How to evaluate an automation vendor before you sign
Most failed automation projects were decided in the sales call, not the build. Someone demoed an impressive agent, nobody asked who owns the code, and six weeks later a workflow breaks on a Saturday with no one on the hook. The checks below are the ones that actually predict whether you get a working system.
They apply whether you are buying AI consulting services from a large enterprise firm, hiring a solo consultant, or bringing a four-person shop in to build one workflow. Run them in order. A vendor who cannot answer the first four in a single call is not ready to quote you, and a vendor who answers all nine in writing has already done more thinking than most.
None of this requires you to understand machine learning. It requires you to be specific about your process and stubborn about ownership, pricing and support. Vendors who genuinely help clients will welcome the questions, because answering them is cheaper than a rescue project later.
What AI consulting services actually include
AI consulting services — advisory and delivery work built around artificial intelligence — cover four separate jobs, and very few vendors do all four well. Strategy decides which use cases are worth funding and how they align to a business outcome. Build writes and deploys the thing. Integration wires it into the systems and business processes you already run. Operations keeps it alive after launch, including the internal adoption work that decides whether anyone uses it.
A firm that only sells AI strategy hands you a roadmap and a slide on AI maturity, then leaves before anything ships. A shop that only builds will happily automate a process that should have been deleted. The firms that help organizations most are the ones that accelerate the two or three jobs you genuinely cannot staff and say so plainly about the rest.
Vocabulary hides scope here. Predictive analytics on your historical data is a different project from generative AI drafting replies, which is different again from agentic AI taking actions in your systems. Our guide to AI agents versus chatbots versus RPA draws the lines. Ask which category each deliverable falls into, in plain words.
Nine checks to run on every automation vendor
Score each one yes, partial or no. Anything below six yeses is a vendor you are gambling on.
1. Do they scope a use case, or sell a platform?
A good vendor asks what happens today, step by step, before proposing anything. A weak one arrives with a platform and looks for somewhere to put it. Ask them to describe your highest-value workflow back to you at the end of the first call. If they cannot, they were selling, not listening. Vendors who lead with a platform tend to price the licence and treat the automation as an afterthought.
2. Who owns the code, the accounts and the AI models?
Get this in writing before money moves. You want to own the source, the automation platform account, the integrations and the API keys. The most common trap is a vendor who builds inside their own tenant, so leaving means rebuilding. On model access, insist on bring-your-own-key so usage bills land on your card at cost. Our BYOK explainer covers why marked-up model usage is the quietest way to overpay.
3. What happens the first time a workflow breaks?
Every automation breaks. An API changes, a field gets renamed, a supplier sends a malformed file. Ask who gets alerted, how fast someone responds, and what failed records do meanwhile. The right answer includes retries, an error queue a human can see, and a named support path. "We monitor it" is not an answer.
4. How do they price, and what is deliberately not in the price?
Ask for a fixed scope with a fixed number, plus a written list of exclusions. The exclusions are where the truth lives: data cleanup, extra integrations, employee training, change requests after sign-off. Hourly billing on a build you cannot specify is how a $10k project becomes a $40k one. For reference, AutomateNexus builds start at $7,500 with a typical build timeline around 30 days.
5. Can they show a working system, not a deck?
Ask to watch a real automation run end to end, live, including a failure. Anonymised is fine. Recorded demos of the happy path prove nothing. If a vendor cannot show one running workflow they built and still support, you are funding their first one. This single check eliminates more candidates than the other eight combined.
6. What governance and data handling do they commit to in writing?
Governance is not a compliance checkbox on an AI project, it is the difference between an agent that can only read and one that can email your customers. Ask which systems the automation can write to, what data leaves your environment, whether prompts and outputs are logged, and how long they are retained. If regulated data is involved, get retention and subprocessor answers before the kickoff, not after.
7. How will you measure whether it worked?
Agree a measurable definition of success before the build starts: hours returned per week, error rate on a specific step, time from request to resolution. Baseline it now, because nobody reconstructs a baseline afterwards. A vendor who resists this is protecting themselves from the number.
8. Will they tell you when automation is the wrong call?
Ask directly: which of these ideas would you refuse to build, and why? A consultant who has never talked a client out of anything has no judgement to sell you. Low-volume, high-variation, high-consequence work is a bad first candidate. So is a process nobody has agreed on, because you will automate the disagreement. Ask who they expect to collaborate with on your side too, since adoption stalls when the people doing the work first hear about it at launch.
