The First Ninety Days: What Actually Ships

A ninety-day plan becomes an operating system only when decisions can move between the fractional leader's working days. The weekly unit has two constant parts: a leadership working session and an asynchronous decision queue. A delivery evidence review becomes a third part only while delivery is active. Those are three different jobs. Combining them creates a meeting where a vendor choice, a late data export and a failed test all compete for the same final ten minutes.

The leadership working session starts from the decision log, not a presentation. Each open item names the decision, the person who can make it, the evidence available, the evidence missing, the deadline and the consequence of waiting. The room closes or explicitly defers the highest-consequence items. An update that asks for no decision stays asynchronous. A delivery issue that can be resolved inside the approved scope stays with the delivery owner. This protects executive time without hiding blocked work.

The asynchronous queue is what makes part-time ownership real. A new request enters through one form, gets one owner, and is classified as decide, investigate, deliver, monitor or reject. The fractional leader can resolve decisions inside the written mandate. A risk acceptance, material budget change, new data class or scope expansion goes to the named executive or functional approver. Every decision lands in the log with its rationale and review trigger, so the next working day begins from the record rather than from memory.

The delivery evidence review looks at a small set of real cases, corrections, exceptions, access failures, cost and reviewer time. A demonstration is useful for communication but cannot close an acceptance condition. If there is no approved build in flight, cancel this review rather than inventing delivery theater. The risk register and adoption plan are reviewed alongside the workflow because an accurate system that nobody can safely use is not ready.

Monthly, the executive sponsor reviews the maintained roadmap, scorecard, risk register, vendor register and unresolved decisions. The purpose is to reallocate budget, stop work, change an owner or accept a documented risk. It is not a repeat of four weekly updates. At days 30, 60 and 90 that monthly review becomes a formal gate with a written decision.

Quarterly, day 90 is both an acceptance point and a reset. Leadership keeps, changes or stops each active workflow; re-ranks the remaining opportunity register; confirms the governance matrix; resets the next-quarter capacity and budget; and names the artifacts that transfer if the engagement changes shape. Nothing rolls forward merely because it was already on the roadmap. The next ninety days begin with a fresh decision record, not an extended promise.

The roadmap is the ordered portfolio, not a list of ideas. The Fractional Chief AI Officer maintains it; the executive sponsor approves material changes; and a row changes when new evidence alters value, feasibility, risk, capacity or timing. Every active item points to a decision, a functional owner and the next evidence due. Parked work names the condition that could bring it back. Rejected work keeps its rationale so it does not return under a new title next month.

The decision log is the memory of the operating system. The leader records the call, decision owner, options considered, evidence, rationale, date, consequence and review trigger. A decision changes only through a new entry that links back to the old one. Editing history until the original call disappears defeats the purpose.

The governance matrix names who proposes, reviews, approves, operates, monitors and can stop each use case. The relevant functional, security, privacy, legal, records and executive owners approve their own domains; a fractional leader does not manufacture authority over them. The matrix changes when the workflow, data, user population, consequence or governing requirement changes.

The scorecard keeps the baseline and the operating measures together: quality, direct cost, human time, adoption, exceptions and the business measure the sponsor actually uses. The measurement owner updates it on an agreed cadence and records missing data rather than filling gaps with estimates. The monthly review can change a target or method, but it must preserve the old definition and explain why comparison across the change is limited.

The risk register records a risk statement, affected workflow, owner, likelihood and consequence judgment, current controls, missing evidence, response, due date and residual-risk decision. The owner updates it after an incident, evaluation failure, vendor change, new data use or control change. A low score is not permission; the named risk acceptor still makes the call.

The vendor register records the product and plan, purpose, contract owner, data terms, access, subprocessors or model dependencies where documented, renewal, exit path, open questions and last verification date. Procurement or the named commercial owner updates it on a contract, plan, material documentation or dependency change. Marketing copy is not silently converted into a control claim.

