Tier one is deterministic automation. Same input, same output, every time. Scheduled extracts, template-driven loads, validation rules, reconciliations that either tie or do not. No model in the path and no explanation required beyond the code. Most of the volume in a finance or operations function belongs here and always did.
Tier two is an AI-enabled workflow. A model does the judgment-shaped work, reading something unstructured, matching it, classifying it or drafting from it, at defined points inside a sequence somebody designed. The path is fixed before the job runs. You can draw the flowchart.
Tier three is agentic. The system determines its own sequence. It decides which tools to call and in what order, judges whether a step worked, and changes approach when one fails. You cannot draw the flowchart in advance, because the flowchart is the output.
The line between tier two and tier three is the sequence. It is not the sophistication of the model, and it is not the presence of a person. A tier two workflow can run on the same model as a tier three agent. What separates them is who decides the order of operations: a designer, in advance, or the system, at run time.
The practical test is whether someone can draw the flowchart before the job runs and whether it covers the cases. Bounded branching that a designer enumerated is still tier two, however many branches there are. A step reading "the model decides what to do next" is tier three, even if it appears once in an otherwise fixed path.
The test follows from that. Escalate a tier only when the tier below probably cannot do the job, rather than when it would be inelegant or when a vendor demonstration was impressive.
Run that assessment honestly across a portfolio and most candidates resolve to tier two. A meaningful share land in tier one, because they were never an AI problem. Both outcomes are good news. Tier one and tier two systems cost less to run, clear governance faster, explain themselves to an auditor without difficulty, and reach production at a far higher rate.
Oversight does not set the tier and treating it as though it does is the most common and most expensive error in these conversations. Two examples make the point.
A nightly document extraction job that posts every record above a confidence threshold and queues the rest for review touches no human on most of its volume. It remains tier two, because the sequence never varies.
An agent that plans its own approach but must obtain sign-off before each action has a person on every single step. It remains tier three, because the agent chose the steps.
Whether a person sits in the loop, on the loop, or in command is a control selected independently of the tier. An approval gate limits exposure, which is valuable. It does not change what was bought, what it costs to run, or what has to be proven before it goes live. Deciding the two questions separately, and in that order, is what keeps a governance conversation from becoming a scoping argument.
The tiers differ in ways that show up on an operating budget and in a control environment, which is the part most proposals leave out. Cost at tier one and tier two is forecastable. At tier three it is a function of decisions the system has not made yet.
Tier one. Run cost is compute and maintenance, essentially fixed per execution. Testing means confirming the right output for known inputs. Failure is loud: the job errors, or the reconciliation does not tie, and somebody knows immediately.
Tier two. Run cost scales with volume but stay predictable, because the number of model calls per execution is known before the job runs. Testing means sampling the model’s judgment at each defined point against a labelled set and setting confidence thresholds. Failure is quiet: a misclassification posts and looks like an ordinary record, which is why sampling and threshold design carry the control weight rather than the model choice.
Tier three. Run cost is variable and difficult to forecast, because the agent decides how many steps to take and which tools to call. Testing cannot enumerate the paths, so assurance moves to outcome evaluation, tool permissions, run-time limits and an audit trail of the sequence the agent chose. Failure can compound across steps before anyone sees it.
Escalating costs and inadequate risk controls are the two reasons agentic projects are most often abandoned, and both are visible in that comparison before a contract is signed. A proposal that has not addressed run-cost variance or path assurance has not addressed the two things most likely to end the project.
The classification is easier to apply than to describe. Four candidates of the kind that appear in most finance and operations portfolios:
Three-way match between purchase order, receipt and invoice. The rules are explicit and the comparison is exact. Tier one. Any AI in this path is decoration, and it adds a run cost and an explainability burden to work that was already solved.
Coding supplier invoices to general ledger accounts. The input is unstructured, the judgment is real, and the sequence is identical every time: read, classify, apply threshold, post or queue. Tier two, unambiguously, however sophisticated the model doing the classification.
