Short answer
Establish first whether you are talking to a builder, a consultancy or a reseller — they charge similarly and deliver different things. Then judge on five criteria: comparable work you can verify, willingness to narrow the scope, who owns and hosts what, what happens after delivery, and whether one named person in your company will be able to operate it. Start with a single process at a fixed price with a written measure of success. Walk away from anyone who quotes before understanding the process, or who hosts everything under their own accounts.
Disclosure worth making plainly: LYVIA is one of these agencies. The criteria below are the ones we would want a client to apply to us.
Three businesses, one label
The term covers at least three distinct offers, and confusion between them explains most disappointment.
| Type | What you get | Judge them on |
|---|---|---|
| Builders | Your systems connected, and something that runs. | What exists in production afterwards, whether your team can operate it, and how they scope — see commissioning custom software. |
| Consultancies | Assessments, roadmaps, governance. Legitimate work, valuable at a certain size. | The quality of the document — and whether you meant to buy a document. |
| Resellers and platform implementers | A specific product, implemented well. | Whether they say which product they represent before you ask. Not disqualifying; undisclosed is. |
A company of ten to a hundred people with an obvious manual process to remove almost always wants the first. If the proposal in front of you is mostly analysis, check that you meant to buy analysis. And if what you are actually after is being found and cited inside AI answers rather than an internal process removed, that is a different purchase altogether — see what a GEO agency does.
The five criteria that predict the outcome
- Comparable work you can verify. Not logos — a description of a process they automated, what it replaced, and what broke along the way. An account that includes a difficulty is more informative than a polished case study, and much harder to invent.
- Willingness to narrow. The strongest signal available in a first conversation. A partner who reduces your list to the one process worth doing first, and says plainly that some of it is not worth doing, is behaving like someone expecting to be judged on a result.
- Ownership and hosting on your side. Workflows, prompts and configuration under accounts in your name. This is the criterion that determines the cost of leaving.
- An explicit answer about after delivery. Who watches it, who fixes it when a connected tool changes its interface, and what that costs. Automations break at the edges, not in the middle.
- Operability by your team. One named person on your side should be able to see what the automation did and why. If understanding it requires the agency, you have bought a dependency along with the workflow.
Questions for the first call
Each of these is short, and the shape of the answer matters more than its content.
- “Which single process would you start with, and why that one?” A good answer picks one and justifies it by frequency and reversibility. A weak one proposes a phased program covering everything you mentioned.
- “What would make you tell us not to automate this?” Someone with delivery experience has a ready answer — unstable process, contested rules, volume too low to repay the build.
- “What happens to this if we stop working together in a year?” Listen for accounts, credentials, export and documentation. Hesitation here is the answer.
- “Tell me about one that did not work.” Everyone building has at least one. Nobody having one means either very little delivery or a rehearsed conversation.
- “Who is actually going to build it?” Worth asking whenever the person selling is impressive and the delivery team is unnamed.
A quoted price before anyone has looked at how the process actually works is not a competitive advantage. It means the number is a template, and the scope will be renegotiated once the real work is visible.
Pricing models and what each rewards
- Fixed price per deliverable. The safest way to start with a partner you have not worked with, because it forces the scope to be written down. Its weakness is rigidity when the process turns out to be more complex than described — which is why the first project should be small.
- Time and materials. Honest and flexible, and it puts the risk of discovery on you. Reasonable once trust exists, uncomfortable as a first engagement.
- Monthly retainer. Sensible for continuous work and maintenance, and a poor way to begin. A retainer signed before any delivery is payment for availability with no evidence behind it.
- Outcome-based pricing. Attractive in principle and rarely clean in practice, because attributing a business result to one automation invites argument. Where it works, the measure was defined narrowly and in writing before the start — the method for that is in our guide to measuring the ROI of automation.
Ownership: code, accounts, data
This is the section most often skipped and the one that determines what leaving costs. Four points belong in the contract before work starts.
- The workflows and configuration are yours, with the right to modify them and to have someone else modify them.
- Hosting under your accounts. Subscriptions to the automation platform, the model provider and any storage in your company’s name, with credentials you hold.
