An AI proposal should describe the work it will do and the boundaries around that work. “Agentic” can cover anything from drafting a reply to changing customer records. Those purchases have different risks, costs and review needs. The useful question is what the proposed system is allowed to do in your business.
A forecast is context, not a verdict on your project
In June 2025, Gartner forecast that more than 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. That is a dated forecast, not a measured failure rate for a specific kind of small-business deployment.
Gartner also used “agent washing” to describe products being relabeled as agentic without substantial agent capabilities. The practical response is to inspect the proposed workflow. A familiar automation tool may be exactly what the job needs; an agent label does not make it better.
Describe the capabilities before choosing the label
A chatbot is a conversational interface. It may follow a script, use a language model or connect to tools. An agent can use a model to choose among permitted actions, while deterministic automation follows configured rules. A single implementation can combine these approaches.
Ask the supplier to show the trigger, inputs, permitted actions, destination records, approval steps and failure route. A demonstration should include an exception, such as missing data or an unavailable integration, so you can see how the system behaves beyond the successful example.
Four levels of scope to compare
The following framework is a buying aid, not a ranking of proven success rates. More autonomy increases the range of behavior you need to specify and test. A narrow workflow can still fail, and a useful system does not need to climb every level.
| Scope | What to specify and test |
|---|---|
| Answer and capture | Approved sources, request fields, operating hours, capacity and a route for questions it cannot answer. |
| Book and update | Read and write permissions, valid booking rules, duplicate prevention, rollback and staff handoff. |
| Coordinate multiple steps | The shared record, order of actions, retry behavior and approvals before consequential changes. |
| Pursue an open-ended goal | The limits on possible actions, spending and access, monitoring, stop authority and recovery from unexpected behavior. |
For Voice AI or Chat AI, define coverage and concurrent capacity rather than assuming every call or message will be handled. Test peak demand and the fallback when a provider or integration is unavailable. Customer questions outside the approved sources need a configured human route.
A scheduling or record-update agent needs more than a correct answer. It must act on the right record, respect the permitted fields and handle repeated requests without unintended duplicate actions. A wrong booking can have real consequences; do not assume every action is cheap or reversible.
For a workflow spanning several tools, agree where a person approves consequential actions and what they can inspect before doing so. Approval is a control to test, not proof that the remaining error rate is negligible. Keep review time and exception handling in the operating cost.
Ask for evidence that matches the proposed job
A relevant case study can show that a supplier has implemented a similar workflow. It cannot guarantee your outcome. Our dental and HVAC comparison reports 93 and 23 month-one bookings; its later call-log observations use their own windows. Those counts do not establish universal capacity or financial return.
For your pilot, define a correct outcome, a representative test set and acceptance criteria. Keep the task counts, dates, exceptions and human interventions visible. Compare the baseline and pilot on the same scope, and investigate changes in workload before attributing the difference to AI.
Keep ownership, oversight and exit terms concrete
Get written answers about the accounts, source code, workflows, prompts, records and phone numbers involved. Identify third-party licenses and dependencies, who can export the data and what work is required to move providers. An account in your name is useful, but it does not make every dependency disappear.
Define who can pause the system, who receives error alerts and who reviews changes to its sources or permissions. For consequential financial or customer actions, specify the required approval rather than assuming a model will recognize every situation needing human judgment.
The AI project checklist and workbook includes scope, test and handover questions. Use it to record evidence and unresolved decisions, not as a certificate that a system is safe or ready.
Decide what to buy from your own baseline
Start with the repetitive job your records show needs attention. Compare removing a step, improving the process, using an existing integration and building an AI workflow. Consider the cost of implementation and ongoing review alongside the work it could remove.
Agree a limited pilot and review date. Continue when the measured quality, cost and outcome meet the stated criteria; revise or stop when they do not. Market adoption forecasts do not establish a deadline to buy, and a narrow scope does not guarantee that a deployment will remain useful in 2027.
For a more detailed implementation plan, read the 90-day implementation playbook. If you need help choosing the first project, the AI Consulting service describes the diagnostic work and how its scope is agreed.