The adoption plan belongs to the functional leader whose team is changing its work. It names affected roles, the workflow change, approved and prohibited use, training, manager reinforcement, support path, feedback, adoption evidence and rollback communication. It changes when observation shows the work is not usable, safe or understood. A higher license count alone does not justify a change.

The runbook belongs to the operating owner and covers access, normal operation, inputs, human review, known failure modes, escalation, incident handling, recovery, vendor contact, monitoring and shutdown. Delivery drafts it while the workflow is built; the operator verifies it before acceptance; and it changes after any operational, model, prompt, data, integration or control change. A runbook assembled on the final day is an untested handover.

The day-30 gate asks whether the company has enough evidence to authorize one bounded validation or delivery path. Required inputs are the verified inventory, opportunity register, named process owner, current baseline, approved data path, initial risk classification, acceptance measures, external delivery scope, budget and explicit exclusions. The gate does not require a live or production workflow. The choices are authorize, validate one missing assumption first, repair a prerequisite, or stop. "Continue discovery" is allowed only when it names the unanswered question, owner, deadline and decision it will unlock.

The day-60 gate decides whether a real workflow can enter bounded production. It requires representative evaluation cases, documented results and failure classes, access and review controls, observed user workflow, current cost, operating-owner evidence, an incident and rollback path, an updated risk register, and a runbook draft exercised by the intended operator. The choices are proceed to bounded production, narrow the scope, redesign and retest, hold for a dependency, or stop. A polished demo, a vendor benchmark or a handful of selected examples cannot substitute for this evidence.

The day-90 gate asks two separate questions: should this workflow remain in operation, and should the next-quarter portfolio receive more investment? The workflow decision needs acceptance results on real approved cases, named operating and functional owners, completed governance decisions, a verified runbook, monitoring and review cadence, known residual risks with explicit acceptance, current cost, adoption evidence and a tested shutdown path. The portfolio decision needs the maintained roadmap, scorecard, decision log, vendor register and capacity view.

The available day-90 outcomes are accept and operate, accept with a dated condition, extend a bounded test for one named evidence gap, return to redesign, or stop operation and preserve the learning record. The next-quarter portfolio outcomes are fund, defer, reject or investigate. Each outcome has an owner and review trigger. No engagement renewal, testimonial or claim of success is itself a gate outcome.

These gates are deliberately evidence-shaped rather than threshold-shaped. A company must set its own acceptable quality, cost, risk, time and adoption thresholds before testing. This article cannot supply those numbers, and it does not report a hidden customer result. It supplies the questions, records and consequences needed to make the decision without invented proof.

All of them promise the same shape, and most of them promise it without a single citation. We read the pages currently ranking for this question. They run from about 1,800 to 4,500 words. They split the quarter into days 1 to 30, 31 to 60, and 61 to 90. They name between four and ten deliverables per phase. The longest of them carried no external source at all. Just its own cost bands and a proprietary methodology name it never actually explains.

Here is the pattern nobody comments on. Go through the deliverable lists and count the items that are a working system. Scorecard, assessment, inventory, roadmap, blueprint, framework, training kit, measurement framework, operating model, execution plan. Every one of those is a file. The word "pilot" shows up, usually as "the first initiative moves from plan to pilot", which is a verb doing a lot of hiding.

That is not a small stylistic complaint. It is the whole difference between a quarter that compounds and a quarter that ends with a leadership team politely thanking someone. A fractional AI officer who bills for three months and hands over ten documents has sold the same thing a consultancy sells, at a discount, with a better title on it. We wrote about that split at length in the question of whether the hire comes with delivery capacity, because it is the one thing buyers get wrong before they sign anything.

So the framework below is organized around a single distinction. In each phase, what is running, and what is only written down.

