Choosing an AI Automation Provider: Expertise Over Influencers

There is a lot of money moving into AI automation right now and very little agreement about what competence looks like. That combination reliably produces a market where confidence is cheap and results are not.

Here is how I would evaluate someone, including me.

Ask what has been running, not what has been built

Building an impressive AI demo is genuinely easy now. That is not a criticism of anyone, it is a statement about how good the tooling has become. A capable person can produce something that looks like magic in a weekend.

What remains hard is the same thing that has always been hard: something that runs for a year, handles the inputs nobody anticipated, degrades sensibly when a dependency fails, and does not require its author to be awake.

So the question is not what have you built. It is what have you had running in production for over a year, and what broke. The second half matters more than the first. Anyone who says nothing broke either has not run anything long enough or is not being straight with you.

Ask what they talked someone out of

A provider who has never advised a client not to build something is not advising. They are selling, and everything they tell you should be read in that light.

Good technical partners refuse work regularly, because a meaningful share of what gets requested is either solved by an existing product, better fixed as a process problem, or not yet worth the money. Someone who can tell you a specific story about that has judgment they are willing to apply against their own revenue.

Watch where the proposal spends its words

Read any proposal you receive and count what it is actually about.

If most of it concerns models, capabilities and what the AI can do, you are being sold a prototype. The model is the commodity layer. It is the part that will be better and cheaper in six months regardless of who you hire.

The work is everywhere else: what systems it has to reach, what happens when they are down, who reviews what, how errors surface, what the audit trail looks like, what it does when it does not know. A proposal that treats those as implementation detail has not thought about them, and you will meet all of them anyway.

The demo problem

Demos are run on chosen inputs by the person who built the thing. Every failure mode has been avoided by construction.

If you are being shown one, ask to type your own input. Ask what happens with a badly formatted request, an ambiguous one, or one in a second language. Ask what it does when the system it writes to is unavailable. The response to those questions is far more informative than the demo itself, because it tells you whether the person has thought past the happy path.

Price signals

Genuinely senior work in this market runs somewhere around $100 to $250 per hour. That is a wide band and it reflects real variation in scope and seniority.

Substantially below that, somebody is either junior, subcontracting to people you will never meet, or pricing to win the work and intending to make it back on change requests. Any of those can be fine if you know which it is. The problem is when the price implies a level of attention that the delivery cannot match.

Substantially above it, ask what you are getting for the premium. Sometimes the answer is good, such as deep domain knowledge or regulatory experience. Sometimes it is an account manager.

Certainty is the red flag

The clearest signal I know is a confident price given before anyone has scoped anything.

It feels like competence. It is usually the opposite: either a guess dressed up, or a template being sold regardless of whether it fits. In both cases the number is fiction and the reconciliation happens later, as change requests, at your expense.

The honest version is slower and less satisfying to hear. It sounds like: I do not know yet, here is what I would need to look at to find out, here is what that costs, and the plan is yours whether or not I build it. That is what an MVP Roadmap is, and you should expect something equivalent from anyone competent.

The short version

Hire the person who tells you what they do not know, has run something long enough to have scars, and can point at work they refused. That filter removes most of the market and almost all of the risk.