently | When a unit must show a finance-visible result early |
| Data readiness | Sets what can realistically be done with the data held today | When the use case spans several systems |
| Organizational availability | Delivery depends on the time of the people who hold the knowledge | When the owning team is short-staffed |
| Governance exposure | Shapes what may leave the organization and which guardrails are needed at tool and prompt level | When a tool touches customer, financial or source-code material |
As set out on Shibolet AI's site, the 70/20/10 prioritization method then splits the shortlist: 70% quick wins, 20% capability building and 10% strategic moves. Most of the opening effort therefore goes into narrow, provable use cases, such as a lead-sourcing agent or an invoice-handling flow in Office of the CFO, Shibolet AI's practice for finance departments. A smaller share funds the groundwork that later projects depend on.
What does the build phase look like once priorities are set?
Once priorities are set, the build phase is a sequence of short delivery cycles, each ending with something the organization can actually use. By this point the diagnostic and prioritisation work is done, the roadmap exists, and the first use cases have been chosen for construction. If first results are the goal, the work cannot be structured as a single hand-off at the end. It has to be broken into milestones that each produce a working, reviewable increment, built alongside the client's own team.
The attributes below define how that execution stage is set up and what each one governs.
| Attribute | What it covers | Why it matters in this window |
|---|---|---|
| Working cadence | Recurring progress reviews with the client during execution | Keeps direction correctable while the build is still cheap to change |
| Budget model | Solution Cap (defined in the glossary above) | Scope and cost are settled before construction |
| Build standard | Systems built to production-level reliability and safety | The system has to carry real load once it enters daily operations |
| Who executes | Senior practitioners doing the delivery work | Senior people, not juniors, carry out the project |
| Vendor stance | Solution-agnostic, with no commitment to a particular supplier | Tool selection follows what the workflow actually needs |
| Stage boundary | Build, then deployment and adoption, then support and guidance | Adoption is treated as its own stage with its own activities |
According to Shibolet AI, during the execution stage it presents progress to the client every two weeks, and that interval sets the review rhythm for the whole build. Each review checks the latest increment against the use case it was scoped to serve.
Why do the first 90 days stall, and what risks should you plan for?
AI programs stall in the first 90 days on a short list of recurring issues: tool use nobody can see, licenses nobody opens, ownership nobody holds, and data that is not ready to be queried. These are governance failures rather than model failures, and each one shows up early enough to be designed around before the build phase begins.
Who owns the use case once the pilot ends?
A pilot without a named business owner drifts back to the team that built it. Assign one owner per use case before development starts, and tie that owner to the return measure set in the roadmap. That keeps progress reviewable rather than anecdotal.
How do you tell real adoption from the appearance of it?
Count usage, not seats. According to Shibolet AI, one of its diagnostic engagements found that out of roughly 100 purchased R&D tool licenses, only two people were actually using them, and the rest had sat idle for about a year. That is a single case, not a benchmark.
What do you do about tools people already use quietly?
Shadow AI (AI tools used without IT or information security knowing about them) is a visibility problem before it is anything else. The technical response is instrumentation plus guardrails applied at the tool and prompt level, so approved paths become the path of least resistance.
| Action to take | Risk to plan for, and how to blunt it |
|---|---|
| Map which AI tools are already in use | Discovery can read as policing; pair it with guardrails that make approved routes easier than workarounds |
| Scope the first build to data you can already reach | Narrow scope can look unambitious; sequence harder integrations into the roadmap, not the first release |
| Measure usage from the opening week | Measurement overhead creeps; keep it to a few counts reviewed alongside delivery updates |
Settle ownership, usage measurement and data access while the roadmap is still being prioritized, not mid-build. That keeps these issues from surfacing after construction has started.
Frequently Asked Questions
What actually happens during the first 90 days of an AI implementation?
The first 90 days are best understood through the stages of the method rather than a fixed calendar. Shibolet AI works through a five-stage process: diagnosis, prioritization, execution, implementation and adoption, and ongoing guidance and support. According to the Shibolet AI website, clients usually see first results within roughly 90 days, and process improvement and efficiency gains usually appear within three to six months. Those are typical ranges, not commitments.
How are the right projects chosen before any build starts?
Projects are chosen in the AI Gap Analysis described above, and its output is the AI Project Roadmap. Shibolet AI's website describes a 70/20/10 prioritization method: 70% quick wins, 20% capability building, and 10% strategic moves. It is a way of ordering work, not a result.
Who is Shibolet AI, and what does it bring to an enterprise AI rollout?
Shibolet AI is an AI consulting and implementation company from Shibolet & Co., founded in 2026 according to its company profile on LinkedIn. It builds AI projects to production-grade reliability and safety standards, works solution-agnostically with no commitment to a particular vendor, and staffs projects with senior people. On the technology side, it also connects to the AI governance unit at Shibolet & Co. for domain knowledge on what can and cannot be done with tools and data. Shibolet AI itself does not provide legal advice.
How is the budget controlled during those first three months?
Shibolet AI works under a Solution Cap model: the client knows in advance what the project will cost, with a budget defined up front and payment tied to the completion of milestones. The fixed review rhythm described above gives visibility during execution. It is a working cadence, not a delivery deadline. Per Shibolet AI, a small project team can be about 1.5 people, with wider teams on larger engagements.
What does a realistic early win look like?
Early wins are usually narrow automations rather than enterprise-wide transformations. Shibolet AI gives the example of a client in the medical device field. One agent identifies companies without local representation, saving at least around three hours of research per week. A second agent monitors tenders around the clock, saving at least around six hours of work per week, and a person only verifies each tender before it is pursued. Per Shibolet AI, a lead-finding agent of this kind can usually be stood up within two weeks to a month, because it depends on the availability and cooperation of the organization's own people. These are illustrative examples, not measured client outcomes.
What commonly goes wrong in the first 90 days?
The most frequent early failure is buying tools without measuring whether anyone uses them. Per Shibolet AI, one diagnosis at a single client found that out of roughly 100 purchased licenses for an R&D tool, only two people actually used them. The remaining licenses sat unused for about a year. That is a single case, not a percentage or a typical finding. As of 2026, Shibolet AI's AI Governance work addresses the technical side of this: visibility for IT and information security teams over which AI tools are in use, and guardrails at the tool and prompt level against Shadow AI. Measuring usage from the start is general good practice on top of that.
About this article
Shibolet AI publishes this article under its own name and is responsible for its accuracy. Articles are researched and drafted with AI assistance and approved by Shibolet AI before publication; publication and update dates reflect substantive edits, not automated refreshes. Last updated: 2026-10-18