Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and model capability is not among the reasons. ADAPTA is the decision framework we use to match the right class of automation to a business process before the budget is committed.
Every automation decision now starts in the same place: a demo that works, a team asking for budget, and no reliable way to tell whether the technology on offer actually fits the process being automated. Getting that judgement wrong is the largest single predictor of failure, and it is a judgement most organisations make by instinct.
On 25 June 2025, Gartner predicted that more than 40% of agentic AI projects would be cancelled by the end of 2027, naming three causes: escalating costs, unclear business value, and inadequate risk controls. Model capability is not on that list.
More than a year on, that prediction reads as conservative. Gartner's 2026 Hype Cycle places agentic AI at the Peak of Inflated Expectations, with roughly 17% of organisations having deployed agents against more than 60% expecting to within two years. Industry surveys through 2026 put pilot-to-production conversion anywhere between 15% and 35% depending on methodology. The spread is wide enough that no single figure should be quoted as fact, but the direction is consistent across every source: adoption climbed, success rates did not follow.
The failure causes are consistent too. They cluster on governance, evaluation and observability, data readiness, and running costs that were never budgeted because they barely registered during the pilot. Very few projects fail because the model could not do the task.
Gartner also named the supply-side problem: agent washing, the rebranding of assistants, RPA and chatbots as agentic without substantial agentic capability. Of the thousands of vendors positioning agentic products, Gartner estimated only around 130 were real. Its blunter observation is the one worth pinning above the desk: many use cases positioned as agentic today do not require agentic implementations.
Under-powering the process. A bank modernising a legacy core of roughly 400 interconnected systems could not get there with manual coding or scripted automation. The work was genuinely adaptive: dependencies surfaced during execution, and decisions changed the path downstream. Agentic workflows, with engineers supervising squads of agents against shared objectives, were the right tool because the process type demanded it.
Over-powering the process. The more common error, and the more expensive one. A team adopts generative or agentic tooling for work that is deterministic, well specified and repetitive: extraction from fixed-format documents, code quality checks, reconciliation against a known rule set. The pilot demos beautifully. Then the running costs arrive, the outputs prove non-deterministic in a process that required determinism, and the controls needed to make it safe cost more than the automation ever saved. Static analysis, a testing framework and RPA would have produced the same outcome at a fraction of the price, with a cleaner audit trail.
Both failures have the same root: the class of automation was chosen before anyone characterised the process.
Evolve AI Labs has built automation across the spectrum, from loan underwriting to invoice reconciliation to compliance document generation, using programmatic, predictive, generative and agentic approaches. ADAPTA (Automation Decision And Process Type Assessment) is the framework we use to decide which of those a given process actually needs. It sits at the intersection of business process management and the current automation stack, so it stays legible to leadership without being naive about the technology.
Type | Description | Input | Output | In plain terms |
|---|---|---|---|---|
Programmatic | RPA, scripting, conventional code | Expected and well defined | Fixed, rule based | Follow these exact steps, every time |
Pattern based | Predictive machine learning | Expected, structured, varied | Fixed, pattern based | Learn from data, predict outcomes |
Zero shot | Language model, single step | Novel and unstructured | Varied, knowledge based | Understand language, generate a response |
Agentic | Language models orchestrating tools over multiple steps | Complex, novel, unstructured | Knowledge and decision based | Orchestrate complex, adaptive processes |
The distinction between zero shot and agentic has hardened since 2025. A model answering a question in one turn and a system given a goal, a set of tools and the autonomy to take steps toward it are different products with different failure modes, different cost profiles and different control requirements. Vendors routinely conflate them. Ask which one you are buying.
BPMN is an open standard maintained by the Object Management Group. It gives business analysts, developers and process owners a shared language, which is precisely what is missing from most automation conversations.
Type | Description | Example |
|---|---|---|
Atomic process | A single well defined task that cannot be broken down further | Looking up matching records across database tables |
Sequential workflow | Atomic processes in order, each completing before the next begins | New hire onboarding |
Process workflow | Predictable, repetitive steps forming a cyclical graph | Inventory management and tracking |
Case workflow | A non-cyclical graph where the path unfolds as information arrives, with terminal nodes | Legal documentation and compliance |
Adhoc workflow | Complex matrix or network processes handling non-standard work | M&A due diligence |
Crossing the two gives the mapping. Fit marks the classes of automation suited to each process type. Most process types have more than one, which is deliberate: the choice within a row is then made on cost and risk, not capability.
Process type | Programmatic | Pattern based | Zero shot | Agentic |
|---|---|---|---|---|
Atomic process | Fit | — | — | — |
Sequential workflow | Fit | Fit | — | — |
Process workflow | — | Fit | Fit | — |
Case workflow | — | — | Fit | Fit |
Adhoc workflow | — | — | — | Fit |
Read the diagonal. As processes become less predictable, the automation that suits them becomes less deterministic. The corollary is the useful part: if your process sits in the top two rows and someone is proposing agentic automation, the burden of proof is on them. An Adhoc workflow can have atomic processes which are implemented as programmatic tools that are provided to the Agent Pipeline, or can have any of the other process types.
Step 1, business process visualisation: Map the current state of the process as a workflow before deciding anything about technology. This tells you which BPMN type you are dealing with, and that understanding drives both implementation design and risk mitigation. Teams that skip this step are choosing technology for a process they cannot describe.
Step 2, complexity testing: Can the process be simplified before it is automated? A significant section of processes cannot be described as optimised for the latest evolution of the organization. They carry redundancies, silos and bottlenecks that are artefacts of change over time. Lean six sigma is enough for this. Automating an unexamined process encodes its inefficiencies permanently and at higher cost.
Step 3, process and technology mapping: Use the ADAPTA matrix. Where a row offers more than one fit, take the leftmost option that meets the requirement. Determinism is a feature, not a limitation.
Step 4, cost and benefit check: Estimate build versus buy, return on investment, and value at risk. Two costs are routinely missed at this stage. Inference and orchestration costs scale with volume in a way that pilot budgets never reveal, and the evaluation, observability and human review needed to run a non-deterministic system safely in production are a permanent operating line, not a one-off build cost. Aim for the simplest automation that works, and only add sophistication where the value at risk justifies it.
Step 5, recording the decision: Consolidate the output into a single document covering process analysis, technology recommendation, implementation timelines, risk assessment, decision rationale, and approval and next steps. Record what you rejected and why, because that record is what protects the project when someone asks six months later why it is not agentic. Start small and graduate to larger processes.
Three things have changed since this framework was first written. Agent washing has made vendor claims materially harder to evaluate, so the buyer now needs an independent view of what a process requires before entering a procurement conversation. Evaluation and observability have emerged as the dominant production blocker, which shifts cost from build to run and changes the business case. And the regulatory position has settled in a way that rewards this kind of documented reasoning: Australia's December 2025 National AI Plan opted for technology-neutral regulation under existing law rather than a standalone AI Act, which means existing regulators assess your decisions using existing standards. A defensible record of why a technology was chosen is now part of the compliance artefact, not just good practice.
Pick one process you are currently considering for automation. Run steps 1 to 3 on it this week. If the matrix points somewhere less sophisticated than the proposal on your desk, you have just found the most valuable output this framework produces.
If you would like our decision record template, or want to run this across a portfolio rather than a single process, get in touch. We run an intensive workshop built around exactly this exercise.
Learn how Australian businesses are maximising value from their AI investments with Evolve bespoke solutions
ROI
faster invoice processing
We build a custom solution to maximise your business revenue, reduce costs and add operational efficiency
Speak to an expert©Evolve AI Labs
Terms and Policies
