How to choose an AI vendor
The vendor is the last decision, not the first. Six of these seven steps happen before you compare companies, and they are what make the comparison fair.
| Step | Do this | Detail | Where |
|---|---|---|---|
| 01 | Name the problem, not the technology | Write one sentence: "We lose X because Y happens." Find the problem sheet. If you cannot write the sentence, buy an audit (S-01), not a system. | /problems/ |
| 02 | Measure the baseline for one week | Enquiries answered within an hour; invoices posted within three days; calls answered. Whatever the sheet says to measure — count it now. It becomes the pilot's success test. | /problems/ |
| 03 | Match shape, size and budget | Use the matcher. Its job is to tell you what not to buy at your budget: a build under US$5,000 is usually a tool plus training. | /match/ |
| 04 | Get three quotes against the same brief | One page: outcome, measure, scope in/out, systems to connect, data available, languages, hosting preference, budget band. Send the same page to each vendor. | /services/ |
| 05 | Ask the questions; write the answers down | Pick the risk areas that apply. Vendors who answer "yes" to everything have not understood the question. Note who could show, not just tell. | /vendor-questions/ |
| 06 | Score with the matrix below | Weights first, scores second, price last. Price is one row. | #matrix |
| 07 | Contract for the risk you found | Pilot before build where the answer depends on your data. Fixed scope for what everyone understands, a capped pocket for what nobody does. Ownership, exit and data clauses in — see the pricing models and the licensing guide. | /pricing-models/ |
Score each vendor 1–5 per row
| Criterion | What you are judging | Weight | Vendor A | Vendor B | Vendor C |
|---|---|---|---|---|---|
| Outcome fit | Did they show the outcome on your problem, or a generic demo? | 20 | _ ×20 | _ ×20 | _ ×20 |
| Evidence on your data | Pilot results or a test on your documents, calls or messages | 20 | _ ×20 | _ ×20 | _ ×20 |
| Integration realism | Named method per system; failure behaviour described | 15 | _ ×15 | _ ×15 | _ ×15 |
| Data and licensing | Roles, sub-processors, location, deletion, training opt-out | 15 | _ ×15 | _ ×15 | _ ×15 |
| Ownership and exit | Accounts, number, keys, prompts and export in your name | 10 | _ ×10 | _ ×10 | _ ×10 |
| Support model | Hours, response times, monthly report, change budget | 10 | _ ×10 | _ ×10 | _ ×10 |
| Price and price shape | Six layers separated; volume behaviour; currency terms | 10 | _ ×10 | _ ×10 | _ ×10 |
| Total | Weighted score out of 500 | 100 | — | — | — |
Adjust the weights to your risk: a clinic or bank raises "Data and licensing" to 25; a shop with one WhatsApp number raises "Outcome fit".
Ends the conversation
The WhatsApp number, cloud account or API keys will be in the vendor's name "to make it easier".
The demo runs on their data and they decline a test on yours.
An "AI invoicing" product that issues fiscal invoices without a fiscal device.
A quote with one line and no split between licence, usage, labour, integration, training and support.
"Your data is safe with us" without a country, a sub-processor list or a deletion procedure.
A Shona or Ndebele voice claim with no recording, no error rate and no pilot.
For large, regulated organisations the due-diligence bar is higher; the governance framework at enterpriseai.co.zw covers policy, model inventory and audit trails. This page is the buyer's minimum for any size.