Our starting question is what the person needs to accomplish, how often it happens and where the information lives. A writing task inside one person's working day may need an existing assistant. A repeatable process across two systems may need automation. A task with several possible routes may justify an agent. None is automatically the most mature choice.
Compare the job, not the label
| Approach | Good starting point | Check before committing |
|---|---|---|
| Existing AI assistant | Drafting, research or questions using supported company sources | Plan availability, permissions, supported actions and team training |
| Ordinary automation | A predictable trigger and a repeatable set of rules | API access, field mappings, retries and duplicate handling |
| AI-assisted workflow | A fixed process with an interpretation or extraction step | Validation, human review and what happens when input is uncertain |
| Custom agent | A bounded job that needs context and a choice among allowed actions | Tool permissions, evaluation, action logs, operating cost and ownership |
| Custom AI software | A distinct product or operational system with its own interface and data model | Product scope, integration, adoption, maintenance and commercial case |
When an existing assistant is enough
Try the assistant your team can already access when the task is well supported and a person remains in control of the output. Summarising a document, preparing a draft or finding approved information may not need a separate product. The implementation work may be source permissions, training and a consistent review process rather than custom software.
ChatGPT apps can connect supported services and, where supported, take actions. Availability depends on the app, plan and workspace. Microsoft Copilot capabilities vary by licence, integration and data access. Check the current OpenAI connected-app documentation and Microsoft Copilot overview against the accounts you actually have.
A fair trial uses a short set of real tasks, including difficult ones. Ask the people who do the work to record usefulness, correction time and gaps. If the tool does the job within the required permissions and cost, a custom build needs a specific additional reason to exist.
When a workflow is a better fit
Consider an enquiry arriving in an inbox. The business wants its details checked, the relevant customer found, a follow-up drafted and a task queued for approval. The overall route is known. AI may help interpret the text, while ordinary code controls validation, record matching and delivery. A free-ranging agent would add complexity without necessarily adding value.
A workflow implementation should specify the event that starts it, the accepted inputs, the destination, duplicate handling and the exception queue. Someone must own that queue. A process that works for the happy path but silently loses difficult requests has not solved the operational problem.
When an agent earns the extra work
An agent becomes useful when the next step depends on context. It may need to decide whether to consult a document, look up a record, ask a person for missing information or prepare an approved action. The allowed choices should still be finite and understandable to the process owner.
The extra cost is not simply calling a model. It is connecting identities and systems, defining permissions, handling failures, testing behaviour and supporting the deployment. Our custom-agent service starts with one bounded job, and our cost guide separates implementation from support and usage.
Five checks for any option
- Data: is there an authoritative source, and can the intended user access it?
- Actions: which steps are read-only, which change records and which need approval?
- Evaluation: what real examples must pass, and what is an unacceptable failure?
- Adoption: where will the team use it, and who teaches and supports them?
- Economics: does the measured benefit justify licences, build cost, review effort and ongoing support?
These checks apply whether you buy a licence, use an automation platform or commission an agent. No product label removes the need for process ownership. Use the implementation checklist to record the decisions before rollout.
How we make the recommendation
In an opportunity sprint, we review the work and its constraints, identify the simplest feasible option and return a costed recommendation. Sometimes that is a custom build. Sometimes it is better configuration of something already in use. A useful implementation partner should be able to explain either conclusion without hiding the alternatives.