Gartner says more than 40% of agentic AI projects will be killed by the end of 2027. The reasons are predictable if you have been in the room when these projects get funded: cost overruns, unclear business value, inadequate risk controls. None of these are technology problems. Every one of them is a decision-making problem — and the decision-making breaks down long before anyone picks a model or signs a vendor contract.
After twelve years of building enterprise AI in surgery, nuclear operations, defense, and commercial enterprise, we see the same root cause behind almost every failed project. The team never described the work. Not the department. Not the job title. The actual work — the workflows that move inputs to outputs, the handoffs that stall, the exceptions that eat most of the labor hours. Until that description exists, every AI investment is a guess dressed up in a business case.
AI investment is not an AI question
This is the first thing that goes wrong. A finance leader gets told by her CFO to “do AI” in orders-to-cash. Three vendors quote three different solutions. None of them have described the actual work her team does. And this is not unusual — it is the norm.
The conversation starts in the wrong place. It starts with the model, the platform, the vendor demo. But the model question is downstream. The vendor question is downstream. The dashboard question is downstream. What sits at the root is how the work itself is shaped and where it creates value.
Most teams skip this step, often because their vendors encourage them to. The vendor has a product to sell. The product has features. The demo shows the features working on a clean, controlled scenario. The contract gets signed. And then the real work begins — the work of figuring out what the tool should actually do inside the buyer’s operation.
This is backwards. You do not buy a surgical instrument and then decide what surgery to perform.
The unit of decision is the workflow
An accounts receivable team does not have one AI problem. It has six or eight: collections prioritization, invoice matching, customer follow-up, exception handling, cash application, dispute resolution, reporting, escalation. Each is a different shape of work. Each routes to a different investment. If you pile all of them into a single RFP, you get a mediocre tool that does maybe one of them well and covers the rest poorly.
The same pattern holds everywhere. A product team’s work includes user research synthesis, spec drafting, backlog grooming, design review, experiment analysis, roadmap judgment, launch coordination, and customer escalation. These are all different shapes of work. Some might be builds. Some might be buys. Treating them as a single “AI for product” initiative guarantees a poor fit.
The unit of decision for AI is not the department head. It is not a particular role. It is the workflow — the entire operating loop. What information comes in, what the system is allowed to do, what good output looks like, who checks it, what gets escalated, who owns the result. The AI model is a small part of that loop. An important part, but a small one. If you want to invest correctly, you have to think less in terms of models and more in terms of workflows, because the workflow is what you are actually investing in.
Every workflow can be scored
Once you decompose the work to the workflow level, the investment question becomes tractable. Every workflow can be evaluated on a handful of dimensions:
- Frequency. How often does this workflow repeat? High-volume, repetitive workflows are the most obvious automation targets.
- Error cost. How costly is a mistake in this workflow? High-stakes workflows need more human oversight, not less.
- Judgment required. How much human judgment does this workflow demand? Workflows that require extensive contextual reasoning are harder to automate reliably.
- Specificity. How specific is this workflow to your company? Generic workflows often have mature market solutions. Company-specific workflows require custom builds.
- Market maturity. Does the market already have a solution for this? If mature tools exist, buying makes more sense than building.
This scoring produces a prioritized map, not a wish list. It tells you which workflows justify investment now, which should wait, and which are better served by a hire than by a tool.
Five levers, not one
When teams talk about “doing AI,” they usually mean automation. But automation is only one of five investment levers available for any workflow:
Automate
Delete the workflow entirely by replacing it with a system. This is the right call when the work repeats often, follows a clear pattern, has recognizable exceptions you can define, and you can verify output cheaply. Do not automate when the exceptions are where most of the value lives — this is where bad enterprise AI demos fall apart. The vendor shows the routine case. The buyer signs the contract. Production traffic turns out to be mostly exceptions. The accuracy number craters. Nobody was lied to. The buyer just bought the wrong thing.
Build
Create a custom AI-powered workflow using agents, tools, connectors, and data. This makes sense when the work is company-specific, has edge cases and context that matter, and no off-the-shelf product covers it. The hard question is whether you can define what “good” looks like for the output. If you cannot describe inputs, outputs, standards, and exceptions, you cannot evaluate whether what your team builds actually works. Your team will be incentivized to tell you the AI thing is amazing. You need to know enough about the workflow to be the honest third-party eyes.