Two weeks, one deliverable, and it is a written one: a list of every AI tool and every half-built automation already running inside the company, with an owner and a cost against each line. It skips the strategy, the roadmap, the forty-slide future-of-work deck. Just a list. The plan document is a different object and it lands earlier: what should be on the table by Friday of week one is the scope, the gates and the acceptance test, not the finished inventory.

It takes two weeks because the list is never the list leadership has in their head. UpGuard's 2025 state of shadow AI research surveyed 500 security leaders and 1,000 employees and found 81% of employees and 88% of security leaders reporting use of unapproved AI tools, with 40% of employees using them daily despite having sat through the training. So the gap is rarely one or two tools. It is the zap that has been quietly moving customer records since spring, the browser extension nobody in IT has heard of, and the account somebody set up on a personal card because procurement was going to take a month.

The second half of the two weeks is ranking, and the rank order is where the argument happens. The AI ownership and opportunity assessment capability explains the fit, prioritization, and evidence questions behind that work: function by function, where do the hours actually go, and which work may be safe to hand to a machine. The output of this operating phase is one page with a numbered list on it and one item circled. That circled item is what weeks three to six build.

What does not ship in weeks one and two: a governance framework. Writing rules for AI work before you have any AI work in production produces a policy that regulates a hypothetical, and it is a comfortable place to hide for a full month.

Representative approved cases are used in controlled validation, explicitly not production, to assemble evidence for one bounded workflow. That is the deliverable for weeks three to six: documented quality, cost, review, failure and access evidence that a formal gate can use.

The separate Production sprint offer is $5,000 fixed for one full workflow live in about two weeks. That published project scope is not proof that every workflow inside a broader engagement fits the same timing or that the operating plan reaches production before its evidence gate. The four-week window here reserves time for access, representative cases, exception handling, review, and adoption evidence.

The delivery work is concrete rather than hypothetical, and it is the kind of thing our embedded engineers support outside a notebook. Read the inbox, draft against the rules, escalate what does not fit. Exercise the CRM-to-billing path on approved cases so reviewers can see where judgment, access and exception handling break down. This is validation work, not a production claim.

Then comes the week-six controlled-use readiness check. Has the workflow been exercised on representative approved cases in controlled validation, explicitly not production, with documented quality, cost, review, failure and access evidence rather than as a selected demo? If not, narrow, redesign, wait on a named dependency, or stop. This evidence feeds the formal day-60 gate; it does not replace it.

The back half turns an approved workflow into controlled operation and prepares the company to own it. Additional builds remain separately scoped; this operating system does not promise a fixed workflow count.

The rules are written now rather than in week two because they can be specific: what the approved workflow may touch, what stays with a person, where the audit trail lives, and who gets paged when a run fails. A policy written against observed operation can stay concise. A policy written against a hypothesis becomes a chapter nobody uses.

A named internal owner should be required before operation is accepted. That owner sits in the reviews and holds the runbook and evaluation criteria. Grant Thornton's 2026 AI Impact Survey found 46% of operations leaders at $1.1B to $5B companies naming workforce skill gaps as the primary obstacle to AI, against 29% in the midmarket. Handing back a working system to a company that cannot operate it is a slower kind of failure, and it does not show up until the quarter after the invoice.

Day ninety is a read, not a celebration. Record what shipped, what got cut and why, what each accepted workflow is worth in hours, errors, or cycle time, and what the next quarter should attempt. Naming what got cut is the part that makes the rest of the report believable.

Model choice is not on the list. It is the thing buyers ask about first and it is close to the least consequential decision in the quarter.

Data access. Somebody has to grant a service account read access to the system the workflow depends on, and that request lands in a queue behind everything else IT is doing. Cloudera's numbers put the scale of it plainly: nearly 80% of 1,270 IT leaders say limited data access across environments is holding AI back, while 18% say their data is fully governed. Gartner's 2025 research found 63% of 248 data management leaders either lacking or unsure of the right data management practices for AI, and predicted that through 2026 organizations would abandon 60% of AI projects unsupported by AI-ready data.

