Assess before you deploy.
A working model for the first month of an AI deployment: the assessment phase that decides whether anything built afterward is worth building at all.
Companies label “transformation” as an initiative to unlock some untapped value.
Whether it’s transforming the organizational structure or transforming how work is done with new technology, someone, typically a sponsor at the C-suite or VP level, wants to change something so that there is some new result. The hope is that the change is a net positive, landing in one of three places:
These initiatives are not always fully scoped projects. Instead, companies will purchase a new technology, usually a SaaS product, thinking that the technology itself will transform their work and solve their problems.
The reality is that software is designed to function as it’s programmed, which does not always align with what operators need.
Enter AI, which many leaders expected to be the silver bullet to their technological transformation initiatives: we can use AI to automate everything, so that people can do the tasks that increase revenue, lower costs, or mitigate risk.
Many companies gave employees free rein to use AI, assuming that they would put it to good use. Some companies had unlimited, and unstructured use. Employees ended up running many tasks through a model regardless of fit, which resulted in blowing through the company's annual AI budget within the first half of the year.
Without a thoughtfully structured transformation initiative, companies will continue to burn through their AI budget without seeing the massive rewards AI could unlock.
AI should not be applied to every workflow by default.
That’s how you tokenmaxx and blow your budget. Work should be redesigned and handled in three ways:
When the rules and inputs are predictable, the work is automated by software.
When the objective is clear but the inputs, path, or required actions vary, the work is handled by AI where appropriate.
When the decision carries material ambiguity, accountability, or irreversible consequences, it stays with a person.
Companies that want to deploy AI effectively should follow this structure:
Assess and document every workflow and process end to end, to understand how work gets done.
Identify where the pain points are, like highly repetitive and manual processes.
Capture company context and the secrets that live in your employees’ heads, and convert those into rules, instructions, and decision logic.
Start with the workflows where AI can deliver the biggest impact: high-volume, time-consuming, spread across multiple systems, driven by repetitive decisions, dependent on institutional knowledge, and tied to a meaningful business outcome.
Redesign the workflow into three parts: what deterministic software should own outright, where an agent should apply judgment within clear boundaries, and what has to stay with a person. Most of what looks like “AI work” turns out to belong in the first category.
Validate before scaling autonomy. Run it in a sandbox, then in shadow mode alongside the people doing the job today, and only then in supervised production — logging every correction so accuracy compounds instead of resetting with each new workflow.
A growing set of AI-native advisory and implementation firms now exist specifically to help companies capture this kind of value. This is how I would structure that engagement.
Every engagement starts the same way: someone believes AI can create value somewhere in the business, and nobody yet knows exactly where, how much, or what it would take to implement. The first thirty days exist to answer that with evidence.
No agent gets recommended, no software gets built, and no roadmap ships until the operating picture underneath it holds up.
The range of outcomes at day thirty includes a recommendation not to proceed. That option must be on the table, or the diagnostic is just a sales pitch with extra steps.
Assessments require engaging with process owners, approvers, front-line workers, and shared services teams to fully understand how they handle their day-to-day work, and to determine whether the organization is positioned to sustain a change.
Interviews
Collection
Tracing
Study
Construction
Sequencing
Deliverables
The goal for this phase is to get a full picture of how work is done today.
Collect all the documented artifacts: SOPs, spreadsheets, exception logs, system exports. Then —
Trace the entire workflow, and make no assumptions.
Capture how the actual workflow aligns, or doesn’t, with the expected one, and how exceptions get handled. Every step gets sorted into one of three piles:
High-volume, repeatable, manual tasks.
Tasks that require reasoning and judgment.
Tasks that, if handled wrong, pose a risk to the business.
Frontline workers keep the secrets of how real work gets done in their heads. Often the full workflow was never written into a single SOP, and the secret to completing it lives with the employee, not the document.
Now that you have the full picture of how work is done, analyze where the gaps and opportunities live.
Time allocation study: quantify how the organization spends its time. Categorize every hour into:
Process map construction: build side-by-side process maps for every sub-function. Document the SOP against the actual workflow, and identify every gap, manual handoff, and undocumented step.
The SOP, the org chart, the written policy.
What actually happens, exceptions and all.
Gap analysis: compare documented policy against observed behavior.
This phase is dedicated to building the transformation roadmap.
Automation opportunity identification: map every manual workflow to an automation opportunity, and score each one.
ROI Potential
Implementation Complexity
Dependency
Risk Exposure
A high-ROI workflow with real compliance or financial exposure gets a separate review track, instead of sailing through on the strength of its number alone.
ROI modeling and business case: a bottom-up model that rolls up into the same three categories from the opening.
Every assumption gets validated against data observed during the assessment, not outside benchmarks.
Develop the roadmap: sequence every opportunity into an implementation timeline, mapped by dependency.
Internal QA and deliverable assembly: cross-check every finding against source data. Peer-review every process map, exception categorization, and ROI calculation before it leaves the room.
Not every engagement ends with a green light.
The ones that do are the ones where the sponsor walks away with three things they can act on without further explanation:
A current-state map of how work happens today, a ranked list of workflows worth redesigning with evidence to support it, and clear boundaries for what the AI can and can’t do.
A phased roadmap that starts with the workflow already validated by both the data and the interviews.
The expected value unlock from implementing AI — hours, cost, and errors, quantified — and a plain answer to whether the organization is ready to sustain what gets built.
Because the ones that get a green light are the ones where it all shows up on the same page: