How to choose an AI partner comes down to one question most vendors will not want you to ask: have they built AI into their own business before they built it for anyone else? If the answer is no, you are paying them to learn on your systems. That single filter will cut your shortlist in half before you get to pricing, timelines, or slide decks. The rest of this guide is how to run the remaining half through a real evaluation.
Key takeaways
- Ask a prospective partner to name an AI system running inside their own operating company today, and what record it reads or writes.
- A real engagement starts with a fixed-fee audit before any build; Throughline's AI Capability Audit is $7,500, quoted in writing.
- A builder writes into the software you already run (ERP, CRM, accounting). A reseller connects Zapier to a chatbot and calls it AI.
- Every write to a live record should require a person on your team to approve it. Full autonomy over money-touching systems is a red flag, not a feature.
- If the problem is a broken process or dirty data, do not hire an AI partner yet. Fix the process first, then automate it.
The four kinds of partner you will meet
You will encounter roughly four categories in your search, and they are not interchangeable. Sorting them upfront saves weeks.
The enterprise consultancy sells strategy decks and roadmaps. The work is real, the fees are large, and the deliverable is usually a document, not a running system. Good fit if you need executive alignment and have an internal team to build.
The offshore dev shop writes code by the hour. You bring the spec, they bring the hands. Cheap on paper, expensive in rework, and you are the systems architect whether you signed up for it or not.
The no-code or chatbot agency wires together tools like Zapier, Make, or n8n and layers a chat interface on top. Fine for a marketing autoresponder. Not what you want touching an invoice, a purchase order, or a rebate accrual.
The builder-operator writes AI into the software you already run and hands the system to your team. That is what we do at Throughline, and it is what an AI systems integrator actually looks like when the work is real.
Most guides on how to choose an AI partner skip this taxonomy because it makes the shortlist uncomfortably short. It should.
Questions that expose fit or bluffing
Send these in an email before the first call. The answers, or the dodges, will tell you almost everything.
"What AI system are you running inside your own business right now, and what does it read or write?" A builder answers with a specific system, the software it lives inside, and the record it touches. A reseller answers with a client anecdote or a product demo.
"Show me a system where a person on the client's team approves every write to a live record." Human approval on money-touching systems is the professional standard. If a vendor pitches full autonomy on invoices, rebates, or GL entries, walk.
"What software does the system integrate with, and how?" You want to hear the name of your ERP, your accounting system, your CRM, and the mechanism (API, database view, file drop). If the answer is "we can connect anything," they mean Zapier.
"What does the first engagement cost, and what is the deliverable?" A real answer is a fixed number and a written scope. Ours is $7,500 for an AI Capability Audit, quoted in writing, with the build priced separately after the audit. Vague hourly ranges are a signal that scope will drift and so will the invoice.
"When have you told a prospect not to build?" A partner who has never talked anyone out of a project is a partner who sells regardless of fit.
What a real engagement looks like
A real engagement has three phases and one honest gate between each.
- Audit. A fixed-fee assessment of where AI actually fits in your operation. Not a strategy deck. A short list of specific chores, ranked by feasibility and payback, with a written build proposal for the top one or two. Throughline's audit is $7,500.
- Build. A separate fixed fee, agreed in writing after the audit. The system is written into the software you already run: your ERP, your document store, your accounting package. This is where intelligent document automation and ERP automation stop being slide-deck words and start being lines of code that touch real records.
- Handover. The system is handed over to your team with documentation and a human-approval step on any write. Cloud costs (models, storage, compute) are metered like any other cloud service and quoted in writing before the build starts.
The honest gate is that you can stop after the audit. If the payback is not there, we say so, and you keep the assessment.
First-hand proof
Here is the work we point to when someone asks the question we just told you to ask.
Kelsan is a multi-state distributor running Epicor P21, and it is inside our own group. We built the rebate-and-margin tool we built inside Kelsan directly into P21. The system reads supplier rebate documents, matches them against invoices already sitting in the ERP, and surfaces entries for a person on the Kelsan team to approve before anything posts. To date, it has recovered over $100,000 in margin that would otherwise have been lost to missed or misapplied rebates. Nobody on our team gets to write into P21 unattended. A human approves every entry that changes a record. That is the pattern we bring to any distribution engagement, and it is the pattern any partner touching your ERP should follow.
Red flags
A few signals should end the conversation, not extend it.
Promised ROI numbers before an audit. Anyone quoting savings before they have seen your data is guessing or lying, and either is expensive.
A demo that never touches a real record. If the pitch is a chatbot answering questions about a PDF, that is a search feature, not a system. Ask to see something write into a database.
No named production system inside their own business. If the partner cannot point to AI running in their own operation, they are practicing on you.
Pricing by the hour with no fixed scope. Fine for a staff aug arrangement. Not fine for a build.
Reluctance to include a human-approval step. On anything that touches money, this is not optional.
When not to hire an AI partner
Sometimes the honest answer is: not yet.
If the underlying process is broken, AI will automate the brokenness faster. Map the process on paper first. Fix what is obviously wrong. Then decide whether to automate.
If the data is scattered across spreadsheets, email threads, and one person's head, you have a data problem, not an AI problem. A cheaper first step is a shared system of record.
If the chore happens twice a year and takes an afternoon, the payback math does not work. Save the audit fee.
If you have an internal engineering team with capacity and curiosity, a good use of an outside partner is a short advisory engagement, not a full build. Sometimes the right answer is custom AI solutions built by your own team with a builder-operator as a sounding board.
Book the call
If you are shortlisting partners, run every one of them through the questions above, including us. Book a call with Throughline, or start with a fixed-fee AI Capability Audit ($7,500). The call is free. Call 865-417-3554.
About the author
Throughline is a small team of builders inside Keller Group. We build AI systems into our own operating companies first, then into yours.