9. Does it scale past the first workflow?
Ask how the second and tenth automations reuse the first one's plumbing. Shared authentication, a common logging layer, one place where customer records are resolved. Vendors who build every project as an island make scaling expensive on purpose. A scalable setup means workflow number ten is cheaper than workflow number two, not the same price forever.
Red flags that should end the conversation
Guaranteed percentage savings with no access to your data. Nobody can promise a number before seeing the workflow, so a vendor who does is either guessing or planning to redefine the metric later.
Refusal to name the underlying tools. If they will not say whether it runs on a mainstream automation platform, a custom service or a pile of scripts, you cannot assess risk or hire anyone else to maintain it.
Locked accounts, marked-up model usage, and month-to-month retainers with no deliverables attached. Retainers are fine for real support coverage; they are not fine as a subscription to keep your own system running. And a proposal with no exclusions section, no acceptance criteria and no named contact is a brochure, not a contract.
What AI consulting and automation builds cost
Independents and boutique firms quote hourly, fixed-fee or retainer, and the same work gets packaged all three ways. Fixed project pricing is safest for a buyer with a written scope, because the vendor absorbs the estimation risk. Hourly is fine for exploratory discovery and dangerous for delivery.
Enterprise consultancies price AI initiatives on blended day rates across a team, which is why the same automation can differ by an order of magnitude depending on who you ask.
Concrete reference points from our side: AutomateNexus builds start at $7,500, an MVP runs 4 to 8 weeks, and model usage is BYOK straight to the provider at roughly $30 to $150 a month for most small business workloads. Our paid two-week strategy audit is $2,500. Ask every vendor for the same three numbers.
A scoring sheet you can run in an afternoon
Put the nine checks in a spreadsheet, one column per vendor, and score yes, partial or no. Weight checks 2, 3 and 4 double, because ownership, support and pricing are what cost real money later.
Then send every shortlisted vendor the same one-page description of the workflow, the volume, the systems involved and your success metric. Identical inputs are the only way to compare quotes. Wildly different numbers against an identical brief mean the vendors are scoping different things.
If you do not yet know which workflow to put in that brief, our free automation health audit takes about three minutes, returns ranked quick wins and an annual cost of manual work, and needs no call. The automation playbook covers how to sequence the first three projects once you have picked one.
Frequently asked questions
How much does an AI consultant cost?
It depends on packaging more than skill. Small firms quote hourly, fixed-fee or retainer; large consultancies quote blended day rates. Ask for a fixed price against a written scope with an exclusions list, and treat any quote without one as an estimate. AutomateNexus builds start at $7,500 and our paid strategy audit is $2,500.
What are AI consulting services?
Advisory and delivery work that helps organizations pick, build and run AI solutions. In practice it splits into strategy, build, integration and operations. Confirm which parts your vendor covers and who is accountable for the rest before you sign, because the gaps are where projects stall.
What is the 30% rule in AI?
There is no formal 30% rule in AI. The phrase floats around as a loose heuristic, usually about how much of a budget goes to data work rather than modelling. Treat it as folklore, not a planning input, and budget from your own scoped workflow instead.
How do I choose the best AI consulting company for my organization?
Send three vendors the same one-page workflow brief, score them on the nine checks above, and weight ownership, support and pricing transparency highest. Prefer the one asking the most uncomfortable questions about your process. Reference calls matter more than logos: ask a past client what broke and how fast it was fixed.
Can an automation vendor work with my existing IT infrastructure?
Usually yes, if your systems have an API or a supported connector. The real constraint is legacy software with no integration surface, where the alternative is screen automation or a database-level workaround, both more fragile. Ask the vendor to name the integration method for each system in the proposal.
How do vendors keep my data secure in an AI project?
Ask three specific questions: what data leaves your environment, which subprocessors touch it, and how long prompts and outputs are retained. Then check whether the automation has write access it does not need. Self-hosting the automation layer and keeping model calls on your own key narrows the exposure considerably.
Can AI solutions be scaled as my business grows?
Only if the first build was designed for it. Shared authentication, a single logging layer and reusable integrations make the tenth workflow cheaper than the second. Isolated one-off builds do not scale, they accumulate. Ask to see the reuse pattern before you buy, not after you want the fourth.