- Your data remains yours, exportable in a usable format without the agency’s cooperation, and not used for anything beyond your project.
- A named exit procedure. What is handed over, in what form, within what period. Cheap to agree at the start and contentious to negotiate at the end.
None of this is adversarial. A partner who works this way has thought about what happens after the engagement, which is itself a quality signal.
Red flags
- A quote before understanding the process. Covered above, and the most common single warning sign.
- Guaranteed percentages. A promised saving before anyone has measured the current cost is a sales figure, not an estimate.
- Everything hosted under their accounts, with vague answers about access.
- A proposal that reads generically. If your process, your tools and your constraints do not appear in it, it was not written for you.
- No mention of maintenance. Automation connects systems that change without warning. A proposal silent on what happens then has priced only the easy part.
- Pressure to sign quickly. Rare in serious delivery work, common where the revenue matters more than the outcome.
Agency, freelancer or in-house
The deciding factor is how much of this work is recurring. A single well-specified automation you can evaluate yourself is a good fit for a freelancer. A steady stream of them justifies hiring, with the caveat that a single internal specialist recreates the dependency problem inside your own walls.
An agency fits the middle case: several workflows over a period, broader capability than one person, without a permanent hire — and, done properly, leaving you able to operate what was built. If none of the three appeals, the honest fourth option is to start with tools your team can use directly, which is the subject of our guide to automating processes without developers.
Structuring the first engagement
Small, fixed, and complete end to end. One process — chosen with the method in our process audit — a fixed price, a deliverable running in production rather than demonstrated, and a written measure of success agreed before the start.
The first project is a test of the working relationship as much as of the automation: how they behave when something turns out to be harder than expected tells you more than any reference. If it goes well, widen the scope from a position of evidence. If it does not, you have lost a small fixed amount and learned something for the cost. Deciding what should follow it across a year is the subject of our AI strategy roadmap.
Frequently asked questions
What should an AI automation agency actually deliver?
Something running in production that a named person in your company can operate, plus documentation short enough to be read and access to everything it depends on. A slide deck, a strategy document or a demonstration environment are not deliverables — they are stages. If the engagement ends and nothing is running that would survive the agency disappearing, you bought consulting and should have known that going in.
How much does it cost to work with one?
Enough variation that any single figure would mislead, because the range covers a two-week workflow and a bespoke internal system. The more useful question is what the price is attached to: a fixed price for a defined first deliverable makes scope explicit and is the safest way to start with someone you have not worked with. Beware of a monthly retainer signed before anything has been delivered — you are paying for availability without evidence.
Should we hire an agency, a freelancer, or build in-house?
It depends on how much of the work is recurring. A freelancer suits a well-specified one-off where you can judge the output yourself. In-house makes sense once automation is continuous enough to fill a role and you can hire someone who will not be the only person who understands it. An agency fits in between — several workflows over a period, more capability than one person, without a permanent hire. The failure mode to avoid in all three is a system only one individual understands.
Who owns the automations they build?
Whatever the contract says, which is why it should say so explicitly before work starts. At minimum you want the workflows and configuration to be yours, hosted on accounts in your name, with credentials you hold and an export you could hand to someone else. An agency that hosts everything under its own accounts has made leaving expensive, whether or not that was the intention.
What is the strongest signal in a first conversation?
Whether they try to narrow the problem or widen it. A partner who spends the first call establishing which single process is worth automating first — and is willing to say that some of what you asked for is not worth doing — is behaving like someone who intends to be judged on a result. Enthusiasm for everything you mention is a sales posture, not an assessment.
How should the first engagement be structured?
Small, fixed and complete. One process, a defined deliverable in production, a fixed price, and a written measure of success agreed before the start. As a rule of thumb from LYVIA's own engagements rather than a published benchmark, a first project that cannot be described in two sentences is unlikely to finish cleanly. The point of the first project is as much to test the working relationship as to deliver the automation.
If you would like to apply these criteria to us, the first conversation is about which single process is worth automating first — and whether it is worth it at all. Book a call.
