What an AI Agency Does in the First Two Weeks

You have chosen an agency and signed. A week later you have had two calls, filled in a form and sent a few logins. Nothing is built. Is that normal?

Mostly, yes. The first stretch of an AI project is not building. It is deciding what to build, getting access to what it will touch, and agreeing how you will know it works. Agencies that skip this tend to build the wrong thing quickly.

No standard says what must happen in which week, and agencies differ. Public guides from AI vendors and consultancies put discovery anywhere from one to four weeks, and they are marketing as much as research. So treat what follows as our view of a sensible first fortnight for a small business, not a rule. It gives you something to hold your agency to.

Quick answer: in the first two weeks a good agency should confirm one job to automate, map the tools and data it will touch, get access set up in a way you control, build a narrow first version on tools you can keep, and test it with a person checking the output. If week two ends with a long slide deck and no working draft, ask why. If it ends with a live system touching customers and no testing, ask that too.

Days 1 to 3: confirm the job

The agency should start by restating the job back to you in plain words. One workflow, one team, one thing you can measure. If your brief said "customer enquiries", they should narrow it: which enquiries, from which channel, answered by whom today.

Expect questions about how the work runs now. Who does it, how long it takes, where it goes wrong, what a good result looks like. We think success measures should be agreed, and signed off by whoever owns the budget, before any building starts. It is a short conversation and it saves arguments later.

What you should hear is a narrower job than you started with. That is a good sign. A first project that tries to do everything is the one that stalls.

Days 3 to 6: map the tools and the data

Next the agency needs to know what you already run: the inbox, the CRM, the booking system, the accounting package, the shared drive. This is where a good agency saves you money, because it builds on what you have instead of adding a new platform.

It also needs to know where the information lives and what state it is in. Is the price list current? Are customer records in one place or three? Are the answers your staff give written down anywhere? A model can only work with what it is given, and old or scattered data is a common reason a first version disappoints.

Ask the agency to write down what it found. A one-page map of tools, data and gaps is a reasonable deliverable at this point.

Days 4 to 8: access, and who owns it

This is the part to watch most closely, and the one most clients treat casually.

The agency will need access to your accounts. There are better and worse ways to give it.

• Use their own login, not yours. Ask for a named user on each tool, with the lowest permissions that let them do the work, rather than sharing your password. You can switch it off later.
• Know which account holds the automations. If the workflow is built in the agency's account and not yours, it leaves when they do. Some tools have a feature for this. Make, for example, has credential requests, which let a builder collect connections from a client without seeing the client's passwords or keys. Make's help page says the feature is limited to its official partners and enterprise customers, and that the client can revoke access at any time. Whatever tool your agency uses, ask the equivalent question.
• Know where it runs. With n8n, the self-hosted version runs on a server someone has to look after, and n8n's own hosting guide covers the databases, security, monitoring and updates that come with that. Our review's catch is that it is free to run but not free to maintain. If the agency hosts it for you, ask what happens to it when the contract ends.

Our reviews of these tools point the same way: the headline plan is rarely the whole bill. Ask who will pay for the subscriptions underneath, and in whose name.

Days 6 to 10: a narrow first build

Now the agency builds, and it should be small. One workflow that does one job, end to end, on real examples from your business.

For a workflow that moves information between tools, the agency will often reach for an automation platform such as Make or n8n, plus a language model for the parts that need judgement. Our guide to Zapier, Make and n8n covers how they differ. Using a platform like these is a good sign for a small project. It means you can see the workflow, and another person can pick it up if the agency leaves.

For a voice job, such as answering calls, the build looks different. Vapi describes each voice assistant as a transcriber, a language model and a voice, with phone numbers and your systems connected around them. Retell AI runs those layers itself and offers a test playground, simulation tests and a way to transfer a caller to a person. Our reviews of both carry the same catch: the headline rate is only part of the cost per minute, because speech, the model and phone charges stack on top. Ask the agency for an estimate at your call volume. Our comparison of the two is a starting point, and our guide to what to ask before buying a voice agent goes further.

Whatever the job, ask to see the first version working on your own examples, even if it is rough.

Days 9 to 14: test it with a person in the loop

Before anything touches a customer, the output should be checked by a person on your side. Run it on real past cases. Read what it produced. Note where it was wrong and what the agency changed.

The Office of the Australian Information Commissioner says AI output should be treated as probabilistic rather than fact, with human oversight and the ability to verify and override it. It also says to record where AI-generated information came from, to tell people when they are dealing with an AI tool such as a chatbot, and to take care with personal information entered into AI products. Its guidance is general and your obligations depend on your situation, so read the OAIC guidance on commercially available AI products and get advice that fits your business.

Practically, by the end of week two you want four things. A first version that has run on real examples. A list of what it got wrong. A named person on your side who checks the output. And a way to switch it off.

What you should have at the end of week two

• A written scope for one job, with a measure of success.
• A map of your tools and data, and any gaps.
• Access set up under named logins, with a clear answer on who owns the accounts and where the workflow runs.
• A narrow first build, tested on real cases with a person checking.
• An agreed next step: go live in a limited way, extend it, or stop.

You do not need all five to be perfect. You do need to be able to see that each one is happening. If you cannot, ask the agency to walk you through where they are. Our guide to red flags lists the signs that the answer is not good enough.

Where to start

1. Check your own side. Before the project starts, work through our five readiness checks so that access and data are not the cause of delay.
2. Compare the tools your agency proposes. Put them side by side on the compare page and read the Our take on each.
3. Still choosing an agency? Use the Concierge and a person on our team will match you to a reviewed agency, or browse the agencies page yourself.