An AI automation agency should deliver a working business process, its tests, and the instructions your team needs to run it. A model connection or an impressive demonstration is only one part of that purchase. The proposal should explain where work starts, what the software changes, and who handles failures.
This guide is for buyers comparing implementation proposals. It separates a useful first release from optional expansion, so you can ask agencies to quote the same outcome. Catapult's service scope covers process mapping, approved integrations, evaluation, and the handoff agreed for your project.
Related service: AI automation services- 01Map the current task and its exceptions.
- 02Confirm data access and the first release boundary.
- 03Build the workflow with review and recovery paths.
- 04Test ordinary, uncertain, and failed cases.
- 05Train the owner and agree launch support.
SECTION 01
Ask for evidence at each delivery stage
Discovery should produce a short process map using your team's vocabulary. Record the trigger, required information, business rules, approvals, destinations, and definition of done. Ask the agency to identify which steps use ordinary rules and which need AI interpretation. This makes the work understandable to the person who approves it.
Next, request a small working connection to the most uncertain system. Confirming that a third-party API permits the required action is more useful than polishing screens around an unavailable connection. The agency should record access restrictions and propose a practical alternative where necessary.
| Stage | What you should receive | Your decision |
|---|---|---|
| Discovery | Process map, sample inputs, exclusions, and named owner | Is this the right task to automate? |
| Implementation | Reviewable workflow, connections, and exception handling | Does it perform the agreed task? |
| Acceptance | Test cases, observed failures, and unresolved limitations | Is the release ready for its intended users? |
| Handoff | Account register, runbook, training, and support boundary | Can our team operate and stop it? |
SECTION 02
Decide what to configure and what to build
If an existing tool already supports the trigger, review screen, and destination, configuring it may be the right first project. Custom development becomes useful when a workflow needs unusual permissions, several systems, special validation, or a user interface the available tools cannot provide. A mixed approach is also possible.
Compare alternatives using recurring subscriptions, maintenance responsibility, access to your data, and the effort of changing vendors. Ask what you retain if the engagement ends. For U.S. buyers working with a remote team, document review overlap and the escalation channel; a globally distributed team does not automatically provide round-the-clock support.
SECTION 03
Worked example: triage incoming service requests
Consider an illustrative maintenance company receiving requests through a shared inbox. Its first release reads new messages, extracts the site and issue, suggests a category, and creates a draft job. A dispatcher checks missing information and approves the record. This is an example scope, not a Catapult client result.
The quote includes one inbox, one job system, approved categories, a review queue, and a record of the source message. It excludes automatic scheduling and commitments about arrival times. A message describing an urgent hazard goes to the existing human escalation process rather than receiving an AI-generated promise.
Acceptance tests include a repeated message, two sites in one email, an unreadable attachment, an unavailable job API, and a message that does not describe a service request. The buyer checks that no failed item disappears and that repeating a request does not create duplicate jobs.
SECTION 04
Make authority and review part of the scope
Give the workflow only the permissions its job needs. OWASP identifies excessive functionality, permissions, and autonomy as sources of agent risk. Translate that into a clear purchasing question: can this release read, draft, send, update, or delete—and which of those actions require approval?
NIST's AI Risk Management Framework treats evaluation and ongoing use as part of risk management. Your project can apply that principle through a repeatable test set and an assigned reviewer. Record errors by type, including unsupported answers and wrong record updates, rather than accepting a vague promise of accuracy.
SECTION 05
Accept the operating handoff, not just the demo
Have the workflow owner pause an automation, inspect a failed item, correct its input, and safely resume it during handoff. Ask where logs live, how long they are kept, who can access them, and which alerts reach a person. Identify the owner of model, hosting, and connector accounts.
An affordable release can cover one reliable task before expanding. Keep the quote's implementation work separate from provider usage and ongoing support. Contact us for pricing; the linked cost guide explains the variables to compare without treating every workflow as the same package.
SECTION 06
Frequently asked questions
What should an AI automation agency proposal include?
A defined workflow, data and integration requirements, permissions, acceptance tests, delivery stages, handoff, and support responsibilities. It should also identify exclusions and recurring third-party costs.
Do we need custom software for every automation?
No. Configuration may be enough when existing tools support the full task. Custom work is useful when requirements exceed those tools or need a different interface, data model, or control boundary.
Can we begin with an affordable pilot?
Yes. Choose one task, a limited source collection, and a reviewable output. Preserve essential failure handling and permissions while deferring optional workflows.
PRIMARY REFERENCES
Sources and further reading
These references cover the standards, platforms, or published prices discussed in the guide. Worked examples and checklists are our editorial guidance.
EDITORIAL METHOD
About this guide
We use AI to assist with drafting and editing. Catapult AI Work is responsible for the published content. Examples illustrate possible approaches; they are not client case studies unless identified as such.
Budget examples are not Catapult package prices. Check linked provider pages for current fees and plan limits before making a purchase.
Read the editorial policy