The workflow redesign. Wiring a model into the process exactly as it exists today gets you a faster version of a process that was already bad. Redoing the process is where the return lives, and it is also the part that requires the people who do the work to change how they do it. BCG's guidance, published across several of its 2026 pieces including Scaling AI Requires New Processes, Not Just New Tools, puts 10% of the effort on algorithms, 20% on technology and data, and 70% on people and processes. Deloitte's 2026 survey of 3,235 leaders across 24 countries found 37% still using AI at surface level with minimal process change.

The named owner. Somebody inside the company has to hold this after the engagement ends, and that person has a day job. If the owner is named in week eleven, the handover is a calendar invite. If they are named in week two and sit in every review, the handover is a formality.

Sponsor attention. The executive who wanted this is the same executive with a reorg and a board meeting. Attention decays, and it decays fastest in the fortnight after the first thing goes live, which is exactly when the next operating decision needs attention.

The formal decisions happen at days 30, 60 and 90. Two earlier readiness checks keep the company from wasting the time before them. A check matters only if a miss creates an owner, evidence request and consequence; otherwise it is a status update with a dramatic name.

End of week two: delivery readiness. Is there one workflow scoped to a named process owner, with a current baseline, approved data path, initial risk classification and external delivery owner? If not, the company names the prerequisite, owner and date rather than pretending the build has begun. The day-30 gate can then authorize, validate, repair or stop with evidence.

End of week six: controlled-use readiness. Has the workflow been exercised on representative approved cases in controlled validation, explicitly not production, with documented quality, cost, review, failure and access evidence? If not, the work narrows, redesigns, waits on a named dependency or stops. This check feeds the day-60 bounded-production decision; it does not replace it.

That is the difference between a Fractional Chief AI Officer and a retained advisor. An advisor can report the blocker. An owner records the decision, consequence and next evidence, then carries that state into the formal gate.

Because a case study we cannot show you the evidence for is worth nothing, and we would rather publish nothing than publish that.

This matters more than it sounds like it should. The archetype this post belongs to is the first-party delivery study, and the honest version of it for a firm our size is the framework and the gates, not a set of numbers with a logo next to them. Every provider in this category can produce a page of results. Almost none of them can produce the raw material behind those results, and you have no way to tell the two apart from the outside.

So take this instead. The cadence above is what we run and what we will be held to. The gates are stated in advance, with the failure condition written next to each one, which means you can hold up your own engagement against them at days 30, 60 and 90 and get a straight answer. That is a harder thing to publish than a testimonial, and it is checkable in a way a testimonial never is.

Where the numbers in this post came from is stated on every chart. All of them are third-party research. None of it is ours to claim.

Three cases, and we would rather say them here than three weeks into an engagement.

If you have one clearly defined workflow and no ambition beyond it, ninety days of leadership is overkill. Buy the build. A starter build runs $1,500 to $2,500 fixed and a two-week sprint puts one full workflow into production for $5,000, and neither of those needs anybody sitting in your leadership meetings.

If nobody will give the work a budget and a veto, ninety days will produce a very good plan that loses every subsequent planning meeting to the product roadmap, because the roadmap has an owner and a bonus attached and your AI work does not. That is not a delivery problem and no cadence fixes it.

And if your data lives in a system that nobody has been able to get an export out of in two years, the first gate will fail and you already know it will. Fix that first. It is unglamorous work and it is the actual prerequisite, whatever anybody's phase-one deliverable list says.

The operating system below is for a non-software company where no employee writes software, firmware, or embedded code. The company supplies executive sponsorship, functional ownership, access decisions, and reviewers; approved implementation is commissioned externally. It is not a product-engineering playbook.

The cadence and artifacts are a reviewable specification. They are not customer evidence, and this article contains no invented client proof. A real engagement must replace every sample field, threshold, and gate with the company's own evidence.