Abstracting key terms from a lease portfolio. Read the document, extract defined fields, flag low-confidence extractions for review. The field list is fixed in advance, so the flowchart is drawable. Tier two, and a common target for a tier three proposal because the documents are complex. Document complexity is not sequencing complexity.
Investigating a month-end variance of unknown origin. The second query depends on what the first one returned, and the investigator cannot know in advance which systems will need to be interrogated. Nobody can draw this flowchart. Tier three, genuinely, and worth funding as such.
The pattern across a portfolio is consistent. Complexity in the input pushes work from tier one to tier two. Only unpredictability in the sequence pushes it to tier three, and that is far rarer than proposal language suggests.
Tier three is not a marketing category. It solves a real class of problem, and the organizations that reserve their agentic budget for that class tend to get value from it.
The signature is a branch space too large or too unknown to enumerate, where the next step depends on what the last step is found. Incident triage across an unfamiliar estate qualifies. So does research synthesis where the useful sources are not known at the outset, and migration work across a heterogeneous estate where each component reveals its own constraints. In each case a human expert would also improvise, and improvisation is precisely what the tier buys.
The discipline is reserving that spend rather than defaulting to it. An organization that runs the test across twenty candidates and funds the two that genuinely need tier three has a better chance of succeeding at both than one that funds all twenty at tier three and cancels most of them in year two.
Tier inflation is rarely a technical error. It enters through procurement, and it enters early. The word agentic appears in a requirements document because it signals that the organization is current. Vendors respond in kind, because a proposal that does not use the word looks dated next to ones that do. A demonstration then shows the hardest case in the set, which is usually a genuine tier three problem, and the pricing that follows covers the whole scope at that tier.
Nobody in that sequence behaves unreasonably. The question that never gets asked is whether the hard case is representative or exceptional, and whether the remaining ninety percent of the volume needs the same architecture.
The fix is cheap and structural. Put the flowchart question in the requirements document rather than in the evaluation. Asking each respondent to state which parts of the scope require a dynamic sequence, and why a fixed one fails, converts a positioning contest into a comparable technical answer, and it does so before the pricing anchors.
In a June 2025 release, Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The same release estimated that only about 130 of the thousands of vendors marketing agentic products meet the definition, a gap Gartner calls agent washing. Other research published that same month reported that fewer than 10% of deployed use cases get past the pilot stage, with function-specific applications the ones most often stranded there. PwC’s April 2025 survey of 308 US executives found 57% of adopters reported cost reductions.
Those sources are worth reading with attention to what each is arguing. McKinsey and PwC both treat the modest returns as a reason to push further into agents. That is their thesis. From a buyer’s seat the same figures support a narrower and more useful conclusion: the projects that reach production are the ones whose scope was set correctly at the start.
Note also that the cancellation reasons Gartner lists are not model failures. Escalating costs, unclear business value and inadequate risk controls are all scoping and governance outcomes, decided before any code runs. None of these argue against the technology. It argues for matching the tier to the work, which is a decision available to any organization at no cost.
The test reduces to three questions, asked in that order. They hold up in a technical review and in front of an investment committee, which is unusual enough to be worth the discipline of asking them in sequence:
The first two questions settle the tier. The third settles the controls and keeping them apart is what stops a governance conversation from turning into a scoping argument halfway through.
The sequence also gives a sponsor something specific to say when the proposal comes back smaller than the original ask. The scope did not shrink. The architecture was matched to it, and the gap between the two numbers is the gap between two ways of solving the same problem. Right-sizing an AI investment is not a decision to do less. It is what makes the spending defensible, the run cost predictable and the go-live date credible.
A proposal that answers the first two questions clearly is describing a system. One that cannot be describing a budget. The organizations getting value from AI right now are largely the ones that ran this test early, bought the tier the work required, and reserved genuine agentic spending for the small number of problems that defeat a fixed sequence. It is an unglamorous discipline, and it is currently the strongest available predictor of whether an AI program reaches production at all.
When you choose Opportune, you gain access to seasoned professionals who not only listen to your needs, but who will work hand in hand with you to achieve established goals. With a sense of urgency and a can-do mindset, we focus on taking the steps necessary to create a higher impact and achieve maximum results for your organization.