Buy
Purchase a solution off the shelf. This works when the work is common across your industry and the market has mature offerings. For company-specific work where the market has some useful primitives, buy the primitives — the building blocks, the connectors, the model access — but own the workflow. The distinction matters: buying a complete workflow solution from a vendor requires you to confirm that there is 80–90% overlap between how they envision the workflow and how yours actually operates. If there is not, integration will cost more than anyone budgeted.
Hire
Fill the talent gap with people. Many companies right now are searching for a purple unicorn — a domain expert who is also an AI builder, a systems architect with executive experience, and a change leader. That person rarely exists. The better question is what human capability your workflows actually need in six to twelve months, then hire for that specific missing piece. If nobody on your team can define the standard, describe what good looks like, or frame the problem — that is a signal that your next investment is a person, not a platform.
Wait
Prioritize elsewhere first. This is the most counterintuitive lever when the world is telling you to move fast on AI. But you have limited resources for change management. Apply them where you get the most leverage. Stack your investments so the high-priority workflows get attention first and deliver disproportionate returns. Waiting is not the same as doing nothing. It is strategic sequencing.
Most real-world situations require a combination of these levers. You might need to hire before you can build. You might buy primitives and build the workflow around them. You might automate the routine portion of a workflow and keep a human in the loop for exceptions. The point is that “do AI” is not a lever. It is a category that contains five very different investment decisions.
Do not automate what you cannot describe. If you cannot articulate the inputs, the outputs, the standards, the exceptions, and the ownership, you cannot make a good investment decision.
The investment matrix
When you plot workflows across two axes — how specific the work is to your company and how mature the market solution is — the investment decision clarifies:
- Common work + mature market: Buy. This is Workday for commodity HR, Stripe for payments, a standard help desk for standard help desk needs.
- Common work + immature market: Prototype narrowly or wait. The category is still defining itself. Do not sign a five-year contract for a tool category that will look different in twelve months.
- Company-specific work + useful primitives available: Buy the building blocks, own the workflow. This is where most ambitious teams should be operating. You buy the connectors, the model access, the orchestration tooling. You own the standard and the logic.
- Company-specific work + thin market: Build and own the category. If the work is unique to you and nobody is solving it, you have both the need and the opportunity to build something proprietary.
Hiring cuts across this entire grid. If the work requires trust, framing, or standard-setting that nobody on your team can do yet, your next investment is a person who can define the workflow well enough to make all the other decisions possible.
The cheap debate and the serious one
There is a version of this conversation that turns into AI versus people. That version is not useful. It generates drama and headlines but no operational clarity.
The serious version asks different questions. Where should people be maximizing their time? Where should they be leveling up? Where are the talent gaps you can hire for? Where are bundles of tasks within a job family that are getting picked up by automation, so the remaining human work can be higher-impact and more leveraged?
This is not AI replacing workers. This is figuring out how to allocate investment so that both human effort and AI capability produce disproportionate value. It starts and ends with understanding the workflow.
Map first, invest second
The companies that avoid the Gartner kill list share a common trait: they describe the work before they decide how to invest in it. They decompose departments into workflows. They score each workflow on frequency, error cost, judgment, specificity, and market maturity. They match each workflow to the right lever — automate, build, buy, hire, or wait. And they make investment decisions at the workflow level, not the department level and not the “AI strategy” level.
This is what our discovery process is designed to do. Fifteen to twenty structured interviews across an organization surface the actual workflows — not the org chart version, not the employee handbook version, but how the work really moves. Each workflow gets scored. Each gets matched to an investment lever. The output is a prioritized catalog that tells you exactly where your next dollar of AI investment will produce measurable returns.
It takes two to three weeks. It costs a fraction of a single failed AI deployment. And it replaces the guesswork that characterizes most AI investment decisions with something you can actually defend in a board meeting.
The executive who can describe their workflows specifically, match them to the right investment lever, and sequence them by impact is the one who gets real returns from AI. Everyone else is buying instruments before diagnosing the patient.