Choose the first workflow.
We count the candidates against your own records, record the baseline and recommend one to start with.
AI agency consultingBuyer’s guide
How to choose
an AI agent
for your business.
Choose by the workflow, not by the demo.
Seven questions decide whether an AI agent will hold up in production, and whether you need a platform, a custom build or your own team. None of them is about the technology.
Written by VoxDonna, which builds the custom side. Where a platform is the better buy, the guide says so.
Start with the workflowVoxdonna · buyer’s guide
FIRST QUESTIONIllustrative example
The assessment is useful whether or not you build it with us.
To choose an AI agent, start with the workflow. Seven questions decide it: how often the task repeats, whether the rules are written down, who owns the exceptions, which system holds the truth, whether the agent only reads or also writes, which languages it needs, and how it will be tested. If a platform already fits your process, buy the platform. If the integration or the brand voice is the product, look at a custom build. VoxDonna is an AI agency that builds the custom side, and for a deeper comparison see custom AI agent vs. conversational AI platform.
Seven questions
What to settle
before you shop.
The first three decide whether the task is worth automating at all. The last four decide how safely an agent can be connected to it.
See how we run the assessment- 01
Does the task repeat enough?
A task that runs two hundred times a month justifies the rules it needs. One that runs twice does not, however irritating it is. Count it from your own records before you shop.
- 02
Can the decision be written down?
If your team can state when to act and when to stop, an agent can enforce it deterministically. If the rule is really judgement, the honest scope is to collect and route the work, not to decide it.
- 03
Who owns what goes wrong?
Every workflow has a tail: the ambiguous request, the missing record, the system that is down. If no one owns the tail today, automating the happy path just moves the mess somewhere less visible.
- 04
Which system holds the truth?
The agent should answer from the real record in your ERP, CRM or helpdesk. Records that are missing or unavailable should route to your team rather than becoming a confident guess.
- 05
Does it only read, or also write?
Read access and write access are different decisions. Write access should be enabled for one agreed process, after the acceptance tests have been run.
- 06
Which languages must it handle?
Agree and test the required languages against your own terminology before launch, rather than assuming a vendor’s list covers you.
- 07
What does it do with the bad input?
Anyone can demo the clean request. Ask what happens with the duplicate, the unreadable attachment and the ERP timeout. We publish six such cases for an order agent: read the acceptance pack.
Platform, custom or in-house
Three ways to get
the same agent.
A platform gives you a faster start. A custom build gives you depth in your own systems. Building it yourself puts both on your team. Which is right depends on how close your use case is to a standard template.
| Factor | Platform | Custom build | In-house |
|---|---|---|---|
| Time to first launch | Faster for a standard, template-shaped use case | Slower upfront; scoped to your exact workflow | Depends on your team |
| ERP and CRM depth | Standard connectors; works well for standard configurations | Built for your specific instance, modules and approval rules | Built by your team |
| Who maintains it | The platform vendor maintains the core product | Your build vendor, end to end | Your own team |
| Ceiling on customisation | Bounded by what the platform exposes | No platform ceiling; bounded by what you commission | Bounded by what you build |
| When it is the better buy | Your use case fits a standard template, you need to be live in days, or no one can own an ongoing integration | The integration or the brand voice is the actual product | Someone on your team can own the integration and its exceptions |
Evaluating companies
How to evaluate AI agent
companies and AI agencies.
Seven checks that work on any vendor, including us. A vendor that cannot answer them cleanly has told you something.
- Do the people who advise also build and run the agent, or does the engagement end at a roadmap?
- Is the success measure agreed before the build, with the baseline recorded at the same time?
- Will they tell you not to automate something, or to buy a platform instead?
- Are the acceptance tests published, and does write access follow them?
- Are the build cost, the running cost and the launch plan in writing before work starts?
- Can the agent be deployed into your own cloud account and region where data location matters?
- Can you read a deployment written up in full? Two are: Le Marquier and Usha Martin.
Where to go next
From the questions
to a decision.
Pick the page that matches where you are.
See how an agent is built and tested.
Deterministic checks, a simulated backend first, and write access only after the acceptance pack passes.
AI agent development servicesTry a live agent.
Live voice and WhatsApp demos across industries, with no signup required.
Browse the demo libraryBefore you choose
The practical
questions.
How do I choose an AI agent for my business?
Start with the workflow, not the demo. Check how often the task repeats, whether the rules can be written down, who owns the exceptions, which system holds the truth, whether the agent only reads or also writes, which languages are needed and how it will be tested. If a platform already fits your process, buy the platform.
What should I ask an AI agent company before hiring it?
Ask what the agent does with the duplicate, the unreadable attachment and the system timeout. Ask who owns the exceptions, whether write access follows the acceptance tests, and what the build cost, the running cost and the launch plan are before work starts.
How do I evaluate AI agencies?
Check whether the people who advise are the ones who build and run the agent, whether the success measure and its baseline are agreed before the build, and whether they will tell you not to automate something. Ask to see a deployment written up in full.
Is a custom AI agent always better than a platform?
No. A platform is the better call when your use case is close to its standard templates, you need to be live in days rather than weeks, and no one on your team can own an ongoing integration. A custom build earns its cost when the integration or the brand voice is the actual product.
What makes a good first workflow for an AI agent?
One that repeats often enough to justify its rules, has a decision that can be written down, and has an exception path someone owns. A task that runs twice a month, changes shape every time, or depends on data nobody trusts is a bad first candidate.
Start where it repeats
Name the task
your team keeps redoing.
We will tell you whether an agent fits it — and when it does not.
Scope a workflowA short qualification form first. A discovery call if there’s a fit.
Prefer email? hello@voxdonna.com